University of Southern California Dissertations and Theses

Towards green communications: energy efficient solutions for the next generation cellular mobile communication systems

(USC Thesis Other)

Towards green communications: energy efficient solutions for the next generation cellular mobile communication systems

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content Towards Green Communications: Energy Efficient Solutions for the Next Generation Cellular Mobile Communication Systems by Luhao Wang A Dissertation Presented to the FACULTY OF THE GRADUATE SCHOOL UNIVERSITY OF SOUTHERN CALIFORNIA In Partial Fulfillment of the Requirements for the Degree DOCTOR OF PHILOSOPHY (Electrical Engineering) May 2019 Copyright 2019 Luhao Wang Acknowledgements First and foremost, I would like to thank my Ph.D. advisor, Prof. Massoud Pedram, for his mentorship, passion and encouragement throughout my graduate studies. He in- troduced me to the research world and taught me how to think, investigate and propose solutions. His enthusiasm for discovering emerging topics, conducting scientific explo- rations, and tackling difficult challenges would be forever inspirations for my career. Prof. Pedram gives me the freedom to investigate new topics, and has been extremely supportive of my work through his kind guidance. At my junior year, he reminded me that I need to look at the most fundamental and critical aspects of every problems. When I got lost in details, he taught me that I should always come back to the big picture. I still can remember many early mornings when he was diligently working in his office while I just came to mine and did not move the project fast enough. I have been always moti- vated by not only what he says but also what he does in practice. Without Prof. Pedram I would not have been anywhere where I am now. Next, I would like to thank other committee members in my thesis and qualification exam, Prof. Sandeep Gupta, Prof. Aiichiro Nakano, Prof. Bhaskar Krishnamachari and Prof. Paul Bogdan, for providing valuable feedback to my research and helping me push the boundary. ii During my study at USC, I have the fortune to work with many talented people. My sincere thanks first go out to my SPORT lab current and alumni members. They include Prof. Shahin Nazarian, Prof. Yanzhi Wang, Shuang Chen, Tiansong Cui, Alireza Shaefei, Prof. Xue Lin, Di Zhu, Naveen Kumar Katam, Ting-Ru Lin, Bo Zhang, Hassan Afza- likusha, Mahdi Nazemi, Ghasem Pasandi, Soheil Nazar Shahsavani, Mohammad Saeed Abrishami, Amir Erfan Eshratifar, Marzieh Vaeztourshizi, Arash Fayyazi, Amirhossein Esmaili, Souvik Kundu and Haolin Cong. Without them I could not accomplish what I have done. Specially thanks to Prof. Shahin Nazarian and Prof. Yanzhi Wang for the external help in providing valuable insights to my research. Another special thanks go to Shuang Chen, Tiansong Cui and Alireza Shaefei for the collaboration and contribution to my work. Outside SPORT lab, I would like to also thank Yuankun Xue, Fangzhou Wang, Ji Li, Kun Xue, Yang Zhang, Jianwei Zhang, Ramy Tadros, Jizhe Zhang for the help and cooperation in different areas. Finally, I would like to express my sincerest appreciation to my parents and my fi- ancee for their unconditional support and love. Without their understanding and encour- agement, this dissertation would not be possible. iii Table of Contents Acknowledgements ii List Of Tables vii List Of Figures viii Abstract x Chapter 1: Introduction 1 1.1 Key Enablers for 5G Cellular Networks . . . . . . . . . . . . . . . . . 2 1.1.1 Small Cell Base Stations . . . . . . . . . . . . . . . . . . . . . 3 1.1.2 Massive MIMO . . . . . . . . . . . . . . . . . . . . . . . . . . 3 1.1.3 Network Virtualization . . . . . . . . . . . . . . . . . . . . . . 4 1.1.4 Caching at the Edge . . . . . . . . . . . . . . . . . . . . . . . 5 1.1.5 Device to Device Communication . . . . . . . . . . . . . . . . 7 1.1.6 Cognitive Radio . . . . . . . . . . . . . . . . . . . . . . . . . 7 1.1.7 Cooperative Multi-Point (CoMP) Systems . . . . . . . . . . . . 8 1.2 Sources of Energy Inefficiency . . . . . . . . . . . . . . . . . . . . . . 9 1.2.1 Excessive Small Cells . . . . . . . . . . . . . . . . . . . . . . 9 1.2.2 Cell Association . . . . . . . . . . . . . . . . . . . . . . . . . 10 1.2.3 Resource Management . . . . . . . . . . . . . . . . . . . . . . 10 1.2.4 Non-Ideal Hardwares . . . . . . . . . . . . . . . . . . . . . . . 11 1.3 Energy Efficient Solutions . . . . . . . . . . . . . . . . . . . . . . . . 12 1.3.1 Dynamic Cell Switching . . . . . . . . . . . . . . . . . . . . . 12 1.3.2 Energy Harvesting and Renewable Energy . . . . . . . . . . . . 13 1.3.3 Interference Cancellation . . . . . . . . . . . . . . . . . . . . . 14 1.3.4 Big Data Driven Solutions . . . . . . . . . . . . . . . . . . . . 14 1.4 Organization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16 Chapter 2: Context-Driven Power Management in Cache-Enabled Base Sta- tions using a Bayesian Neural Network 18 2.1 Background and Prior Work . . . . . . . . . . . . . . . . . . . . . . . 19 iv 2.2 System Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23 2.2.1 Channel Model . . . . . . . . . . . . . . . . . . . . . . . . . . 25 2.2.2 Service Model . . . . . . . . . . . . . . . . . . . . . . . . . . 26 2.2.3 Cache Model . . . . . . . . . . . . . . . . . . . . . . . . . . . 27 2.2.4 Power Consumption Model . . . . . . . . . . . . . . . . . . . 29 2.3 Problem Formulation and Solution Method . . . . . . . . . . . . . . . 30 2.3.1 Problem Description . . . . . . . . . . . . . . . . . . . . . . . 30 2.3.2 Control Problem Formulation . . . . . . . . . . . . . . . . . . 33 2.3.3 Solution Method to the Online Control Problem . . . . . . . . . 35 2.3.3.1 Bayesian Neural Network . . . . . . . . . . . . . . . 35 2.3.3.2 Variational Inference . . . . . . . . . . . . . . . . . 37 2.3.3.3 Proposed Context-Driven Online Power Mangement Framework . . . . . . . . . . . . . . . . . . . . . . . 39 2.4 Experimental Results . . . . . . . . . . . . . . . . . . . . . . . . . . . 41 2.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47 Chapter 3: Power Management of Cache-enabled Cooperative Base Stations Towards Zero Grid Energy 48 3.1 Background and Prior Work . . . . . . . . . . . . . . . . . . . . . . . 49 3.2 System Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51 3.2.1 Service Model . . . . . . . . . . . . . . . . . . . . . . . . . . 53 3.2.2 Channel Model . . . . . . . . . . . . . . . . . . . . . . . . . . 54 3.2.3 Energy Consumption Model . . . . . . . . . . . . . . . . . . . 55 3.3 Problem Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . 57 3.4 Lyapunov Optimization Solution . . . . . . . . . . . . . . . . . . . . . 59 3.5 Experimental Results . . . . . . . . . . . . . . . . . . . . . . . . . . . 63 3.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66 Chapter 4: Concurrent User Association and Dynamic Switching in Cells 69 4.1 Background and Prior Work . . . . . . . . . . . . . . . . . . . . . . . 69 4.2 System Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72 4.3 Problem Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . 74 4.4 Proposed Solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76 4.4.1 Factor Graph Model . . . . . . . . . . . . . . . . . . . . . . . 77 4.4.2 Message Passing Max-product Algorithm . . . . . . . . . . . . 79 4.4.3 Proposed Solution . . . . . . . . . . . . . . . . . . . . . . . . 81 4.5 Experimental Results . . . . . . . . . . . . . . . . . . . . . . . . . . . 82 4.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85 Chapter 5: Joint Transmission and Charging Optimization in Wireless Pow- ered IoT Systems 87 5.1 Background and Prior Work . . . . . . . . . . . . . . . . . . . . . . . 87 5.2 System Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89 v 5.3 Problem Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . 92 5.4 Proposed Solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94 5.5 Experimental Results . . . . . . . . . . . . . . . . . . . . . . . . . . . 97 5.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99 Chapter 6: Final Remarks and Looking Ahead 101 Bibliography 104 vi List Of Tables 2.1 Summary of Different Type of Base Stations. . . . . . . . . . . . . . . 19 2.2 Simulation Parameters. . . . . . . . . . . . . . . . . . . . . . . . . . . 43 3.1 Simulation Parameters. . . . . . . . . . . . . . . . . . . . . . . . . . . 66 4.1 Simulation Parameters. . . . . . . . . . . . . . . . . . . . . . . . . . . 83 vii List Of Figures 1.1 5G network system overview. . . . . . . . . . . . . . . . . . . . . . . . 2 2.1 Graphical illustration of a two-tier cache-enabled HetNet. . . . . . . . 24 2.2 Service model of the two-tier cache-enabled HetNet. . . . . . . . . . . 27 2.3 Illustration of the proposed BNN. . . . . . . . . . . . . . . . . . . . . 39 2.4 Cumulative energy consumption to serve requests from connected users of the sBS under study. . . . . . . . . . . . . . . . . . . . . . . . . . . 44 2.5 Cumulative regret in terms of energy consumption. . . . . . . . . . . . 44 2.6 The relationship between total energy consumption and sBS/MBS den- sity ratio S M . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45 2.7 Number of cache hits on the sBS if active v.s. actual decision made by the sBS for the first 5000 time slots. . . . . . . . . . . . . . . . . . . . 46 3.1 Graphical illustration of a two-tier cache-enabled HetNet. . . . . . . . 52 3.2 Energy flow in an off-grid base station. . . . . . . . . . . . . . . . . . . 55 3.3 Comparison of average download throughput per user . . . . . . . . . . 67 3.4 Trace of state of charge in the MBS . . . . . . . . . . . . . . . . . . . 67 4.1 System model. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72 4.2 Factor graph representation of the JointEMIN problem. . . . . . . . . . 79 4.3 Energy consumption for 8 BSs and 4000 UEs at each time slot. . . . . . 84 4.4 Throughput for 8 BSs and 4000 UEs at each time slot. . . . . . . . . . 85 viii 4.5 Normalized energy efficiency for different number of BSs. . . . . . . . 86 5.1 System model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90 5.2 The average throughput of network system of 20 UEs at different time slott. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99 5.3 The average battery state of charge (SoC) of 20 UEs at different time slott.100 5.4 The average throughput of network system for different number of UEs. 100 ix Abstract The proliferation of multimedia infotainment applications and smart consumer elec- tronics (e.g., cellphones, tablets, wearable devices, laptops) has remarkably impacted how we entertain, interact and communicate. Video streaming, HDTV and social net- works fuel the exponential growth of global mobile data. Consequently, current network systems already reach their capacity limits in highly populated area during peak hours. Congestion problems arise especially in the backhaul links. Academia and industry in the field of communications have reached a consensus that incremental improvements on the existing infrastructure fail to meet the explosive data demands of the foreseeable future. The goals of next generation mobile network (5G) are broad and are presumed to include much greater throughput, much lower latency, and lower power consumption. Promising solutions such as heterogeneous network (HetNet), cell sleeping techniques, cache-aided base stations (BSs) and renewable energy power supply have been proposed and deployed to address those challenges. Although the deployment of dense small cell base stations (sBSs) in HetNet can efficiently offload traffic from the existed macro cell base stations (MBSs), the further benefits have not be harnessed well enough and the potential overhead should be carefully alleviated. This thesis aims to address these challenges by embracing novel mathematical frame- works in control theory, statistical machine learning and probabilistic graph model and x advocate for intelligent, energy-efficient and low complexity methodologies for the man- agement of entities (BSs, user equipments, power beacon) for next generation network systems. This thesis proposes a variational inference (VI) based Bayesian neural network (BNN) to learn from user’s contexts to drive the decision of dynamically switching an sBS between the active mode and the sleep mode to minimize the total energy consump- tion by solving a multi-armed bandit problem. This distributed approach is independent of cache updating policy and can adaptively learn from the requests from users. A pos- terior distributions for every weights and biases are updated by VI which is efficient especially for neural networks with large scales. No special parameter tuning is needed for the scenarios that user’s preferences change rapidly. This thesis derives an efficient online near-optimal control policy with proved opti- mality gap for cooperative BSs powered solely by renewable energy in HetNet by Lya- punov optimization theory and convex optimization. In our studied framework, all BSs are powered by energy harvesting modules and rechargeable battery packs are deployed to deal with the intermittent energy supply, such as solar energy, so that harvested energy can be used at a later time. A virtual queue and an energy storage queue are maintained at each BS node for the purpose of balancing QoS and preventing the quick depletion of battery. This is achieved by solving their independent energy minimization sub-problems before making an active/sleep decision. The developed efficient and scalable framework can maximize the average throughput with a guaranteed optimality gap by avoiding fre- quent power outage. This thesis proposes a distributed framework for a cellular network consisting of one MBS and a set of sBSs and user equipments (UEs). UEs request to download files from BSs through the down-link. We jointly consider the problem of user association and xi dynamic-switching of BSs. The energy minimization problem with power and channel constraints is formulated as an integer non-linear programming problem. Then, a belief propagation based distributed approach is proposed to solve this problem. The effec- tiveness of the proposed framework in practice is demonstrated by experimental results based on realistic user request traces by outperforming other baseline algorithms. In further, a throughput maximization problem in wireless powered IoT network sys- tems is investigated. This work proposes an opportunistic joint transmission and charging management framework. This work considers a wireless powered IoT network system consisting of a power beacon (PB), a set of IoT UEs and a BS. The PB transmits RF energy beams to IoT UEs. Each UE is equipped with a battery of limited capacity and is solely powered by the harvested energy provided by the PB. Then, a online opportunis- tic control solution is proposed and its optimality gap compared to the optimal solution has also been proved. The average throughput is improved against the greedy algorithm baseline. In summary, this PhD work explores energy efficient solutions for the next generation network systems. Several power management and resource allocation policies with dif- ferent capabilities and under different system scenarios have been carefully investigated. With the theoretical and computational contribution combined, this PhD work aims to serve as the basis for the intelligent core of future autonomous network communication with self-organizing, self-learning and self-optimization capabilities towards the goals of high throughput, low latency and energy-efficient future network systems. xii Chapter 1 Introduction The proliferation of multimedia infotainment applications and smart consumer elec- tronics (e.g., cellphones, tablets, wearable devices, laptops) has remarkably impacted how we entertain, interact and communicate. Video streaming, HDTV and social net- works fuels the exponential growth of global mobile data. According to Cisco’s report [4], the annual mobile traffic volume will reach 587 exabytes by 2021, which is 122 times more than that in 2011. Usually, MBSs are designed for the coverage of large areas and would have scarce resource capabilities to handle high data rate transmission for crowded metropolitan hotspots and indoor environments. To accommodate the ex- plosion of mobile data traffic aiming to reduce the gap between network capacity and data transmission rate, the evolving fifth generation (5G) cellular networks have been under extensive study in telecommunication industries and academia. With 1000 more connected devices on site, core characteristics of 5G cellular networks are achieving 100 data transmission rate, 1=5 end-to-end latency, 1=10 energy consumption and 1 1=5 network management operation expenditure [1]. In order to address these chal- lenges, emerging technologies are proposed, investigated and even deployed aiming to be smarter, faster and more efficient. Here are some of the most promising ones. 1.1 Key Enablers for 5G Cellular Networks Figure 1.1: 5G network system overview. 2 1.1.1 Small Cell Base Stations Due to the limited capacity of existing network infrastructures, the undesired large la- tency could be induced during peak-traffic hours especially at urban areas, stadiums that with massive user requests. Thus, users could suffer from the poor quality-of-service (QoS) especially those at the edge of base stations’ coverage. In order to enhance the network capacity and provide an ubiquitous coverage, the dense deployment of compact, low-power small cell base stations (sBSs) is proposed [45]. An sBS is basically a minia- ture base station overlaid within the coverage of existing macro base station (MBS). This umbrella term could include femtocells, pico cells and micro cells. As a way of adding capacities to an existing network, the dense deployment of sBSs, i.e., network densifica- tion, could be in door or outdoor, especially hot-spot areas heavy with users. With con- nections to core networks, sBSs could offload data traffic from already congested MBS and provide higher data rate with reduced transmission distance. However, the increased inter-cell inference between sBSs must be carefully mitigated in the ultra-dense deploy- ment scenario even the transmission power is low compared to MBS [88]. Moreover, the more frequent handover of users at cell edges requires new mobility management scheme [37]. 1.1.2 Massive MIMO Massive multiple-input multiple-output (MIMO) [50] is a compelling technology for further improving the capacity of network system. The main concept is to deploy large antenna arrays at base stations (BSs) to simultaneously transmit data to multiple user equipments (UEs) using the same time and frequency resources. Since the performance 3 of data transmission is significantly influenced by noise and fading in propagation chan- nels, multiple paths from BSs to UEs can avoid the deterioration of transmission for being trapped in channel imperfections. Therefore, massive MIMO can improve wire- less communications by leveraging the law of large number. Other benefits of massive MIMO include extensive use of inexpensive low-power components, reduced latency, simplification of the MAC layer, and robustness against intentional jamming. The key characteristics for designing deploying the massive MIMO BSs are the number M of antennas at the BS, the numberK of active UEs can be served, and the transmit power that uniformly cover the interested area (M >> 1 andK >> 1). Many previous work [16] focus on the derivation of appropriate power models from those key characteristics to gain insights on how to design energy efficient massive MIMO systems. 1.1.3 Network Virtualization Network virtualization is a technique that abstracting physical infrastructures and ra- dio resources in a software as-a-service manner. By splitting the physical resources to support different types of services, the differentiation could optimize the performance considering particular characteristics of services. However, the user experience would not be influenced because of the isolation of layers. Cloud radio access network (C- RAN) is one instance of network virtualization at the BS level. In a C-RAN, the base- band processing unit (BBU) is decoupled from BSs and implemented on a centralized processing resource (could be cloud), rather than on the local hardware of BSs. This allows a scalable deployment of remote radio heads (RRHs), which retains only radio functionalities and simplifies the architecture of full-fledged BSs. The key advantages 4 are scalability, low-cost and easy to do resource management [76]. Other key ways to help the system achieve network virtualization are (i). network slicing, which will en- ables network operators to customize the quality-of-experience (QoE) for different type of services even when sharing common computing, storage or connectivity resources. (ii). network function virtualization, which enables network operators to better manage network infrastructures and simplify the delivery of new services. 1.1.4 Caching at the Edge One major reason that network densification can help reduce the traffic volume on backhaul links is that the deployed small base stations (sBSs) introduce an additional level of content caching which further lowers the cost of acquiring the data requested by users. Since it has been observed that a small set of popular multimedia content files usually contribute to a large portion of the network traffic, saving popular files in base stations is preferred over repetitively fetching these files via backhaul links because of the saving of both backhaul capacity and energy consumption. It has been proved that caching of various types of data including text, image, and video is applicable in 3G [32] and 4G LTE [70] cellular networks and can reduce up to 50% of traffic volume in the core network. There has been a large body of research addressing the management of cache-enabled base stations [38, 69, 13, 56], which try to improve the QoS in terms of round-trip time and/or throughput subject to limited backhaul bandwidth. Stemming from the observation that a large amount of mobile multimedia traffic is due to duplicate downloads of a small portion of popular contents (e.g., popular music videos) with large sizes, researchers and engineers have been investigating effective ways 5 to reduce the duplicate content transmissions by adopting intelligent caching strategies inside the mobile networks, and enabling mobile users to access popular content with less traffic load. More concretely, cache-enabled base stations would have (i) large storage capacity, (ii) localized, high-bandwidth communication capabilities which enable high frequency reuse, and (iii) low rate backhaul links which can be wired or wireless. They can cache popular files and serve requests from mobile users by enabling localized com- munication and hence frequency reuse. The key point is that if there is enough content reuse, i.e. many users are requesting the same video content, caching can replace back- haul communication. From the perspective of internet service providers (ISPs), this also helps reduce traffic exchanged inter- and intra-ISPs. For users, this would significantly reduce the response time required to fetch wanted files. On the other hand, the passion of deploying cache entities is also driven by the trend that the capacity of modern-day stor- age units, such as solid-state-drives (SSDs), has increased exponentially with continued declining costs. The concept of caching has been in existence for almost 20 years with its origins traced back to the widespread adoption of the Internet and consequently the proliferation of internet congestion as a major bottleneck. The bottleneck was eased by the invention of content delivery networks (CDNs) and the exploitation of web caching. We know that the key performance metric for caches is hit rate, and the idea of using local storage to reduce the pressure on the backhaul link would only work if relatively large portion of requested content files reside in the local caching entities. Due to the vast amount of content available in multimedia platforms, not all avail- able content can be stored in local caches. Therefore, deciding which files to be cached 6 and how to update the stored content files aiming to improve hit rate is the key to fur- ther benefit from this idea. Therefore, by smartly exploiting the statistical traffic patterns and users context information (i.e., file popularity distributions, age, location, type of terminals and mobility patterns), it would result in a better prediction of when users con- tents are requested with the amount of resources needed, and at which network locations contents should be pre-cached. 1.1.5 Device to Device Communication The paradigm of having BSs as the core data transmission source unavoidably needs to be shifted due to the spectral scarcity and capacity limitation. Devices can refer to smart phones, vehicles or any type of UEs with the connectivity to the network. Device to device (D2D) communication is a concept of implementing devices as relaying nodes for each other. Without the involvement of BSs or with limited involvement, nearby devices can share data (commonly interested contents and services), resources (spectral resource or computation resources) [79]. Constructing a network mesh with D2D capabilities will improve the spectral efficiency and data transmission rate due to reduced geographical distance. Furthermore, D2D communications can ultilize the licensed spectrum bands, and thus can guarantee uniform service provision and QoS. Moreover, D2D can offload data traffic from BSs and alleviate the network congestion [35]. 1.1.6 Cognitive Radio In conventional cellular networks, different types of services have different QoS re- quirements. For example, voice service is low volume but delay sensitive and multimedia 7 downloading is high volume but delay in-sensitive. On the other hand, the network traffic is unevenly distributed across space and time. The capacity of existing networks may be insufficient during some peak hours but heavily underutilized during other time. Cog- nitive radio is a technique to allow a cellular network to lease underutilized frequency resources without causing serve interference. In this way, a mobile network operator (MNO) can expand the cellular network during high traffic time without purchasing ex- pensive licensed band resource [42]. With the emerging technologies to address challenges, energy efficient approaches are desirable taking into account the carbon emission and energy cost. According to [44], the ICTs (information and communication technologies) consume about 3% of the worlds electrical energy which accounts for approximately 2% of the CO2 emission foot- print in the whole world; 9% of this consumption of ICTs is caused by communication networks[31]. We first take a look at sources of energy inefficiency in 5G cellular net- works. 1.1.7 Cooperative Multi-Point (CoMP) Systems Coordinated Multiple Point transmission and reception (CoMP) was initiated in Septem- ber 2011 in Third Generation Partnership Project (3GPP) [3] to support coordinated transmission in the downlink and coordinated reception in the uplink. The assumed deployment scenarios include homogeneous configurations, where the points are differ- ent cells, as well as heterogeneous configurations, where a set of low-power points (e.g., remote radio heads or sBSs) are located in the geographical area served by an MBS. For 8 coordinated transmission in the downlink, the signals transmitted from multiple trans- mission points are coordinated to improve the received strength of the desired signal at the user equipment (UE) or to reduce the co-channel interference. Download link CoMP strategies can be categorized into [51] i). coordinated scheduling/beam forming (CS/CB) scheme in which scheduling and beaming decisions are made in a coordinated manner among all BSs in the region, but data is only transmitted from one BS. ii). dy- namic point selection, namely, the serving BS would change dynamically at a millisec- ond level according to wireless resource availability and channel state information. iii). joint transmission (JT), which is a technique that allows to transmit data from multiple BSs coherently or non-coherently to improve on the received signal quality and through- put. In coherent JT, the transmission signal from multiple BSs is jointly precoded to achieve coherent combining in the wireless channel. On the other hand, in non-coherent JT, UE would receive multiple transmissions individually precoded by each BS without consideration for coherent combining. 1.2 Sources of Energy Inefficiency 1.2.1 Excessive Small Cells While network densification helps improve the QoS, it is associated with a significant power consumption overhead because of the large number of deployed sBSs. According to data in references [11, 89], there can be 100 times more sBSs than MBSs in a cellular network in practice, resulting in twice as much the total peak power consumption. As discussed before, the data traffic can fluctuate geographically and spatially. During a 9 certain period of low traffic time, powering on underutilized sBSs would result in energy inefficiency. Moreover, although placing dense sBSs improve data transmission rate by reducing transmission distance, the addition of interference could heavily deteriorate the network energy efficiency. 1.2.2 Cell Association In cellular network systems, how to associate UEs with BSs is a fundamental prob- lem. Under the scheme of next generation network dominated by sBSs, smaller coverage area incurs more frequent user handover. The computation complexity also increases with more candidate base stations due to the smaller communication radius. Previous maximum reference signal received power (RSRP) based cell association framework is known to be inefficient for heterogeneous network (HetNet) system because the discrep- ancy of transmission parameters. For example, MBSs could always have the largest transmission power and serve the majority of UEs. In order to fully take advantage of network densification, new cell association frameworks need to be developed to address the challenges of load balancing and traffic offloading. 1.2.3 Resource Management With new technologies in next generation networks, network operators need to fully exploit the potential and additional degrees of freedom to achieve the utility maximiza- tion. Intelligently managing energy and spectrum resources is the key to achieve higher energy efficiency. We know that cognitive ratio technique allows the leasing of underuti- lized spectrum resource and unused spectrum bands can be independently identified by 10 UEs. However, new harmful interference are doomed to be introduced for the primary spectrum holders if careful management is absent [29]. 1.2.4 Non-Ideal Hardwares Although massive MIMO is attractive for network deployment by using low-cost components to construct large arrays, the used hardware components are particularly prone to many non-idealities (e.g., amplifier non-linearities, I/Q imbalance, phase noise, and quantization errors [15]). Even the efforts of compensation algorithms are used, the non-idealities cannot be completely eliminated. The non-idealities can be divided into transmitting side and receiving side. At the transmitting side, the aggregation of imper- fections in amplifiers, converters, mixers, filters and oscillators would create a mismatch between the intended transmitting signal and what is actually generated. At the receiving side, received signals could be distorted in the reception process. In the cellular network system, Power Amplifiers (PAs) are used to increase the power level of the transmit sig- nal so that the corresponding received signal can be demodulated by the receiver under an error probability constraint. In conventional cellular networks, PAs are the major power consumer in wireless cellular network, of which 50%-80% of consumption is consumed at PAs [40, 18]. Linearity and efficiency are the two main characteristics of PAs. Lower linearity could result in higher signal distortion that requires larger signal to noise ra- tio (SNR) to decode it. Efficiency is related to the ratio between converted RF power and input DC power. Note that having a higher efficiency could minimize the energy consumption and thermal dissipation. 11 1.3 Energy Efficient Solutions In next generation network systems, new approaches bring not only opportunities but also challenges. Technologies to enhance energy efficiency have become a critical design and operational consideration due to 1) increasing energy prices and 2) growing attention toward environmental factors such as climate change and associated carbon emissions. In order to address the energy inefficiency problems in 5G cellular network systems, promising approaches are under intensive studies. Before diving into some of them, we first review the fundamentals of energy efficiency. In network systems, energy efficiency (EE) is defined as the number of bits transferred per Joule of energy, which can be calculated asEE = Average Throughput [bit=cell=s] Power Consumption [Joule=cell=s] . Energy efficient solutions should focus on reducing power consumption and improving the average throughput at the same time. 1.3.1 Dynamic Cell Switching From an energy efficient standpoint, placing dense sBSs reduces the transmission distance, which is beneficial to EE due to smaller path fading. However, EE would de- grade because of additional interfences and low-utilization BS nodes. Recent efforts have been made related to power saving in cellular networks with the introduction of sleep modes for BSs, i.e. dynamic cell switching. The concept of dynamically switching BSs between active mode and sleep mode can further reduce energy consumption without degrading QoS. In [25], a protocol was proposed for an orthogonal frequency-division multiple-access sBSs to completely switch off its radio communication and associated 12 processing. E. Oh [66] et al. present a implementable switching-on/off base energy sav- ing algorithm to reduce the energy consumption in wireless cellular network. The online and distributed algorithm shows a comparable performance to the exhaustive search op- timal solution. The work [81] reduces the energy per information bit by making sBSs transmit at full load during good channel conditions and go into sleep otherwise. 1.3.2 Energy Harvesting and Renewable Energy Instead of drawing electricity from the power grid, energy harvesting technologies enable communication systems to obtain energy from ambient radio frequency resources and renewable energy resources (e.g., solar and wind). In work [85], authors study the power management problem for a set of “off-grid” base stations in a cellular network hi- erarchy that are powered solely by on-site renewable energy sources. A network through- put maximization problem is mathematically solved by using Lyapunov framework. Re- newable energy resources are generally intermittent in nature and cannot be predicted perfectly. The uncertainty raises new challenges for the pure usage of renewable energy because insufficient energy would cause service outages. One compromise approach is to design a hybrid system that incorporates many resources in a compensation manner to prevent service outage. In work [53], BSs that powered by both the power grid and renewable energy sources are investigated to provide uninterrupted services. User asso- ciation aims to reduce on-grid power consumption by maximizing the utilization of green power harvested from renewable energy sources, as well as enhance network quality of service by minimizing the average traffic delay. Alternatively, in an RF-powered energy harvesting network, energy can be harvested from dedicated low-power RF signals using 13 components of an RF antenna module, impedance matching parts and a voltage multi- plier to collect and convert RF signals into electricity. In order to address the uncertainty in energy generation, batteries are always used to store the excessive energy for the future use. Energy allocation and scheduling need to be carefully done in order achieve higher EE. 1.3.3 Interference Cancellation Full duplex communications is widely recognized as one of the key technologies for 5G wireless communication systems because it is expected to double the spectral effi- ciency of half-duplex systems. However, full duplex technology is prevented from being widely used in commercial systems due to the phenomenon self-interference (signals are received by its own receiver), which makes it impossible to decode. To deal with the self-interference issue in full-duplex systems, researchers have developed several self- interference cancellation techniques, the objective of which is to mitigate or cancel the self-interference to noise level [75]. For example, designing an antenna that relaxes the coupling between a transmitter and a receiver, mimicking the self-interference signal and subtracting from it. 1.3.4 Big Data Driven Solutions Inspired by big data and artificial intelligence, next generation network systems can learn from contexts and make smart decisions about resource allocation, scheduling and 14 caching contents to achieve ultra-low latency and high EE. In the work [84], user con- texts (e.g., ages, genders and professions) are utilized to drive the decision of dynam- ically switching a small cell base station between the active mode and the sleep mode to minimize the total energy consumption. The online control problem is formulated as a contextual multi-armed bandit problem. A variational inference based Bayesian neu- ral network is proposed as the solution method, which implicitly finds a proper balance between exploration and exploitation. In next generation network systems, many com- mon problems can also be formulated into unsupervised learning problem. For example, sBSs have to be carefully clustered to avoid interference using coordinated multi-point transmission (CoMP), while the mobile users are clustered to obey an optimal offloading policy, the devices are clustered in D2D networks to achieve high EE [47]. In the work [87], authors formulate a mixed integer programming problem to jointly optimize both the gateway partitioning and the virtual-channel allocation based on k-means cluster- ing. Alternatively, networks systems are always modeled as Markov decision processes (MDPs). The system states are carefully defined and can make transitions between dif- ferent ones after certain kinds of decisions are made (e.g., transmission power or BSs on/off). In [8], the transmission power control problems of systems were investigated using the MDP model, where the state space includes the battery state, the channel state, the packet transmission/reception states, and an action is defined as transmission power for sending packets. 15 1.4 Organization In chapter 2, a novel proactive and decentralized power management method for sBSs in a cache-enabled multi-tier HetNet is presented. User contexts are utilized to drive the decision of dynamically switching a small cell base station between the active mode and the sleep mode to minimize the total energy consumption. The online control problem is formulated as a contextual multi-armed bandit problem. A variational inference based Bayesian neural network is proposed as the solution method, which implicitly finds a proper balance between exploration and exploitation. Experimental results show that the proposed solution can achieve up to 46.9% total energy reduction compared to baseline algorithms in the high density deployment scenario and has comparable performance to an offline optimal solution. In chapter 3, we investigate the cooperative transmission and power management problem for a set of “off-grid” base stations in a cellular network hierarchy that are pow- ered solely by on-site renewable energy sources. The network throughput maximization problem is mathematically formulated as a mixed-integer non-linear programming prob- lem. In the proposed formulation, a BS can adjust its transmission power in a coordinated multipoint communication scheme and/or switch to a “sleep mode” for energy saving. Based on the Lyapunov optimization theory, an efficient near-optimal solution method is proposed with provable bound of the optimality gap. Experimental results on a realistic setup show that the proposed algorithm can achieve up to 2.96x download throughput per user compared to some baseline algorithms. In chapter 4, we investigate joint user association and dynamic base station ac- tive/sleep switching problem in cache-enable HetNet system. We first formulate the 16 energy consumption minimization problem and prove its NP-hardness. Then, based on factor graph model and max-product algorithm, an efficient distributed message passing solution method is proposed. Experimental results show that the proposed algorithm can achieve up to 92% average energy saving rate compared with some baseline algorithms and improve the energy efficiency up to 5.3X. In chapter 5, we investigate a throughput maximization problem in wireless powered IoT network systems. We propose an opportunistic joint transmission and charging man- agement framework. We consider a wireless powered IoT network system consisting of a power beacon (PB), a set of IoT UEs and a BS. The PB transmits RF energy beams to IoT UEs. Each UE is equipped with a battery of limited capacity and is solely powered by the harvested energy provided by the PB. Then, a online opportunistic control solu- tion is proposed and its optimality gap compared to the optimal solution has also been proved. The average throughput is improved by 4.38X by using our proposed online control framework against the greedy algorithm baseline. Finally, chapter 6 concludes the thesis. 17 Chapter 2 Context-Driven Power Management in Cache-Enabled Base Stations using a Bayesian Neural Network In this chapter, a novel proactive and decentralized power management method for small cell base stations in a cache-enabled multi-tier heterogeneous cellular network is presented. User contexts are utilized to drive the decision of dynamically switching a small cell base station between the active mode and the sleep mode to minimize the total energy consumption. The online control problem is formulated as a contextual multi-armed bandit problem. A variational inference based Bayesian neural network is proposed as the solution method, which implicitly finds a proper balance between explo- ration and exploitation. Experimental results show that the proposed solution can achieve up to 46.9% total energy reduction compared to baseline algorithms in the high density deployment scenario and has comparable performance to an offline optimal solution. 18 2.1 Background and Prior Work With the proliferation of Internet-of-Things (IoTs) and smart-phones, there has been an explosion of mobile data. According to Cisco’s report [4], the annual mobile traffic volume will reach 587 exabytes by 2021, which is 122 times more than that in 2011. As a result of the demand for ever more network capacity, existing backhaul links and core networks are faced with greater challenge. Among all investigated solutions to the soaring mobile data traffic, deployment of ultra-dense small cells, a.k.a. network den- sification, as proposed in some preliminary 5G cellular network standards (e.g. METIS project [67]), is deemed the most promising solution. Network densification not only partially shifts the burden from the core network and backhaul links to the edge of the network and enable higher transmission speed [36], but also synergizes well with other key techniques in 5G cellular networks such as massive multiple-input multiple-output and millimeter wave transmission [6, 14] because it reduces the communication distance between a user and the nearest base station. Small cells are basically miniature base stations that have a smaller coverage area, lower power consumption and less number of concurrent connected users compared with Macro base stations. Small cell is a usually a umbrella term that consists of femtocell, pico cell and micro cell. Their characteristic parameters can be summarized in Table 2.1. Table 2.1: Summary of Different Type of Base Stations. Cell Type Transmission Power (W) Range (km) Users Locations Femtocell 0.001 to 0.25 0.01 to 0.1 1 to 30 Indoor Pico Cell 0.25 to 1 0.1 to 0.2 30 to 100 Indoor/Outdoor Micro Cell 1 to 10 0.2 to 2.0 100 to 2000 Indoor/Outdoor Macro Cell 10 to>50 8 to 30 >2000 Outdoor 19 In addition, another major reason that network densification can help reduce the traf- fic volume on backhaul links is that the deployed small base stations (sBSs) introduce an additional level of content caching which further lowers the cost of acquiring the data requested by users. Since it has been observed that a small set of popular multimedia entities usually contribute to a large portion of the network traffic, saving popular files in base stations is preferred over repetitively fetching these files via backhaul links because of the saving of both backhaul capacity and energy consumption. It has been proved that caching of various types of data including text, image, and video is applicable in 3G [32] and 4G LTE [70] cellular networks and can reduce up to 50% of traffic volume in the core network. There has been a large body of research addressing the management of cache- enabled base stations [38, 69, 13, 56], which try to improve the quality of service (QoS) in terms of round-trip time and/or throughput subject to limited backhaul bandwidth. While network densification helps improve the QoS, it is associated with a signifi- cant power consumption overhead because of the large number of deployed small base stations. According to data in references [11, 89], there can be 100 times more sBSs than macro base stations (MBSs) in a cellular network in practice, resulting in twice as much the total peak power consumption. Hence, from the perspective of energy efficiency, sBSs need to be self-organizing, low-cost, and energy-efficient. Based on the observa- tion that sBSs only help when their cached contents are requested, some opportunistic management policies have been proposed to switch an sBS into a “sleep mode” when no user request can be serviced [91, 30]. However, such techniques can only achieve limited energy saving because the sBS needs to monitor user traffic even in the “sleep mode” in order to resume normal operation upon arrival of new requests. 20 Further energy saving can be achieved by proactively putting some of sBSs into a deeper sleep mode in which sBSs do not serve or monitor user requests. While the cov- erage of the cellular network is not affected when some sBSs are not active because at least one MBS can be reached by any user, the system may suffer from reduced through- put and/or energy efficiency. In order to avoid the case in which an sBS is not active when its cached contents are frequently requested, an accurate estimation of the user demand in the future is required. In fact, designing a predictor to determine whether the cached contents of an sBS will be in demand in an upcoming time period is a challenging task, since the user traffic pattern is dependent on the user context comprised of a large variety of factors such as the number, age, and interests of the users in the service range. More- over, as an sBS constantly refreshes its cache over time in order to include the emerging popular contents, the predictor needs to be able to adapt itself to new cached contents and user contexts. However, aggressive network densification in next generation cellular networks is ac- companied by an increase of the system energy consumption and calls for more advanced power management techniques in base stations. On the other hand, in order to effectively alleviate the traffic directed to macro cells and backhaul networks, deployed small cells should aim at improving the cache hit rate of their cache entities. Observing that content popularity would be largely influenced by the preference of connected users. Some relevant characteristics may consist of age, gender, occupation and the type of user equipment. This kind of information is referred as context information, which can be collected from subscriptions from content provider. Caching policy wisely learning context information would improve cache hit and therefore reduce the burden on core networks. 21 Many studies have been conducted in the area of green cache-enabled base station under a heterogeneous cellular network (HetNet) [72, 83]. For instances, M¨ uller et al. [62] present an algorithm that can learn the time-varying popularity profile of contents over time and update the cached content regularly and observe the demands. J. Liu [57] et al. study a belief propagation based distributed algorithm to solve the cache placement problem. Parallel computations can be performed by individual base stations and only few messages are exchanged between neighboring base stations. D. Liu [55] et al. investigate the energy efficiency of downlink networks of cache en- abled base stations. The consumption of transmission, circuits, backhauling and caching at Base stations are taken into account to derive the expression of energy efficiency. They conclude that caching at pico cells can provide higher energy efficiency than caching at macro base stations. K. Poularakis[68] et al. study how to jointly control the content caching and on/off switching for base stations under heterogeneous cellular networks. A energy-minimizing problem is formulated and solved using approximation framework. Numerical results show that the joint control can perform better than performing caching and dynamic on/off in a disjoint manner. E. Bas ¸tu [13] et al. model the network system with cache-enabled base stations in a stochastic way and derive expressions of outage probability and delivery rate based on variables, such as signal-to-interference-plus-noise ratio (SINR), small cell density and storage size. E. Oh [66] et al. present a implementable switching-on/off base energy saving al- gorithm to reduce the energy consumption in wireless cellular network. The online and 22 distributed algorithm shows a comparable performance to the exhaustive search optimal solution. In summary, previous work on green cache-enabled cellular networks mainly address two issues, i). design a proper caching policy to regularly update cached contents in order to improve hit rates. ii). dynamically switch base stations between on and off state to reduce power consumption. However, all such prior work fails to connect the dynamic user context with the energy consumption of the network. In this chapter, we propose a framework that utilizes the user information collected by an sBS including the gender and age of a user to estimate the energy efficiency benefit of switching an sBS to the sleep mode. To account for the uncertainty of user request pattern in the future, the control problem is formulated as a contextual multi-armed bandit problem. A variational inference based Bayesian neural network (BNN) is used as the solution, which outperforms some well-known algorithms. 2.2 System Model We consider a two-tier cache-enabled HetNet as shown in Fig. 2.1, where each tier- 1 MBS is associated with a number of tier-2 sBSs within its coverage and each sBS is associated with a set of connected user equipments (UEs), e.g. smartphones, laptop computers, tablet PCs, etc. The MBS is also connected to a hub in the core network via a backhaul link to access contents on the Internet. sBSs are usually deployed to enhance capacity in locations of high demand under the coverage of MBSs. Therefore, sBSs don’t contribute to the coverage of MBSs. The spatial distribution of base stations 23 Figure 2.1: Graphical illustration of a two-tier cache-enabled HetNet. follows homogeneous Poisson Point Process (PPP) [7] based on the stochastic geometry approach. The corresponding density of MBSs and sBSs are M and S respectively. In practice, while it is possible that a UE is in the coverage of multiple sBSs, it will only connect to the sBS with the strongest signal strength. An sBS can choose to either operate in an “active mode” in which it can serve incoming user requests or enter a “sleep mode” in which it does not serve or monitor any user requests. If an sBS is in the sleep mode in order to achieve higher energy efficiency, all requests from the UEs connected to the sBS will be served by the MBS. Because the interest of this chapter is the distributed control of sBSs, we will only focus on one sBS along with its connected UEs and associated MBS unless otherwise noted. 24 2.2.1 Channel Model LetS andU denote the set of sBSs and UEs, respectively. For an interested sBS S2S , the signal-to-interference-plus-noise ratio (SINR) when data is transmitted from the sBS to a connected UEu2U, denoted bySINR S (u), is defined as SINR S (u) = P t S g S (u) I S (u) + 2 ; (2.1) whereP t S is the transmission power of the sBS,g S (u) is the channel gain between the sBS and UEu,I S (u) is the interference power received by UEu from other sBSs, and 2 is the noise power. I S (u) can be calculated asI S (u) = P S 0 2SnS P t S 0g S 0(u). The channel gain,g S (u), can be further expanded as g S (u) =h S (u)L S (d S (u)); (2.2) whered S (u) is the distance between the sBS and UEu,L S () is the log-distance path loss function, andh S (u) accounts for the channel fading. Similarly, when data is transmitted from the MBS to a UEu, the SINR, denoted bySINR M (u), can be defined as SINR M (u) = P t M g M (u) I M (u) + 2 ; (2.3) whereP t M is the transmission power of the MBS,I M (u) is the interference power from other MBSs, andg M (u) is the channel gain between the MBS and UEu which can be expressed the same way as in Eqn. (2.2). 25 The channel capacity, i.e. the maximum data throughput in a channel, denoted by TP, can be calculated based on the Shannon-Hartley theorem as follows TP =BW log 2 (1 +SINR); (2.4) where BW is the effective bandwidth of the channel, and SINR is calculated using Eqn. (2.1) or (2.3). BW S and BW M are denoted as the effective bandwidth of sBS and MBS, respectively. In software-defined networks (SDNs),BW M can be modeled as BW M = (1O M )BW , withO M is the overhead cost of MBS andBW is the system bandwidth [90]. For sBSs, BW S W can be achieved with the assistance of MBS. For the backhaul link between an MBS and the core network, the channel capacity is assumed to be constant. 2.2.2 Service Model Fig. 2.2 shows the service model. It is assumed that both MBSs and sBSs are equipped with caches (which are usually solid state drives) with finite capacities. Gener- ally speaking, an MBS has a larger cache than an sBS. When a UE initiates a request for a file, the request will first be served by its associated sBS if it is active. If the requested file is present in the cache of the sBS, it will be transferred via the downlink of the sBS to the UE. Otherwise, the request will be redirected to the associated MBS. If the requested file can be found in the cache of the MBS, then the cached copy will be transferred from the MBS. If the requested file can be served from neither of the cache in the sBS and the MBS, the MBS is responsible for downloading it from the core network via the backhaul link and relaying it to the UE. 26 Figure 2.2: Service model of the two-tier cache-enabled HetNet. In addition, the sBS will routinely attempt to acquire the information of the connected UE such as the type and the location of the equipment. Please note that the sBS may not have access to some context information due to privacy reasons. However, our proposed framework doesn’t depend on any specific context information and it can utilize the in- formation that can be collected from UEs. 2.2.3 Cache Model The purpose of caching popular contents in the base stations is to reduce the traffic load of MBSs. It is critical for sBSs to store the most frequently requested contents locally in caches to possibly maximize the cache hit. However, learning the content popularity and deciding which contents to be stored in caches are difficult considering the huge amount of contents. Especially, the amount of contents is growing exponentially 27 and it is impossible to cache all popular contents. It is hence crucial to decide which contents to cache taking popularity into account. We consider a practical situation that each sBS and MBS is limited in the cache size, which is in terms of how many files can be stored. Many policies for refreshing caches have been widely studied. Least Recently Used (LRU): the algorithm discards the least recently used items first. It requires keep tracking of what was used and when and is based on the observation that data that have been frequently used recently will probably be used again in the future. LRU doesn’t consider the size of files, which could lead to differences in storing them into local caches limited by size. Least-Frequently Used (LFU): the algorithms counts how often an item is needed. Those that are used least often are discarded first. The drawback of LFU is that some items that heavily requested in the past may not be requested any more in the future, and they are not evicted in time. Periodical Update (PU): Considering frequently downloading contents from back- haul networks would cause overhead for base stations themselves, base stations only update their caches after each period of time based on statistics. For example, assume thatP S c andP M c are the percentage of total contents files in sBS and MBS that will be replaced based on the frequency of requests after every timeT S c and T M c for sBSs and MBS, respectively. Please note that our proposed framework doesn’t rely on any specific cache replacing policy and it can adapt to rapid cache updating. 28 2.2.4 Power Consumption Model We adopt a power consumption model for base stations similar to that in reference [89], in which the total power consumption of a base station is comprised of static power, transmission power, and data access power. Note that since the total amount of trans- mitted data is dominated by the requested content, we ignore the power consumption incurred from sending requests or other utility packets. As discussed in Section 2.2.2, an sBS will only transmit data if there is a cache hit for the requested file. Therefore, the power consumption of the sBS, denoted by P tot S , can be calculated as P tot S =x S P 0 S + 1 S P tx;i S +w CA TP S + (1x S )P sleep S ; (2.5) where P 0 S is the static power consumption of an sBS, P tx;i S is the instantaneous trans- mission power of the sBS, S is the transmission efficiency that accounts for the power overhead for transmission such as channel encoding, TP S is the transmission rate that can be achieved from the transmission power, w CA is the energy efficiency (in J/bit or W/bps) of a cache read operation,P sleep S is the power consumption of the sBS in the sleep mode, andx S is binary variable that indicates whether the sBS is active or not. If a request is redirected to the MBS either due to a cache miss on the sBS or because the sBS is in the sleep mode, the MBS is responsible for issuing the transmission whether the requested file is in its cache or not. Therefore, the total power consumption of an MBS, denoted byP tot M , can be calculated as P tot M =P 0 M + 1 M P tx;i M +w CA TP hit M +w BH TP miss M ; (2.6) 29 where P 0 M , P tx;i M , and M are the static power, transmission power, and transmission efficiency for an MBS, respectively. TP hit M is the data throughput of files available in the MBS’s cache, TP miss M is the data throughput of files pulled via the backhaul link, and w BH is the energy efficiency for accessing the backhaul link. Furthermore, if a file has to be downloaded from the core network to the MBS, there is an increase in the power consumption of the core network, denoted by P C , which can be estimated as P C =w CN TP miss M ; (2.7) wherew CN is the energy efficiency for data transmission in the core network. 2.3 Problem Formulation and Solution Method In this work, a slotted time model is adopted for decision making in which the op- eration mode for an sBS is provided for time intervals of equal and constant duration. Since the file requested from a cellular network is relatively small, we assume that all file transmission can be finished within the same time slot in which the request is initiated. 2.3.1 Problem Description The objective of the proposed problem is to minimize the total energy consumption of the network while serving all user requests by controlling the operation modes of the 30 sBSs. In a given time slott, if the amount of requested data (in bits) from UEs connected to the sBS under study, denoted byC tot t , can be expressed as C tot [t] =C S [t] +C M [t] +C BH [t]; (2.8) where C S [t] is the amount of requested data that is cached by the sBS, C M [t] is the amount of requested data that is not cached by the sBS but can be found in MBS’s cache, andC BH [t] is the amount of requested data that is not present in either sBS or MBS and has to be downloaded from the backhaul link. When the sBS under study is in the active mode during time slott, the total energy consumed to serve requests from its connected users, denoted by E ON [t], consists of three components, i.e. the energy consumption of the sBS under study itself, the energy consumed by the MBS to aid the sBS, and the increased amount of energy consumption in the core network to transfer data needed by the sBS. As mentioned above, the spatial densities of sBSs and MBSs are denoted by S and M , respectively, E ON [t] can be derived as follows based on the power consumption model in Eqns. (2.5) – (2.7). E ON [t] = P 0 S + M S P 0 M + 1 S P tx S T S [t] + 1 M P tx M T M [t] +w CA (C S [t] +C M [t]) + (w BH +w CN )C BH [t]; (2.9) whereP tx S andP tx M are the maximum transmission power of an sBS and an MBS, respec- tively, whereasT S [t] andT M [t] are the amount of time needed to transmit files from the sBS and the MBS to the UEs in the coverage of the sBS under study, respectively, which can be calculated using the throughputs derived from Eqns. (2.1), (2.3), and (2.4). Please note that the ratio M S scales the effect of the MB and other sBSs excepting the studied 31 one. The amount of requested files is also scaled explicitly by considering requests are directed to all sBSs and one MB. On the other hand, if the sBS is in the sleep mode during time slot t, the total en- ergy consumed to serve requests from its connected UEs, denoted by E OFF [t], can be estimated as E OFF [t] = P sleep S + M S P 0 M + 1 M P tx M T 0 M [t] +w CA C M [t] + (w BH +w CN ) (C BH [t] +C S [t]); (2.10) whereT 0 M [t] is the amount of time needed to transmit files from the MBS to the users in the coverage of the sBS under study which can be calculated in the same way asT M [t] in Eqn. (2.9). Please note that we assume the cached contents in the sBS and the MBS do not overlap for the convenience of expression in Eqn. (2.10), but it can be trivially extended to the case in which there is overlapping between caches. If we consider a total of T consecutive time slots with t = 0; 1;:::;T 1, the objective function, denoted byE obj , can be calculated as E obj = T1 X t=0 x[t]E ON [t] + (1x[t])E OFF [t] ; (2.11) whereE ON [t] andE ON [t] are calculated as in Eqns. (2.9) and (2.10), respectively, and x[t] is a binary decision variable which is set to 1 when the sBS under study is in the active mode in time slott and set to 0 when the sBS is in the sleep mode. 32 2.3.2 Control Problem Formulation We first start with an offline control problem formulation in which the operation mode of the sBS under study is determined for all time slots within a control horizon, provided all information of user connections and download requests, e.g. URL and size of the requested files. In such an offline setting, the optimal value of decision variablesx[t]’s can be determined using a simple solution method. Since the values ofC S [t]’s,C M [t]’s, andC BH [t]’s in Eqn. (2.8) can be derived from the size of each requested file, one can know the exact values ofE ON [t]’s andE OFF [t]’s. Therefore, for time slott, the value ofx[t] can be set to 1 ifE ON [t]<E OFF [t] and set to 0 otherwise. Although the offline solution yields optimal results, it requires the request pattern to be known before control decisions are made, which is not realistic. In a more realistic setting for online control, the number and type of files that will be requested in future time slots cannot be known a priori, which means that values of C S [t],C M [t], andC BH [t] that are required to calculateE ON [t] andE OFF [t] are unknown at the time of finding the optimalx[t]. However, as discussed in Section 2.2.2, user profiles and information contexts in- cluding the equipment type, gender, age and occupation of each connected user can be obtained and updated by sBSs. In the online control problem, we assume that an sBS will collect the user profile/context, specified by a vector v[t], at the beginning of time slott, after which the operation mode of the sBS will be determined as a function of v[t], while the actual energy consumption over a time slot can only be obtained at the end of the current time slot. 33 The aforementioned problem description fits in the framework of a contextual multi- armed bandit problem [62, 60] in which an agent needs to sequentially select from a set of actions to take in each step based on some observations in order to maximize his/her reward which is non-deterministic and unknown a-priori. For our control problem, the agent is the sBS under study, the action set consists of two different operation modes, and the reward in time slott, denoted byr[t], can be defined as: r[t] = 8 > < > : E OFF [t]E ON [t]; x[t] = 1 0; x[t] = 0 (2.12) Notice thatr[t] is set to 0 when the sBS is in the sleep mode because it is not monitoring the user traffic and cannot estimate the energy saving. The goal of the control problem is then translated into finding the optimal action in time slott, denoted byx [t], such that x [t] = arg max x[t] E [r[t]j[t];x[t]] (2.13) where[t] = (v[0];::: v[t];x[0]:::x[t 1];r[0]:::r[t 1]) is the history of the system at the beginning of time slott. Since the MBS is always on and its energy consumption overhead for helping sBSs is captured by energy model, we assume that the throughput related QoS constraints for UEs can be satisfied by transmitting data from the MBS. In general, a contextual multi-armed bandit problem is hard to solve [20] due to the partial observability of the system and the non-deterministic relationship between the action taken and the reward. Thompson sampling[80] is a popular approach for solving the contextual multi-armed bandit problem by picking actions to balance exploration and exploitation. Generally, given the input context v(t), a set of actionsx(t)2X , rewards 34 Algorithm 1: Thompson Sampling 1 Posit a-prior distributionsp(z), z are parameters of the latent model. 2 for each time slott do 3 Sample new sets of z from the posterior distributionp(zj(t)) 4 Receive the context v[t]. 5 Perform the action with highest expected reward, namely x [t] = arg max x[t] E(r(t)jv[t];x(t); z). 6 Update the model. 7 end r(t) inR and the history of system[t], Thompson sampling is performed following the steps shown in Algorithm 1 . Bayesian treatment is necessary in Thompson sampling. According to Bayes’ rule, the posterior distribution can be calculated as p(zj(t)) = p((t)jz)p(z) p((t)) , where thep((t)jz) is the likelihood distribution,p(z) is a-prior distribution andp((t)) is the evidence. Please note that the latent model is not trivial to construct and requires the knowledge of system. In our work, we adapt Thompson sampling to a learning agent by using a model- free Bayesian neural network (BNN) to approximate the latent model while preserving the Bayesian treatment. In this way, it relaxes the requirement of constructing a precise model and implicitly handles uncertainties in the observation properly. 2.3.3 Solution Method to the Online Control Problem 2.3.3.1 Bayesian Neural Network Bayesian neural network is a neural network with a-prior distributions instead of single fixed point estimate on its latent variables [64]. The BNN can be seen as an en- semble of a large number of neural networks with introduced variabilities on weights and biases. As a result, BNNs are robust to disturbances in the learning especially when 35 the training set is noisy or incomplete. In addition, the over-fitting can be potentially alleviated because the implicitly introduced regularizations, which prevents BNN from over-exploiting the currently obtained observations (training sets) in the contextual ban- dit problem setting [17]. In BNNs, latent variables z consist of weights W and biases b. In analogy with Thompson sampling method, we posit a-prior distributions p(z) on latent variables z, which denotes the assumption before observing the system. Here, observations are the history of system(t). The likelihood distributionp((t)jz) in BNN can be computed as follows: p((t)jz) =r[t] log o(x [t]; v[t]; z) ; (2.14) because the BNN has been trained using historical observations. o() is the value at the output node of BNN, which corresponds to the chosen actionx [t], input context v[t] and latent variables z. The actionx [t] is chosen to be the one producing the largest value at the output node. Please note that the likelihood here is not necessarily a probability within the region [0; 1]. As we know from Bayes’ rule, in order to calculate the posterior distributionp(zj(t)), the denominatorp((t)) has to be computed first. However, calculating the denomina- torp((t)) requires the integral of sum over all possible latent variables, i.e.,p((t)) = R z p((t)jz)p(z)dz. For most of cases of interest, this integral is intractable, therefore approaches are needed to approximately evaluate the posterior probability [39]. One of the most commonly used method is Markov chain Monte Carlo (MCMC), which keeps 36 sampling a Markov chain by traversing the high probability area. However, MCMC is computationally very expensive and lack of a clear stopping criterion [49]. 2.3.3.2 Variational Inference In order to solve the intractable issue of Bayesian learning in neural networks by an efficient way, variational inference [39, 49] has been proposed as an alternative to MCMC. Different from MCMC, variational inference solves the optimization problem, has a better converge rate and is scalable to large problems. Variational inference assumes thatp(zj[t]) can be approximated using a variational posteriorq(z;) where is an unknown parameter to be found and the variational poste- rior distribution of latent variables can be adaptively updated by adjusting the value of. The core idea of variational inference involves two steps. 1). posit variational posterior distributionsq(zj) over the latent variables z. 2).useq(zj) to approximate the posterior distribution p(zj(t)) by optimizing over its parameters to minimize the divergence metric (Dvgnc), = arg min Dvgnc(q(z;);p(zj(t))). For the divergence measure- ment, when the Kullback-Leibler (KL) divergence is used, the optimization problem becomes as follows: = arg min KL(q(z;)jjp(zj(t))) = arg min Z +1 1 q(z;) log( q(z;) p(zj(t)) )dz = arg min E q(z;) (logq(z;) logp(zj(t))): (2.15) 37 Please note that Eqn. (2.15) is still intractable due to the dependence onp(zj(t)). Con- sidering the property that can be derived from Bayes’ rule as follows: logp((t)) =KL(q(z;)jjp(zj(t))) +E q(z;) [logp((t); z) logq(z;); where the logp((t)) is a constant with respect to the variational parameters. Therefore, we can minimize KL(q(z;)jj p(zj(t))) by instead of minimizing the cost function F () below: F () =E q(z;) [logq(z;) logp((t); z)] =E q(z;) [logq(z;) log(p((t)jz)p(z))]: (2.16) The expectation value can be estimated by Monte Carlo integration which samples latent variables z from the distributionq(z;). Then the estimated cost functionF () 0 can be calculated as follows: F () 0 = 1 M s Ms X s=1 [logq(z s ;) log(p((t)jz s )p(z s ))]: (2.17) Here,M s is the cardinality of Monte Carlo integration sampling steps. z s is the sampled values of latent variables z from the distributionq(z;) at the sampling steps. TheM s is usually a small number compared to the sampling count in MCMC. Eqn. (2.17) is tractable because each term can be calculated. 38 Figure 2.3: Illustration of the proposed BNN. 2.3.3.3 Proposed Context-Driven Online Power Mangement Framework To solve the aforementioned contextual multi-armed bandit problem, we propose to train a BNN using the variational inference method. As shown in Fig. 2.3, a BNN has a set of latent variables, z =fz 1 ;z 2 ;:::g, which includes edge weightsW l ij from thei-th neuron at the layerl to thej-th neuron at the layerl+1 and biases b l at layerl. The input nodes of the network correspond to the elements of v[t], which are the input context of all connected users at the studied sBS. More concretely, each kind of user profiles is categorized and encoded into one-hot encoding. Each bit in the one-hot encoding corresponds to one neuron at the input layer. Then, each connected user is binned into input nodes according to their categories. Whereas the two output nodes produce the probabilities that either of the two operation modes yields lower energy consumption, i.e.p(x[t] = 0jv[t]; z) andp(x[t] = 1jv[t]; z), respectively. 39 As discussed before, different from other well-known neural networks, such as arti- ficial neural networks (ANNs) in which z take deterministic values, a BNN assumes that latent variables z have a-prior distributions p(z) and posterior distributions p(zj[t]). While the a-prior distribution is independent of the history of system and remains un- changed, the variational posterior distribution is updated at the end of each time slot after the action is taken and the reward is evaluated. In our implementation, both a-prior distributionsp(z) and variational posterior distributionsq(z;) are assumed to be Gaus- sian distributions. For q(z;), denotes the mean value and standard deviation of the Gaussian distribution. The proposed solution framework is summarized in Algorithm 2. After initializ- ing the prior and variational posterior distribution of z, the BNN will iteratively select the operation mode of the sBS at the beginning of each time slot and update the varia- tional posterior distributions at the end of each time slot using forward propagation and backward propagation operations [71]. In time slott, the proposed solution proceeds as follows. First, at the beginning of the time slot, given the value of v[t] as inputs to the BNN, the value ofx[t] is determined by forward propagating through the network with respect to a sampling of z and choosing the value corresponding to the output node with a larger output probability. After that, the operation mode is selected according to the x[t] value found. At the end of time slott, the value of parameter is updated, which in turn modify the distribution for sampling latent variables in z. We update using a stochastic gradient descent approach which tries to adjust the value of in the opposite 40 direction of the gradient of a loss functionL [t] through back propagation. Base on the Eqn. (2.14) and Eqn. (2.17),L [t] can be computed as follows L [t] = 1 M s Ms X i=1 (logq(z i ;)r[t] logo i logp(z i )); (2.18) where z 1 ;:::; z Ms are M s samples of z according to probability distribution function q(z;),o 1 ;:::;o Ms are the output values on the output node corresponding to the selected x[t] in the M s sampled networks, and r[t] is the reward defined in Eqn. (2.12). The updated will then be used in time slot (t + 1). Algorithm 2: Variational Inference BNN Agent 1 Posit prior distributionsp(z). 2 Posit variational posterior distributionsq(z;). 3 for each time slott do 4 Sample z fromq(z;) 5 Receive the context v[t]. 6 Perform forward propagation for input v[t]. Pick action x [t] = arg max x[t] p(x[t]jv[t]; z). 7 Set operation mode according tox [t] and evaluate rewardr[t] as in Eqn. (2.12). 8 Use Monte Carlo integration to calculate the cost functionL [t] in Eqn. (2.18). 9 Perform backward propagation of the gradient ofL [t] with respect to. 10 Update according to L rL [t] where L is the learning rate. 11 end 2.4 Experimental Results Similar to references [63, 12, 52], we use the MovieLens 1M dataset [41] to char- acterize the pattern of mobile users’ download requests. The MovieLens 1M dataset 41 contains about 1 million ratings of 3952 movies from 6040 users within the year 2000- 2003. Each rating consists the user ID, movie ID and a timestamp along with the user’s gender (2 categories), age (7 categories) and occupation (20 categories). Therefore, we treat each movie as a file entry stored in caches in MBSs and sBSs and each rating as a download request for a file from the UE with a profile described in the dataset. This approach is reasonable since each rating process is usually done after downloading and watching the movie. Our assumption is that users with similar profiles would request similar files. Time slots with 15 minutes each are considered to cover the first 365 days in the dataset. Specifications of the two-tier cellular network are given in Table I.BW is the nominal channel bandwidth in the system. In practice, the effective channel band- width of an sBS can be approximated asBW while the effective bandwidth of an MBS is calculated as (1O M )BW whereO M is the portion of overhead [90]. To estimate the SINR from a user to a base station, the path-loss for an sBS and an MBS are 30:6 + 36:7 log 10 (d) and 35:3 + 37:6 log 10 (d) indB, respectively [56], whered is the distance specified in kilometers from a UE to a base station. The users’ positions are uniformly generated within the service range of an sBS. The interference is estimated based on the signal strength from three nearest base stations. The noise power is set to 2 =95dBm [56]. We adopt a least frequently used replacement strategy in caches in sBSs and MBSs which replacesp S u andp M u percentage of entries every day. N M andN S are the maximum number of files an sBS and an MBS can cache, respectively. Each file is assumed to have a size of 6 MB. The user context, v[t], is comprised of the count of connected users that fall in each category of gender, age, and occupation, resulting in a BNN with 29 input nodes. We implement two hidden layers in the BNN with 200 neurons in each of the hidden layer. 42 Table 2.2: Simulation Parameters. Parameter Value Parameter Value Channel BW 10MHz O M 28.5%[90] M 0:001=m 2 S 0:1=m 2 Power M 0.311 [56] S 0.066[56] P tx M 46dBm [11] P 0 M 724:6W [56] P tx S 30dBm [11] P 0 S 10:16W [56] w CA 6:25pW=bps [56] w BH 0:5W=bps [56] P sleep S 3:87W [56] w CN 70nW=bps [5] Cache P M c 2% P S c 10% T M c 24hr T S c 24hr N M 1000 N S 100 The two output nodes of the BNN correspond to active mode and sleep mode, respec- tively. Rectified linear units are used as the activation function for the input layer and hidden layers while softmax activation is used in the output layer. Gaussian distributions with a mean value of 0 and a standard deviation of 10 3 are assumed to be the prior distributions of all edge weights. The variational posterior distributions on weights and biases are also assumed to be Gaussian distributions parameterized by the mean value and the standard deviation, which are updated in each time slot.M s in Eqn. (2.18) is set to 5 to balance between exploration and exploitation. Stochastic gradient descent is used with mini-batches of 10 data points each with a learning rate of 0.001. The cumulative energy consumption as defined in Eqn. (2.11) and cumulative regret associate with one sBS in one year is shown in Fig. 2.4 and Fig. 2.5, respectively. The regret is defined as the difference in cumulative energy consumption between the offline optimal solution as discussed in Section 2.3.2 and the online solutions shown in the figure. Our proposed variational inference BNN agent is compared with four baseline algorithms, namely, (1) always-active, (2) 10%-greedy-neural, (3) 30%-greedy-neural, and (4) 10%-greedy-simple. In the “alway-active” scheme, the sBS is never switched 43 Figure 2.4: Cumulative energy consumption to serve requests from connected users of the sBS under study. Figure 2.5: Cumulative regret in terms of energy consumption. 44 Figure 2.6: The relationship between total energy consumption and sBS/MBS density ratio S M . to the sleep mode. The three other baselines adopt the well-known -greedy learning strategy [78] in which the agent takes a random action with a probability in each step and takes the best action generated by the algorithm with a probability of (1). In “10%-greedy-neural” and “30%-greedy-neural”, the same neural network structure as the proposed BNN is used without any Bayesian assumption. To avoid over-exploration at later stages, the is set to decrease by half every 100 days. “10%-greedy-simple” is a simple baseline that does not consider any user context but only choose the best action according to the average reward in the history. As can be seen from Fig. 2.4, the energy consumption by using the proposed agent is 606 MJ, achieving 18:7% total energy reduction compared to the “always active” baseline (745 MJ). Furthermore, from Fig. 2.5, it can be observed that the regret of the proposed agent is nearly constant in later time slots while all four baselines see a linear increase of regret over time. Fig. 2.6 45 Figure 2.7: Number of cache hits on the sBS if active v.s. actual decision made by the sBS for the first 5000 time slots. shows the relationship between total energy consumption and sBS/MBS density ratio S M . Total energy consumption reduction can be achieved at 37:9%, 45:3% and 46:9% when the density ratio increases from 100 to 400, 800 and 1000, respectively. It can be seen that the proposed BNN agent can reduce the total energy consumption significantly in a scenario with a high density deployment of sBSs. The relationship between the actual decision of the sBS and the number of cache hits in the sBS if it is active is shown in Fig. 2.7 for the first 5000 time slots. As can be seen from the figure, after an initial “warm-up” period, the sBS will only be active when there are a relatively large number of cache hits, which matches our expectation, suggesting that our BNN agent can accurately interpret the user context and predict whether an active sBS will be energy efficient. 46 2.5 Conclusion In summary, we investigate how to improve the energy efficiency in a cache-enabled two-tier HetNet. An sBS is controlled to switch between the “active mode” and the “sleep mode” to reduce energy consumption driven by connected users’ contexts. The online en- ergy minimization problem is carefully formulated into a contextual multi-armed bandit problem, and a variational inference BNN agent is proposed. With the proposed frame- work, an sBS can achieve a sub-linear increase of regret in energy consumption over time and reduce energy consumption by 46.9% for high density deployment scenario in one year comparing with the “alway-active” baseline. 47 Chapter 3 Power Management of Cache-enabled Cooperative Base Stations Towards Zero Grid Energy With the increasing demand of high speed mobile data transmission, densely de- ployed small cell base stations capable of caching popular contents have recently emerged as a promising technique to improve the quality of service for mobile users. In this chap- ter, we investigate the cooperative transmission and power management problem for a set of “off-grid” base stations in a cellular network hierarchy that are powered solely by on-site renewable energy sources. The network throughput maximization problem is mathematically formulated as a mixed-integer non-linear programming problem. In the proposed formulation, a base station can adjust its transmission power in a coordinated multipoint communication scheme and/or switch to a “sleep mode” for energy saving. Based on the Lyapunov optimization theory, an efficient near-optimal solution method is proposed with provable bound of the optimality gap. Experimental results on a realistic setup show that the proposed algorithm can achieve up to 2.96x download throughput per user compared to some baseline algorithms. 48 3.1 Background and Prior Work While network densification helps improve the quality-of-service (QoS), it is asso- ciated with a significant power consumption overhead because of the large number of deployed (sBSs). According to data in references [11, 89], there can be 100 times more sBSs than macro base stations (MBSs) in a cellular network in practice, resulting in twice as much the total peak power consumption. In terms of environmental impacts, it has been reported that information and communication technology (ICT) already con- tributes around 2% of the global carbon dioxide emission and is expected to increase rapidly in the future. Moreover, 10% of the world’s electric energy is consumed by the ICT infrastructure and base stations contribute to around 60% of the power consumption in cellular networks. From the perspective of energy efficiency, base stations need to be self-organizing, energy-efficient, and environmental-friendly. One promising solution, as already adopted by many telecom operators around the world [2], is to use off-grid base stations that are powered solely by renewable energy generated from energy harvesting systems (e.g. photovoltaic cells) [22]. On the other hand, power management policies can be designed to switch a base sta- tion into a “sleep mode” in order to save energy when there is no user to be serviced. Generally speaking, power management for an off-grid base station is a non-trivial prob- lem because of the sporadic and intermittent nature of commonly used renewable en- ergy source and the limited energy storage capacity in a base station. Without a reliable prediction of energy generation and user request profile, over-conservatively reducing the transmission power of base stations can significantly affect the network throughput, 49 while activating base stations too often will quickly drain the energy storage and may cause failure in service later on. Another potential issue with dense sBS deployment is the inter-cell interference, which is usually addressed using cooperative transmission schemes such as coordinated multipoint (CoMP) communication [73, 46]. CoMP allows a user equipment (UE) to establish connections with several base stations at the same time while the involved base stations can be viewed as a virtual MIMO array that schedule and process the transmitted signals in a coordinated manner [61]. A large body of prior literature aims at solving some aspects of the aforementioned problems of ultra-dense sBSs. In work [26], Chiang et al. consider a cache-enabled base station switch-off scheme for saving energy under the cooperation transmission powered by solar energy. A centralized local search is implemented to reduce the total energy consumption. In work [28], Cili et al. study the combination of CoMP transmission and cell switch-off technique to achieve more energy efficiency. Chamola et al. [23] investigate a downlink power management policy for cellular base stations with hybrid power supplies. The network delay is considered for the energy minimization problem. However, there is no prior research on a complete solution to the power management of cooperative off-grid base stations. Besides, algorithms proposed by prior work usually have high complexity and become intractable when the problem size is large. In this chapter, we propose an online power management framework for a cellular network consisting of cooperative off-grid base stations powered by energy harvesting modules. Rechargeable battery packs are deployed in base stations to deal with the in- termittent energy supply so that harvested energy can be used at a later time. Each base station is equipped with a content cache to directly transmit cached contents to requested 50 users without downloading them from the core network. Based on Lyapunov optimiza- tion theory, the network throughput maximization problem with power and energy con- straints is formulated as a mixed-integer non-linear programming problem. To develop an efficient algorithm, the original problem is decomposed into a number of subprob- lems, each of which can be solved using standard convex optimization techniques with a provable optimality gap. The effectiveness of the proposed control framework in practice is demonstrated by experimental results based on realistic user request traces. The rest of this chapter is structured as follows. In Section 3.2, the system model of a cache-enabled two-tier cellular network is introduced. In Section 3.3 and Section 3.4, we propose the problem formulation and solution method. Experimental results are presented in Section 3.5. Section 3.6 concludes the chapter. 3.2 System Model We consider a two-tier cache-enabled cooperative HetNet as shown in Fig. 3.1, where a tier-1 MBS covers a set of UEsM =f1; ;j; ;Mg and is associated with a number of tier-2 sBSsN S =f1; ;i; ;Ng. We useN =N S [f0g to represent the MBS and all its associated sBSs. In addition, the MBS is also connected to a hub in the core network via a backhaul link to access contents on the Internet. We consider the case in which sBSs are deployed to enhance capacity in locations with high demand under the coverage of MBSs and a UE is in the coverage of at least one MBS. A CoMP communication scheme is used in which an MBS and its associated sBSs are capable of forming a cooperative region and jointly transmitting data to a UE, thus increasing the effective downloading throughput. Coordinated transmission between multiple MBSs is 51 Figure 3.1: Graphical illustration of a two-tier cache-enabled HetNet. not considered and a UE will only connect to the MBS with the highest signal strength if it is in multiple MBS’s coverage. As a result, without loss of generality, we can focus on one cooperative region consisting of one MBS (which will be referred to as “the MBS”) and its associated sBSs and connected UEs. All MBSs and sBSs are considered to be off-grid, i.e. the only power source of a base station is its energy harvesting module. Rechargeable batteries are used to store excessive harvested energy for later use. A base station can choose to either operate in an “active mode” in which it can perform data transmission or enter a “sleep mode” in which it does not monitor user requests or transmit data for energy saving purpose. 52 3.2.1 Service Model In our proposed framework, a slotted time model is adopted in which each time slot has a duration of. Each UE may request for at most one file (e.g. videos, texts, images, etc.) in each time slot, which is considered to be known at the beginning of the time slot. Please note that this model can be easily extended to the case where a UE requests mul- tiple files at the same time by creating “dummy” UEs at the same location. In addition, since caching benefits smaller files such as news or weather information much more than large files such as movies [70], we focus on small files that can be transmitted within the same time slot that they are requested. Both MBSs and sBSs are equipped with content caches capable of storing files that may be requested by UEs. Due to the complex dynamics of the user request pattern and the finite cache space, one cannot guarantee that all requested files can be found in the cache of an sBS or MBS. As a result, files that are missing in the base stations’ caches have to be downloaded from the core network. A file request from a UE will be sent to the MBS and all associated sBSs that are in active mode. If an sBS has the requested file in its cache, it can choose to join the CoMP scheme with a specific transmission power. On the other hand, the MBS can choose to join the CoMP scheme by either transmitting the requested file from its cache or relaying the requested file from the core network. Please note that since the design of caching policy in base stations is not the focus of our work, we assume that a base station is aware of the applied caching policy and knows its cached content. A set of binary variablesF ij [t]’s are used to represent the caching status of requested files. F ij [t] is set to 1 if the file requested by UEj in time slott is present in the cache 53 of base station ofi and set to 0 otherwise. The size of the file requested by UEj in time slott is denoted byO j [t]. In the case that UEj does not request any file in time slott, one can simply setO j [t] to 0. 3.2.2 Channel Model When downloading files in a CoMP scheme, the signal-to-interference-plus-noise ratio (SINR) for UEj2M, denoted bySINR j , is defined as SINR j [t] = P i2N P tx ij [t]g ij P int j + 2 (3.1) whereP tx ij [t] is the transmission power that base stationi assigns to UEj.P int j is the ag- gregate interference seen by UEj from outside the cooperative region,g ij is the channel gain from base stationi to UEj, and 2 is the noise spectral density. The channel gain, g ij , can be further expanded as g ij =h ij L (d ij ) (3.2) where d ij is the distance between base station i and UE j, L() is the log-distance path loss function, andh ij accounts for the channel fading. Furthermore, according to Shannon-Hartley theorem, the data transmission rate achieved at UEj can be calculated as R ue j [t] =BW log 2 (1 +SINR j [t]) (3.3) whereBW is the effective bandwidth of the channel. 54 Figure 3.2: Energy flow in an off-grid base station. 3.2.3 Energy Consumption Model The block diagram of the power flow within an off-grid base station is shown in Fig. 3.2. In time slott, if the amount of harvested energy and the energy consumption of base stationi2N are denoted byh i [t] ande i [t], respectively, then the amount of available energy stored in the battery pack at the beginning of time slott, denoted byE i [t] will be updated as E i [t + 1] =E i [t] 1 D e i [t] + C h i [t] (3.4) 55 where C and D are the charging and discharging efficiency of the battery pack, respec- tively. If the energy capacity of the battery pack in base station i is denoted by C max i , then the dynamics ofE i [t] should satisfy 0E i [t] 1 D e i [t] + C h i [t]C max i (3.5) Generally speaking, the total power consumption of a base station is comprised of a static component (independent of the workload) and a dynamic component (depen- dent on the workload) [9]. It has been proved that the total power consumption can be approximated as a linear function of its data transmission power output [11]. We adopt a similar power model as in work [27], in which the total energy consump- tion of a base stationi2N in time slott,e i [t], is given by e i [t] = X j2M p P tx ij [t] + (1y i [t])P slp i +y i [t]P act i (3.6) where p is the slope of the load-dependent power consumption,P act i andP slp i are the static power consumption of base stationi in the active mode and sleep mode, respec- tively, andy i (t) is a binary variable indicating whether BS is operating in the active mode (= 1) or in the sleep mode (= 0) in time slott. According to the data in reference [56], the energy consumption of cache read and downloading via backhaul is usually much lower than the energy consumption of data transmission. Therefore, we choose not to introduce these terms in the energy consumption model. 56 3.3 Problem Formulation In this section, we present the formulation of the base station cooperative control (BSCC) problem. In the BSCC problem, the goal is to maximize the average download throughput of all user requests by selecting the operation mode and setting transmis- sion power of each base station. Since all base stations are off-grid, the control policy should judiciously balance the energy generation and the energy consumption to prevent frequent service outages. Based on the modeling of system components as in Section 3.2, the BSCC problem can be formally defined as follows Given: cache informationF ij [t] and channel informationg ij [t]. Find: the optimal transmission powerP tx ij [t] and the operation modey i [t]. Maximize: lim T!1 1 T T1 X t=0 X j2M R ue j [t] (3.7) Subject to: 57 0P tx ij [t]y i [t]F ij [t]P max i 8i> 0;8j (3.8) 0P tx 0j [t]y 0 [t]P max 0 8j (3.9) X j2M P tx ij [t]P tot i 8i (3.10) E SINR j [t] th 8j (3.11) 0E i [t] 1 D e i [t] + C h i [t]C max i 8i (3.12) y i [t]2f0; 1g 8i (3.13) whereR ue j [t] is the download throughput of UEj in time slott as defined in Eqn. (2.1) – (3.3),P max i is the maximum transmission power output of base stationi,C max i is the energy capacity of the battery pack in base stationi, and the amount of energy storage in base stationi, denoted byE i [t], is updated according to Eqn. (3.4). Constraint (3.8) states that an sBS can only help transmit data to UE j when it is in the active mode and the requested file is in the base station’s cache. Similarly, constraint (3.9) ensures that the MBS will transmit data only when it is in the active mode. Constraint (3.10) sets the limit of total transmission power to P tot i for each base station i. As a QoS requirement, constraint (3.11) sets the time-average minimum acceptable SINR for each UE to a threshold value th . Constraint (3.12) guarantees that a battery pack’s state of charge is always valid. Constraint (3.13) sets the domain of decision variables. The BSCC problem is hard to solve in general because of the existence of binary decision variables and the infinite horizon. Furthermore, since it is impractical to assume that the amount of harvested energy is known for all future hours, predicted values of 58 h i [t]’s are often used, which makes online solutions that can adaptively update the control policies more desirable than offline solutions. 3.4 Lyapunov Optimization Solution In order to develop an efficient online solution to the BSCC problem, we apply the Lyapunov optimization framework [65] to decouple the original problem formulated in Section 3.3 into a set of independent subproblems, each corresponding to a base station. In order to capture the QoS requirement as in constraint (3.11), a virtual queueZ j [t] is defined for UEj, which is updated at the beginning of time slot (t + 1) as follows Z j [t + 1] = [Z j [t]SINR j [t] + th ] + (3.14) On the other hand, to model the state of charge of the battery pack of base station i, another virtual queue ~ E i [t] = C max i E i [t] is introduced. Deriving from Eqn. (3.4), ~ E i [t] is updated as follows ~ E i [t + 1] = ~ E i [t] + 1 D e i [t] C h i [t] (3.15) SinceR ue j [t] as defined in Eqn. (3.3) is a concave function ofP tx ij [t]’s, Jensen’s in- equality can be used to find a lower bound ofR ue j [t] as follows R ue j [t] BW N + 1 X i2N log 2 (1 + P tx ij [t]g ij (N + 1) P int j + 2 ) (3.16) 59 For the convenience of discussion, we define a “penalty” function, denoted bypen(t), as the opposite of the lower bound found in Eqn. (3.16), i.e. pen(t) = BW N + 1 X j2M X i2N log 2 (1 + P tx ij [t]g ij (N + 1) P int j + 2 ) (3.17) A Lyapunov drift, denoted by [t] can be defined as [t] = 1 2 X j2M Z j [t + 1] 2 Z j [t] 2 + 1 2 X i2N ~ E i [t + 1] 2 ~ E i [t] 2 (3.18) The BSCC problem can be mapped into a Lyapunov drift-plus-penalty minimization problem [65] that aims at minimizing the time average of the weighted sum of the Lya- punov drift and the penalty function, i.e. lim T!1 1 T T X t=1 ([t] +Vpen[t]) (3.19) whereV is a positive constant coefficient that determines the relative importance of the penalty function and the average virtual queue length. Instead of solving the BSCC problem or the Lyapunov drift-plus-penalty problem with infinite horizons, we solve an opportunistic control (OPC) problem at the begin- ning of each time slot to only make the control decision for the next time slot, which significantly reduces the problem size. The OPC problem is formally defined as follows Find:P tx ij [t] andy i [t],8i;j 60 Minimize: Vpen[t] + X j2M Z j [t]( th SINR j [t]) + X i2N ~ E i [t] 1 D e i [t] C h i [t] (3.20) Subject to: Constraints (3.8) – (3.10), (3.12), and (3.13) It can be proved that a near-optimal solution to the BSCC problem can be obtained by solving the OPC problem at the beginning of each time slot. Theorem 1. If eachh i [t] can be viewed as an i.i.d. process over time slots, there exists a finite optimality gap between the solution achieved by the OPC problems and the optimal solution of the BSCC problem as long as the SINRs,e i [t]’s, andE i [t]’s are bounded. Proof. To simplify the expression, we introduce the following functions G 0 (P) = lim T!1 1 T T1 X t=0 X j2M R ue j [t] (3.21) G LB (P) = lim T!1 1 T T1 X t=0 pen[t] (3.22) where P = n P tx ij [t] 1 t=1 ;fy i [t]g 1 t=1 o is the series of control decisions. Note thatG 0 (P) is the objective function of the BSCC problem andG LB (P) is the average opposite value of the penalty function of the OPC problem. We denote the optimal control policies that yield the maximum value ofG 0 (P) andG LB (P) by P 1 and P 2 , respectively. The solution to the OPC problem over all time slots will be denoted by P 0 . 61 From Eqn. (2.1) – (3.3) and constraints (3.8) and (3.9), it can be derived that G LB (P)G 0 (P)c 0 ; 8P (3.23) where c 0 = max 0xSINRm log 2 (1 +x) log 2 (1 +SINR m ) SINR m x MBW (3.24) whereSINR m is the upper bound of SINR for any channel. The value ofc 0 is found by deriving the convex hull of theG LB (P) curve. It has been proved in Theorem 4.8 of reference [65] that G LB (P 0 )G LB (P 2 ) B V (3.25) whereB is a constant. Moreover, applying the bound of G LB (P) in Eqn. (3.23), we have G LB (P 2 )G LB (P 1 )G 0 (P 1 )c 0 (3.26) Combining Eqn. (3.25) and Eqn. (3.26), we can find the optimality gap as follows G LB (P 0 )G 0 (P 1 )c 0 B V (3.27) Note that even whenh i [t] is not an i.i.d. process, the OPC-based control is still effective in practice, which is demonstrated in the experimental results in Section 3.5. 62 To further reduce the complexity of the problem, we rewrite the objective function (3.20) in OPC problem as X i2N Sub i [t]; (3.28) whereSub i [t] is defined as Sub i [t] = BW N + 1 X j2M log 2 (1 + P tx ij [t]g ij (N + 1) P int j + 2 ) X j2M Z j [t]BW log 2 (1 + P tx ij [t]g ij P int j + 2 ) + X j2M Z j [t] th N + 1 + ~ E i [t] 1 D e i [t] C h i [t] (3.29) Note thatSub i [t] is independent ofP i 0 j [t]’s andy i 0[t]’s for anyi 0 6=i, and the constraints related to different base stations are also non-overlapping in the OPC problem. Conse- quently, the optimal solution of the OPC problem can be obtained by solving for optimal Sub i [t]’s for each base station. The subproblem for base stationi can be efficiently solved using a loop over the value of y i [t] (which can be either 0 or 1). In each iteration, the problem of solving for optimalP ij [t]’s becomes a standard convex optimization problem and standard solution techniques such as the interior point method [19] can be used. The proposed solution framework is summarized in Algorithm 3. 3.5 Experimental Results To numerically evaluate our proposed framework, we use the Youtube request trace data from a campus network measurement conducted on the University of Massachusetts’ 63 Algorithm 3: Proposed Online Control Algorithm 1 for each time slott do 2 for each base stationi2N do 3 Observe the virtual queueZ j [t] 4 Observe its energy depletion queue ~ E i [t] 5 Sety i [t] = 1, solve theSub i [t] for active state 6 Sety i [t] = 0, solve theSub i [t] for sleep state 7 Set transmission powerP tx ij [t] and operation modey i [t] to the set of result achieving smallerSub i [t] 8 end 9 for eachj2M do 10 ComputeZ j [t + 1] according to Eqn. (3.14) 11 end 12 for eachi2N do 13 Compute ~ E i [t + 1] according to Eqn. (3.15) 14 end 15 end Amherst campus between June 2007 and March 2008 1 . Each data entry in the trace cor- responds to a download record of video clip from Youtube server, consisting of a times- tamp (in seconds), a client IP address, a requested video ID, and a Youtube server IP address. We parse the data on September 15, 2007 and obtain the information of 19,061 requests from 2,344 unique users. A cooperative region of 1km 2 is considered which is covered by one MBS and four sBSs. All base stations are assumed to be solely pow- ered using photovoltaic (PV) cells. We extract the solar radiation level from the TMY3 weather dataset on September from National Renewable Energy Laboratory (NREL) at 1 http://traces.cs.umass.edu/index.php/Network/Network 64 Hawthorne Municipal Airport, Los Angeles 2 . The energy harvested by PV cells is com- puted using PVWatts calculator 3 with an energy conversion efficiency of 15%. The nom- inal capacity of batteriesC max i in MBS and sBS is assumed to be 1KWh and 0:1KWh, respectively. We set the duration of each time slot to be 10 mins. To estimate the SINR, the path-loss for an sBS and an MBS are modeled as 30:6 + 36:7 log 10 (d) and 35:3 + 37:6 log 10 (d) indB, respectively [56], whered is the distance specified in kilometers from a UE to a base station. User locations are uniformly gen- erated within the cooperative area. The noise power is set to 2 =95dBm [56]. The maximum numbers of cached files in the MBS and sBSs are set to 4,000 and 1,000, respectively. We adopt a least frequently used replacement strategy in caches in sBSs and MBSs which replaces 5% and 2% of entries every two hours. Specifications of other channel- and power- related parameters can be found in Table 3.1. Note that some power-related parameters are set differently for the MBS and the sBSs. Our proposed algorithm is compared against with two baseline algorithms. The first baseline is referred to as “MBS-only”, which only considers the MBS and does not make use of sBSs at all. The second baseline is referred to as “greedy”, in which each base station tries to use its maximum transmission power whenever it is possible. When the battery is depleted and there is not enough harvested energy, the base station will simply shut itself down until more energy is available. For the proposed algorithm, we use two different values ofV as used in Eqn. (3.20). The average download throughput per user and the state of charge in the battery pack of the MBS in each time slot are shown in Fig. 3.3 and Fig. 3.4, respectively. As can be 2 http://rredc.nrel.gov/solar/old data/nsrdb/1991-2005/tmy3/ 3 http://pvwatts.nrel.gov/ 65 Table 3.1: Simulation Parameters. Parameter Value Parameter Value Channel BW 10MHz th 10 dBm Converter D 0.95 C 0.96 MBS [11] P max 0 43dBm p 4:7 P act 0 130W P slp 0 75W sBS [11] P max i 21dBm p 4:0 P act i 6:8W P slp i 4:3W seen from Fig. 3.3, the proposed algorithm achieves an average throughput of 82.5Mbps whenV = 10 5 over all time slot and outperforms the two baselines by 2.96x and 1.41x, respectively. From Fig. 3.4, one can see that the “greedy” algorithm will quickly drain the battery pack when the energy generation is low while the proposed algorithm can maintain a reasonable state of charge. When using different values forV , one can observe from the results that there exists a tradeoff between the download throughput and the state of charge in the batteries. When a largerV is used, the download throughput will be higher. However, the battery pack will also be discharged faster, which can potentially increase the risk of battery depletion and service failure. 3.6 Conclusion In this chapter, we investigate cooperative transmission and power management prob- lem for base stations powered solely by renewable energy in a cellular network. We first formulate the data transmission throughput maximization problem into a mixed-integer non-linear programming problem. Then, based on Lyapunov optimization theory, an efficient near-optimal solution method is proposed which is proved to have a bounded 66 Figure 3.3: Comparison of average download throughput per user Figure 3.4: Trace of state of charge in the MBS 67 optimality gap. Experimental results using realistic setting and data trace show that the proposed algorithm can achieve up to 2.96x average transmission rate compared to some baseline algorithms. 68 Chapter 4 Concurrent User Association and Dynamic Switching in Cells 4.1 Background and Prior Work Due to the limited capacity of existing network infrastructures, undesired large la- tency could be induced during peak-traffic hours especially at urban areas, stadiums that with massive user requests. Thus, users could suffer from the poor quality-of-service (QoS) especially those at the edge of base station’s coverage. To address this issue, the deployment of heterogeneous network (HetNet) has been investigated and integrated to significantly improve end-to-end network communication. Low power, short-range small cell base stations (sBSs), such as femtocells, pico cells, and micro cells are deployed un- der the coverage of macro cell base stations (MBSs). Indoor or outdoor sBSs can offload data transmission traffic from MBSs, and improve spectral efficiency via reduced trans- mission distance. Although the deployment of a large number of sBSs has been proved to be suc- cessful, the overhead of energy consumption in sBSs is non-negligible. According to 69 [44], the ICTs (information and communication technologies) consume about 3% of the worlds electrical energy which accounts for approximately 2% of the CO2 emission foot- print in the whole world; 9% of this consumption of ICTs is caused by communication networks[31]. Within these communication networks, 10% of the energy is consumed by user devices, while 90% is consumed by base stations (BSs). According to data in references [11, 89], there can be 100 times more sBSs than MBSs in a cellular network in practice, resulting in twice as much the total peak power consumption. Hence, from the perspective of energy efficiency, sBSs need to be self-organizing, low-cost, and energy- efficient. Since user requests are observed to be asynchronous and bursting, it is inevitable that BSs are under-utilized during a certain period of time, such as at night. Further energy saving can be achieved by proactively putting some of BSs into a deeper sleep mode in which BSs do not serve or monitor user requests (dynamic on/off switch). In [33], it is proposed to save power by regularly checking the traffic load level and mak- ing sleep decisions based on a certain traffic load threshold. The work [82] proposes to reduce the energy consumption of BSs by transmitting at full load during good channel conditions and go into sleep otherwise. In [77], optimal sleep strategies are determined for BSs based on power consumption minimization and energy efficiency maximization of a stochastic geometry based model. In [24], the problem of minimizing the energy consumption is studied and characterized into a 0-1 Knapsack problem and a central- ized algorithm is proposed to perform a joint user association and dynamic switching to minimize the energy consumption. 70 Some of aforementioned studies solely consider the optimization problem of dynam- ically switching BSs and treat policies of user association as different scenarios in simu- lations. For example, the received power based user association rule is the most prevalent one, where a user will choose to associate with the specific BS, which provides the maxi- mum received signal strength (max-RSS) [54]. The network system with a dense deploy- ment of sBSs inevitably render such a rudimentary user association rule ineffective, and jointly study the problem of dynamic switching of and user association are needed for addressing the unique features of the emerging network system. In addition, a central- ized approach that a central entity collects information, such as channel state information (CSI) and user requests, then make user association and dynamic switching decisions is computationally prohibitive. A distributed approach is more desirable owing to its low implementation complexity and low signaling overhead. In this chapter, we propose a distributed framework for a cellular network consisting of one MBS and a set of sBSs and user equipments (UEs). UEs request to download files from BSs through the down-link. We jointly consider the problem of user association and dynamic-switching of BSs. The energy minimization problem with power and channel constraints is formulated as a integer non-linear programming problem. Then, a belief propagation based distributed approach is proposed to solve this problem. The effec- tiveness of the proposed framework in practice is demonstrated by experimental results based on realistic user request traces by outperforming other baseline algorithms. 71 Figure 4.1: System model. 4.2 System Model We consider a two-tier HetNet system with a set of BSsM which includes one MBS overlaid with some sBSs and UEsN =f1;::;Ng, where one MBS (j = 1) is located at the center of studied area while sBSs (j2f2;:::;Mg) are randomly deployed and overlaid with the MBS without optimizing their locations. Each BS j divides its spectrum intok j orthogonal frequency subcarriers and is connected to the core network via a backhaul link of capacityB j . In this work, a slotted time model is adopted in which each time slot has a duration of. We further assume that the channel state information (CSI) corresponding to each subcarrier is perfectly known to the UE transmitters. We assume each UEi can be associated with only one BSj and request a file with size of f i [t] at the beginning of time slott. An binary indicatorx ij [t]2f0; 1g denotes whether 72 UEi is associated BSj (x ij [t] = 1) or not (x ij [t] = 0). Note that this assumption can be extended to multiple files by adding dummy UEs. The instantaneous transmission rate from BSj to UEi is given by r 0 ij [t] =BW j log 2 (1 + ij [t]); (4.1) where BW j is the bandwidth of each subband in BS j. We assume that when a BS j has multiple associated UEs, the associated UEs are scheduled as temporarily fairly or equally when sharing the bandwidth resource. ij [t] = p tx j g ij [t] I ij + 2 is the instantaneous signal-interference-plus-noise ratio (SINR) between BSj and UEi. Here,p tx j andg ij [t] denote the transmission power of BS i and the channel gain between BS i and UE j, respectively. g ij [t] is the product of log-distance path loss function and channel fading function. 2 is the noise power.I ij is the interference produced by other BSs. When a UE sends a request to access a file, the request will be served by its associated BS if it is in active state. Then, the requested file will be transferred via the downlink of the BS to the UE with a transmission rate r 0 ij [t]. More concretely, the BS is responsible for downloading the requested file from Internet via the backhaul link and relaying to the UE. In such a setting, the data transmission rate at which the requested file of UEi can be delivered from a BSj is calculated as r ij [t] =minfB j ;r 0 ij [t]g; (4.2) whereB j is the backhaul link capacity assigned to the BSj. 73 The binary variabley j [t]2f0; 1g denotes whether BS is operating in the active mode (y j [t] = 1) or in the sleep mode (y j [t] = 0) at the timet. We adopt a similar power model as in work [27, 85], in which the total energy consumption of a BSj2M at the timet, e j [t], is given by e j [t] = X i2N x ij [t] f i [t] r ij [t] p tx j + (1y j [t])p slp j +y j [t]p act j (4.3) where the first term is the load-dependent energy consumption. p act j and p slp j are the static power consumption of BS j in the active mode and sleep mode at time slot t, respectively. is the duration of studied time slot. According to the data in reference [56], the energy consumption of downloading via backhaul is usually much lower than the energy consumption of data transmission. Therefore, we choose not to introduce this term in the energy consumption model. 4.3 Problem Formulation Here, we aim at minimizing total energy consumption by assigning each UEi2N to the BS j 2M and making active/sleep operation decisions for each BS j. Given the information of channels and requested files, the joint energy minimization problem (JointEMIN) is formally formulated as follows Given: file request informationf i and channel state informationg ij . Find: the user associationx ij and the operation modey j . Minimize: T X t=0 M X j=1 e j [t] (4.4) 74 Subject to: M X j=1 x ij [t] = 1 8i;8t (4.5) N X i=1 x ij [t]k j 8j;8t (4.6) x ij [t]y j [t] 8i;8j;8t (4.7) x ij [t]2f0; 1g 8i;8j;8t (4.8) y j [t]2f0; 1g 8i;8t (4.9) where constraint (4.5) makes sure each UE is associated with and only with one BS. Constraint (4.6) states the number of subbands assigned to connected UEs cannot exceed the limit. Constraint (4.7) captures that UEs cannot connect to a BS in sleep mode. Constraints (4.8-4.9) indicatex ij [t] andy j [t] are binary variables. Note that constraints (4.6) and (4.7) can be combined into one constraint P N i=1 x ij [t] y j [t]k j . We know that the accumulative energy consumption across the time horizon T is a sum of each independent energy consumption e j [t] at each time slot. Therefore, we can solve the JointEMIN for each time slot instead. Theorem 2. Single Source Capacitated Facility Location Problem (SSCFLP) [74] can be reduced to JointEMIN problem. Proof. In SSCFLP problem, each customer has to be assigned to one facility that supplies its whole demand. The total demand of customers assigned to each facility cannot exceed its capacity. An opening cost is associated with each facility. The objective is to minimize 75 the total cost of opening the facilities and supply all the customers. Consider a special version of JointEMIN problem that the interference from other BSs are considered to be constant, CSIs are the same for all UEs and file requests have equal sizes. x ij can be seen as the indicator whether a customer i is assigned to a facility j. y j denotes whether the facilityj is open or not. k j is the facility’s capacity. Minimizing the energy consumption is the same as minimizing the total costs in SSCFLP. Therefore, we can solve any SSCFLP using JointEMIN problems solution in a special case. From theorem 2, we know that JointEMIN problem is as hard as SSCFLP, which is a known NP-hard problem. This implies that we cannot find an optimal solution in a poly- nomial time. In previous work, heuristics such as Tabu search, simulated annealing (SA) and genetic algorithms are applied to solve the SSCFLP and its variants [10]. However, difficulty in finding a proper stopping criteria and the computation latency overhead may incur significant transmission delay for all UEs. On the other hand, centralized heuristic search approaches require the existence of central controllers to collect information. For the dense deployment of sBSs network system with a large number of UEs, the signaling overhead and computation complexity is prohibitive. A distributed framework is more desirable due to the less information collecting and being able to harness the computation resources distributed at each side. 4.4 Proposed Solution In this section, we introduce our distributed solution to solve the JointEMIN prob- lem defined in (4.4), by applying a factor graph based message passing framework. As discussed in section 4.3, the JointEMIN problem is solved optimally if we can obtain 76 optimal solution for each time slot. Thus, we ignore the time slott in the discussion here for simplicity. 4.4.1 Factor Graph Model First, we define three functions depending on user association and active/sleep deci- sions as follows ~ F j =exp X i2N x ij a ij +y j ~ e j (4.10) G j = 8 > > > < > > > : 1; if P i2N x ij y j k j 0; otherwise (4.11) H i = 8 > > > < > > > : 1; if P j2M x ij = 1 0; otherwise (4.12) wherea ij = f i r ij p tx j and ~ e j = (p act j p slp j ). The functionH i only depends on user associ- ation variables. Both the function ~ F j andG j depend on user association and active/sleep variables, which allows us to combine them into a single function as follows F j = 8 > > > < > > > : exp P i2N x ij a ij +y j ~ e j ; if P i2N x ij y j k j 0; otherwise : (4.13) 77 Lemma 1. Finding the user association x ij and active/sleep decision y j to minimize the energy consumption in JointEMIN problem is equivalent to solving the following unconstrained problem X; Y = arg max X=f0;1g NM ;Y=f0;1g M1 M Y j=1 F j N Y i=1 H i : (4.14) Proof. The constant value part in Eqn. (4.3) is omitted since it doesn’t not influence the solution. Minimizing Eqn. (4.4) is equivalent to maximizing Q M j=1 exp P i2N x ij a ij + y j ~ e j . Because the exponential termF j is strictly positive, and violations of constraints (4.5), (4.6), (4.7)) would result in a zero product, constraints are implicitly satisfied when maximizing the objective. The factor graph representation of the equation (4.14) for solving the JointEMIN problem is illustrated in Fig. 4.2. Definition 1. A factor graph is a bipartite graph that expresses the structure of the fac- torization Q F j Q H i . A factor graph has a variable node for each variablex ij andy j , a factor node for each local functionF j andH i , and an edge connecting variable node to factor node if and only if the variable node is an argument of the function node. 78 Figure 4.2: Factor graph representation of the JointEMIN problem. 4.4.2 Message Passing Max-product Algorithm According to [48], an undirected factor graph model can represent a joint probability distribution given as follows P = 1 Z Y i=1 i ( i ): (4.15) whereZ is the normalization constant, =fW 1 ;W 2 ;:::g is a set of random variables (RVs). i denotes a subset of . Finding assignments of RVs to maximize Eqn. (4.15) is known as the maximum a posteriori (MAP) inference. Max-product algorithm is a message passing algorithm on factor graph models to solve MAP assignment problem. Let A!B (W =!) denote a message sent from a nodeA to a nodeB that contains the information about the RVW taking the value of!.n(A) denotes the set of neighbouring 79 nodes of nodeA in the factor graph. The message sent from a variable nodeW i to a factor node j is as follows W i ! j (!) = Y j 0 =n(W i )n j j 0!W i (!); (4.16) Basically, the message sent from the variable nodeW i to the factor node j is calculated as the product of messages toW i sent from all neighbouring factor nodes of the variable nodeW i excepting the factor node j . Messages sent from factor nodes to variable nodes are as follows j !W i (!) = max n( j )nW i j n( j ) Y W i 0 =n( j )nW i W i 0! j (!) ! : (4.17) where n( j )nW i denotes the set of neighbouring variable nodes of j excluding W i . Note that the message passing work process in a iteration way. After the completion of iterations, the belief at each variable node can be calculated as follows B W i (!) = Y j =n(W i ) j !W i (!); (4.18) which is the product of all incoming messages for the variable node W i . Finally, the variable values that achieving the maximum belief is the configuration of our MAP as- signment problem. 80 ! = arg max ! B W i (!) (4.19) In our JointEMIN problem, the user association and active/sleep decisions are made by following Eqn. (4.19) for each variable nodex ij andy j in Fig. 4.2, respectively. 4.4.3 Proposed Solution The proposed distributed message passing solution makes use of computation re- source at UEs and BSs and no centralized controller is needed for information collecting and decision making. The information is exchanged between UEs and BSs or BSs and BSs. One iteration is defined as every node has finished sending information to neigh- bouring corresponding node. After the completion of iterations, the UE could make decision which BS it would be associated with and the BS would make decision whether in active or sleep mode. It is noteworthy to mention that in the factor graph model, only the messages F j !x ij and x ij !F j involve communications from BSs to UEs and from UEs to BSs, respectively. When j 0 6= j, y j !F j 0 and F 0 j !y j require the exchange of message from BSj to BSj 0 and inversely, respectively. Whereas, H i !x ij and x ij !H i can be fulfilled locally in UEi, and F j !y j , y j !F j can be completed locally in BSj. i). Broadcasting: UEs broadcast to all BSs within a certain range to establish a connection. If there is no connection between UEi and BSj, the variable nodex ij in the factor graph doesn’t exist which result in a sparser graph. ii). Exchange messages between an UE and a BS: Since the variable x ij is a binary value, the message sending from UE i to BS j is in the form of binary tuple 81 x ij !F j = x ij !F j (0); x ij !F j (1) and the message sending from BS j to UE i is F j !x ij = F j !x ij (0); F j !x ij (1) . Initially, both messages are two scalar tuples (1; 1). iii). Exchange messages between two BSs: The variabley j is also a binary value, the message sending from BSj to BSj 0 is also in the form of binary tuple y j !F j 0 = y j !F j 0 (0); y j !F j 0 (1) and F j 0!y j = F j 0!y j (0); F j 0!y j (1) . Similarly, the mes- sage is two scalar values (1; 1). iv). Belief update and decision making: As mentioned before, one iteration is defined as all nodes have mutually exchanged messages with each other. After each iteration, every BS and UE should update their beliefs following Eqn. (4.18). When updating beliefs converges, a BS j and an UE i can make active/sleep and association decision by looking at the belief at the variable nodey j andx ij following Eqn. (4.19), respectively. 4.5 Experimental Results To numerically evaluate our proposed framework, we consider a two-tier HetNet consisting one MBS and specified number of sBSs in a area with a radius 2km. The MBS is located at the center of the area. The locations of sBSs are randomly generated within the studied area. 4000 UEs are considered in the 200 iterations of 20 UEs, which are uniformly distributed all around the area. To estimate the SINR, the path-loss for an sBS and an MBS are modeled as 30:6 + 36:7 log 10 (d) and 35:3 + 37:6 log 10 (d) indB, respectively [56], whered is the distance specified in kilometers from a UE to a BS. The noise power is set to 2 =95dBm [56]. The backhaul link capacity for all BSs is set to 20 Mbps. Specifications of other channel and power related parameters can be found 82 in Table 4.1. The maximum number of subbands is limited to 5. Note that some power- related parameters are set differently for the MBS and the sBSs. The file size is 80MB. The duration of each time step is 10 mins. For algorithmic parameters of message passing algorithm, the maximum number of iteration is 20 and stopping condition is = 0:001. Table 4.1: Simulation Parameters. Parameter Value Parameter Value Channel BW 10MHz 2 -95 dBm MBS [84] p tx 0 46dBm p slp 0 75:0W p act 0 724:6W sBS [84] p tx j 30dBm p slp j 3:87W p act j 10:16W We compare our proposed message passing algorithm with two baseline algorithms. they are described as follows: 1).First-fit algorithm: A UE choose to associate with the BS which can provide the strongest SINR. If the BS has reached the subband resource limit, the UE chooses the next BS in non-decreasing order until finds an available one. Then, all BSs that are not associated with any UE are put into sleep mode. 2). Simulated Annealing (SA) algorithm: This is centralized heuristic search algo- rithm for approximating the global optimum of a given function. It is proved that SA algorithm can find the global optimum if run for a long enough amount of time. At each iteration of SA, we define a move as follows: Associate each UE randomly with a BS which hasn’t reached its limit of subband resource and put the associated BS into active mode. The energy consumption E 0 of the new user association and active/sleep deci- sions is calculated and compared with the current energy consumptionE. IfE 0 <E, it is a good move, and we accept this move and continue to execute the next iteration. If E 0 E, we have a probabilityexp (E 0 E)=T to accept it. Here,T is also known 83 as temperature, which is initialized at a large value T max and decreases exponentially with the number of iterations to T min . Note that the acceptance probability would de- crease with decreasingT . The algorithmic parameters for SA algorithm are as follows: T max = 500,T min = 1, step limit is 200, iteration limit at each temperature is 500. Figure 4.3: Energy consumption for 8 BSs and 4000 UEs at each time slot. The energy consumption for each time slot 10 mins is shown Fig. (4.3). The average energy consumption for proposed message passing algorithm, first-fit algorithm and SA are 75.6 KJ, 493.9 KJ and 99.1 KJ, respectively. The total energy savings of our pro- posed algorithm are 84.7% and 23.7% compared with first-fit and SA. Note that SA is centralized algorithm that is computation prohibitive to applied to large scales. The av- erage throughput is shown in Fig. (4.4). It can be seen from Fig. (4.4), first-fit algorithm can provide the largest average throughput 2.1 Mbps compared with our proposed mes- sage passing algorithm 0.76 Mbps and SA 0.77 Mbps. However, the large throughput 84 Figure 4.4: Throughput for 8 BSs and 4000 UEs at each time slot. achieved by first-fit algorithm is at the cost of 6.5X energy consumption over message passing algorithm. The normalized energy efficiency is shown in Fig. (4.5), we can see that the energy efficiency is improved significantly by using proposed message passing algorithm against two baselines. 4.6 Conclusion In this work, we investigate joint user association and dynamic base station ac- tive/sleep switching problem in cache-enable HetNet system. We first formulate the energy consumption minimization problem and prove its NP-hardness. Then, based on factor graph model and max-product algorithm, an efficient distributed message passing solution method is proposed. Experimental results show that the proposed algorithm can 85 Figure 4.5: Normalized energy efficiency for different number of BSs. achieve up to 84.7% average energy saving rate compared with some baseline algorithms and improve the energy efficiency up to 2.43X. 86 Chapter 5 Joint Transmission and Charging Optimization in Wireless Powered IoT Systems 5.1 Background and Prior Work In the next generation networks systems, internet-of-things (IoTs) are fundamental elements aiming at providing ubiquitous connections between users and services. Driven by emerging services and applications, new techniques have been proposed and investi- gated to address new challenges. Machine learning, online gaming, augmented reality (AR) applications are usually power hungry and require a large amount of computation resources to process. However, the deployment of those applications on battery assisted IoT devices are hindered by the limitation of battery capacity. Moreover, the advances in battery related innovations still cannot keep up with the stringent demand of powering IoT devices for a long enough time. In order to prolong the longevity of IoT batteries and provide seamless services, intelligent power management and charging methods are highly needed. 87 Being able to charge batteries without power cords and outlets, wireless power trans- fer (WPT) technique is increasingly attracting research efforts from both academia and industries. It allows IoT devices to move freely and replenish energy storage without re- sorting to fixed and inconvenient energy resources. According to the power transmission distance, WPT techniques can be categorized into two types: near field: inductive coupling and resonance coupling. In both cases, the power is transferred between coils of wire by a magnetic filed. The difference is that resonant coupling technique has two resonant circuits in between. far field: microwave and laser. The power transmission via radio waves can be made more directional, allowing longer-distance power beaming. For the mi- crowave technique, a rectenna with high conversion efficiency may be used to convert the microwave energy back into electricity. For laser related techniques, the electricity is converted to a laser beam. Resonant beam charging technique [58], also as known as distributed laser charging (DLC), is able to transmit watt-level power over meter-level distance safely. Compar- ing with other WPT techniques, e.g, inductive coupling, magnetic resonance coupling, radio frequency, and so on, resonant beam charging is deemed to be more suitable for mobile IoT devices [34]. A new method [21], i.e. quasi-static cavity resonance (QSCR), developed by Disney Research for wirelessly transmitting power enables users to seam- lessly charge electronic devices throughout a specially built 16-by-16-foot room. Due to the latest breakthrough in 5G network systems, the communication distance would be significantly reduced because of the deployment of mm-Wave transmission, small cell base stations and massive MIMO techniques. The propagation loss for the wireless 88 energy transfer would be dramatically mitigated with shorter distances, which makes it more attractive for next generation network systems. The work [43] proposes a WPT enabled cellular network architecture and derives tradeoffs between the network parame- ters under the outage constraint. In this new network structure, power beacons (PBs) are deployed in a existing cellular network system for recharging mobile devices and sensors via microwave radiation known as microwave power transfer (MPT). The authors in [86] investigate the power transfer efficiency of WPT in support of simultaneous power and data transfer and find the channel variety among user equipments (UEs) inherent to the wireless environment can be exploited by opportunistic scheduling to significantly im- prove the power transfer efficiency. Ma et al. [59] study the optimal resource allocation in a power beacon-assisted wireless-powered communication network which consists of a set of access points (APs) and UE pairs and a PB. Each UE has no embedded power supply and must harvest energy first from the PB or APs and then transmit the data. The problem is then formulated as an auction game and propose an auction based distributed algorithm by considering the PB as the auctioneer and the APs as the bidders. In this work, we investigate a throughput maximization problem in wireless pow- ered IoT network systems. We propose an opportunistic joint transmission and charging management framework based on Lyapunov optimization framework. 5.2 System Model We consider a wireless powered IoT network system consisting of a PB, a set of IoT devices (UEs) and a base station (BS) as shown in Fig. 5.1. The PB transmits RF energy beams to UEs, which are denoted by a setN =f1; ;i; ;Ng. Each UE is 89 Figure 5.1: System model equipped with a battery of limited capacity and is solely powered by the harvested energy provided by the PB. A time division multiple access (TDMA) scheme is adopted for the PB to transmit energy. More concretely, if the whole operation time isT , the transferred energy via the energy downlink can be calculated asT , i.e.,2 (0; 1), considering the energy beamforming employed by the PB. We use i to denote the normalized allocated time period toUE i , where P i2N i 1. The energy harvested byUE i for a period oft is calculated as h i = i p tx B g e i i t; (5.1) where i is the energy harvesting efficiency at theUE i . p tx B is the energy transmission power of PB.g e i is channel coefficient for the energy transmission channel and is further calculated as p A (d PB i ) . d PB i is the distance from the PB to UE i . A and are 90 constants capturing the channel properties. When the data is transmitted fromUE i to the BS, the data transmission dater i is calculated as r i =BWlog 2 (1 + p tx i g s i 2 ); (5.2) where p tx i is the transmission power from UE i to the BS. g s i is the uplink channel co- efficient from UE i to the BS. sigma 2 is variance of the Gaussian noise. BW is the bandwidth. The power consumption at eachUE i is calculated as P i =p 0 i + tx i p tx i + c i f i ; (5.3) wherep 0 i is the static power. tx i is the dynamic power slope. The second term is related to the dynamic power consumption. The third term is the power consumption of processing the arrival data and computing. f i is the arrival data size. c i is the coefficient related the data size. From the previous discussion, we know that the dynamics of energy stored in each battery of UEs can be described as Q E i [t + 1] =Q E i [t]e i [t] +h i [t]; (5.4) where Q E i [t] is the energy queue of the UE i at the time slot t. e i [t] is the energy con- sumption ofUE i at the time slott.h i [t] is the energy harvested byUE i from the wireless transfer power at the time slott. Similarly, we have another data queueQ D i [t + 1] ofUE i 91 at the time slot t, which indicates the backlog of data needed to be transmitted. The dynamics can be updated as Q D i [t + 1] =Q D i [t]r i [t] +f i [t]; (5.5) wherer i [t] is the transmission rate forUE i calculated in 5.2 at the time slott.f i [t] is the arrival data size at the time slott. 5.3 Problem Formulation In this section, we present the formulation of throughput maximization problem (TMP) with energy and capacity constraints. In the TMP, the goal is to maximize the time average of sum throughput of all UEs. Here, we use a time slotted framwork. Each time slot ist with a duration of. Based on the modeling of system components as in Section 5.2, the TMP problem can be formally defined as follows Given: channel informationg s i [t] andg e i [t] at the beginning of each time slot. Arrival data sizef i [t] is revealed at the end of each time slot. Find: TDMA allocation i [t] for charging. The data transmission power ofp tx i [t]. Maximize: lim T!1 1 T T1 X t=0 X i2N r i [t] (5.6) 92 Subject to: 0p tx i [t]P tx;max i 8i;8t (5.7) 0 X i2N i [t] 1 8t (5.8) 0Q E i [t]C E;max i 8i;8t (5.9) 0Q D i C D;max i 8i;8t (5.10) Q E i [t + 1] =Q E i [t]e i [t] +h i [t] 8t;8i (5.11) Q D i [t + 1] =Q D i [t]r i [t] +f i [t] 8t;8i (5.12) Constraint (5.7) captures the transmission power of eachUE i cannot exceed the maxi- mum transmission powerP tx;max i . Constraint (5.8) ensures the i [t] is normalized. Con- straints (5.9) and (5.10) describe the properties of energy storage and data backlog. Con- straints (5.11) and (5.12) are the dynamics of energy queue and data queue, respectively. It is challenging to solve TMP problem because we don’t have a prior information of the system uncertainties. For example, channel information and arrival data size would randomly fluctuate. Too aggressively transmitting data in the short-term would cause the outage and reduce the throughput in the future. 93 5.4 Proposed Solution In this section, we first use Lyapunov optimization framework to transform the orig- inal TMP problem and then present our solution. We define the Lyapunov function as follows L([t]) = X i2N 1 2 ~ Q E i [t] 2 + 1 2 Q D i [t] 2 ; (5.13) where [t] = ~ Q E i ;Q D i is the combined queue vector of the system state. ~ Q E i = C E;max i Q E i is defined as the battery depletion queue. The Lyapunov drift is then calculated as ([t]) =L([t + 1])L([t]) (5.14) = X i2N ~ Q E i [t] e i [t]h i [t] +Q D i [t] f i [t]r i [t] + 1 2 e i [t]h i [t] 2 + 1 2 f i [t]r i [t] 2 : Assume the arrival data sizef i [t] is bounded, then we can have ([t]) X i2N ~ Q E i [t] e i [t]h i [t] +Q D i [t] f i [t]r i [t] +B; (5.15) where B is a constant. For our proposed control and resource allocation policy, we assume mean rate stable must be achieved for both energy depletion queue ~ Q E i and data queue Q D i . Moreover, the objective function of initial TMP is transformed into the expectation valueE( P i2N r i [t]j[t]). Then the online opportunistic control (OPC) problem is formally defined as 94 Given: channel informationg s i [t] andg e i [t] at the beginning of each time slot. System state vector [t]. Arrival data sizef i [t] is revealed at the end of each time slot. Find: TDMA allocation i [t] for charging. The data transmission power ofp tx i [t]. Minimize: ([t])VE( X i2N r i [t]j[t]) (5.16) Subject to: Eqns: (5:7) (5:12) (5.17) After a further analysis, we have a relationship as follows ([t])VE( X i2N r i [t]j[t]) X i2N ~ Q E i [t] e i [t]h i [t] +Q D i [t] f i [t]r i [t] (5.18) VBWlog 2 (1 + p tx i [t]g s i 2 ) +B Instead of minimizing the previous objective function, we minimize the upper bound (eliminate the constant B). Then the objective function is Minimize: X i2N ~ Q E i [t] e i [t]h i [t] +Q D i [t] f i [t]r i [t] VBWlog 2 (1 + p tx i [t]g s i 2 ) (5.19) At each time slot, we observe the system state queue vector, and solve the Eqn. (5.19 which is a convex function. The algorithm is described in Algorithm. (4). 95 Algorithm 4: Proposed Online Control Algorithm 1 for each time slott do 2 for eachUE i i2N do 3 Observe the data queueQ D i [t] 4 Observe the energy queueQ E i [t] 5 end 6 Solve the convex objective function 7 Set transmission powerp tx i [t] and TDMA allocation i [t] 8 for eachUE i ,i2N do 9 UpdateQ E i [t + 1] 10 UpdateQ D i [t + 1] 11 end 12 end Lemma 2. Assume y* is the optimal solution for the original TMP problem, our control and resource allocation policy can achieve lim T!1 1 T E( X i2N r i [t]j[t]) B V +y; (5.20) This means our proposed control and resource allocation policy can achieve a finite optimality gap comparing to the optimal solution for the original TMP problem. Proof. We know the following relation must be valid ([t])VE( X i2N r i [t]j[t]) X i2N ~ Q E i [t] 1 +Q D i [t] 2 Vy +B (5.21) Then, we take expectation on both sizes and have the following for each time slot: E ([t]) VE( X i2N r i [t]j[t]) X i2N E( ~ Q E i [t]) 1 +E(Q D i [t]) 2 Vy +B (5.22) 96 Summing overt2 [0; 1;:::;T 1] and it yields: E L([T] L([0]) V T X t=0 E( X i2N r i [t]j[t]) (5.23) T X t=0 X i2N E( ~ Q E i [t]) 1 +E(Q D i [t]) 2 Vy T +BT Divide both sizes byVT and take lim T!1 , then we can have the following: lim T!1 1 T E( X i2N r i [t]j[t]) B V +y ; (5.24) Therefore, Eqn. (5.20) is proved. 5.5 Experimental Results To numerically evaluate our proposed framework, we use the a region with a radius of 10 m. The band width BW is 10 MHz To estimate the SINR, the path-loss for data transmission from UE to BS is modeled as 30:6 + 36:7 log 10 (d) [56], whered is the distance specified in kilometers from a UE to a base station. User locations are uniformly generated within the region area. The noise power is set to 2 =95dBm [56]. The static power for UE is 1.2 W . The maximum transmission power of PB P tx;max i is 50 W . The battery on each UE has a maximum capacityC E;max i of 5000mAh, and output voltage 2V . The maximum transmission powerP tx;max i is 21dBm. For the path-loss g e i ,A = 10 3 and = 3 are the path loss exponent. The duration of each time slot is 10 mins. 97 Our proposed algorithm is compared against with two baseline algorithms, which are 1). round-robin: the PB charges every UE with equal time duration in a round-robin way. Each UE tries to use its maximum transmission power to transmit the data to BS whenever it is possible. 2). priority-based: the PB charges the UE which has the lowest SoC first using the whole time slot. Also the the maximum possible transmission power is used for each UE to transmit the data. In both cases, when the battery is depleted and there is not enough harvested energy, the UE will simply shut itself down until more energy is available. The PB charge every UE by equally distributing i . The average data transmission throughput per user and the average state of charge (SoC) in the battery pack of all UEs in each time slot are shown in Fig. 5.2 and Fig. 5.3, respectively. As can be seen from Fig. 5.2, the proposed algorithm achieves an average throughput of 5.7 Mbps whenV = 10 5 over all time slot and outperforms the round-robin and priority baseline by 4.38x and 1.17x for 20 UEs, respectively. From Fig. 5.3, one can see that the round-robin and priority-based charging algorithm will quickly drain the battery pack when the energy generation is low while the proposed algorithm can maintain a reasonable state of charge. Fig. 5.4 shows that two baseline algorithms cannot keep up with the increasing number of UEs because of the limitation of charging capability of PB. As discussed in previous chapter 3, we know that using different values forV , one can observe from the results that there exists a tradeoff between the download throughput and the state of charge in the batteries. When a largerV is used, the download throughput will be higher. However, the battery pack will also be discharged faster, which can potentially increase the risk of battery depletion and service failure. 98 Figure 5.2: The average throughput of network system of 20 UEs at different time slott. 5.6 Conclusion This chapter investigates the joint transmission and charging optimization in wireless powered IoT systems. We first formulate the throughput maximization problem and further study it using Lyapunov optimization framework. Then, a online opportunistic control solution is proposed and its optimality gap compared to the optimal solution has also been proved. The average throughput is improved by 4.38X by using our proposed online control framework against the greedy algorithm baseline. 99 Figure 5.3: The average battery state of charge (SoC) of 20 UEs at different time slott. Figure 5.4: The average throughput of network system for different number of UEs. 100 Chapter 6 Final Remarks and Looking Ahead This thesis introduced frameworks for achieving energy-efficient next generation net- work systems. Four individual works presented different perspectives and focuses. Chap- ter 2 studied how to utilize user contexts to drive the decision of dynamically switching a small cell base station between the active mode and the sleep mode to minimize the to- tal energy consumption. Chapter 3 investigated the cooperative transmission and power management problem for a set of “off-grid” base stations in a cellular network hierarchy that are powered solely by on-site renewable energy sources. A throughput maximiza- tion problem is solved under energy and capacity constraints. The chapter 4 proposed a distributed framework for a cellular network that jointly consider the problem of user as- sociation and dynamic-switching of BSs. The energy minimization problem with power and channel constraints was formulated as a integer non-linear programming problem. Then, a belief propagation based distributed approach was proposed to solve this prob- lem. The chapter 5 investigated a throughput maximization problem in wireless powered IoT network systems. An opportunistic joint transmission and charging management framework based on Lyapunov optimization framework was proposed. For each work, 101 the system model was carefully described and the corresponding optimization solutions were presented. The performance of proposed solutions was evaluated with experimen- tal results, either randomly generated based physical conditions or from realistic data. The experimental results proved the good performance and efficiency of our proposed solutions compared to competitive baseline algorithms. Moreover, the finite optimality gap between some of the solutions and the optimal solutions was carefully analyzed. Since energy efficient networking for various next generation cellular mobile com- munications systems is very critical and a challenging issue. For future studies, we need to address this problem hierarchically and tackle energy inefficiencies at different layers of communication systems. At the cloud layer, data centers would serve as the most important components un- der the software-defined networking scheme. In order to provide a seamless and reliable network that wirelessly connects a huge number of devices, data will need to be stored, transferred, and processed fast enough. Low utilization of servers in a data center is one of the biggest factors in low power efficiency of the data center. From the energy effi- ciency point of view, the non-energy-proportional nature of the current servers motivates data center operators to turn on as few servers as possible and make them highly utilized. Therefore, a management framework with the capability of consolidating and migrating of virtual machines (VM) to different physical servers is highly desired to meet various service level agreements (SLAs). At the network layer, as the virtualized network functions (VNFs) replace the tradi- tional function-dedicated equipment, the energy-aware placement approach of serviced- based functions is highly desired. Energy and computation resource constraints need to be carefully considered, and the end-to-end latency requirement should be satisfied. 102 Since sharing the computation resource even the VNFs themselves among different tasks for users is possible, the energy-aware placement approach should unavoidably leverage the information about the service functions each tasks request. For different topological structures of network systems, carefully placing the requested VNFs to as less servers as possible while ensuring latency constraints are satisfied could reduce the number of under-utilized servers and save energy consumption. Smart routing policies afterwards are critical for serving tasks with different At the device layer, mobile devices are always running on batteries and they are con- sidered limited in terms of processing resources and energy budget compared to the cloud server. Since many emerging applications are power hungry, such as machine learning, augmented reality and high quality video games, the mobile devices may offload some of the computation intensive tasks to the cloud. Since the information of network channel condition, resource, work load may vary from time to time, intelligent scheduling frame- works are needed to satisfy the end-to-end latency requirements and achieve high energy efficiency. 103 Bibliography [1] Europe 5G. https://5g-ppp.eu/kpis/. [2] Global system mobile association (GSMA). (dec. 18, 2015). [3] LTE the UMTS long term evolution: From theory to practice. John Wiley and Sons. [4] Cisco visual networking index: global mobile data traffic forecast update, 2016– 2021. Tech. Rep, 2016. [5] Arsalan Ahmad, Andrea Bianco, Edoardo Bonetto, Luca Chiaraviglio, and Filip Idzikowski. Energy-aware design of multilayer core networks [invited]. Journal of Optical Communications and Networking, 5(10):A127–A143, 2013. [6] J. G. Andrews, S. Buzzi, W. Choi, S. V . Hanly, A. Lozano, A. C. K. Soong, and J. C. Zhang. What will 5G be? IEEE Journal on Selected Areas in Communications, 32(6):1065–1082, June 2014. [7] Jeffrey G Andrews, Francois Baccelli, and Radha Krishna Ganti. A tractable ap- proach to coverage and rate in cellular networks. IEEE Transactions on Communi- cations, 59(11):3122–3134, 2011. [8] A. Aprem, C. R. Murthy, and N. B. Mehta. Transmit power control policies for energy harvesting sensors with retransmissions. IEEE Journal of Selected Topics in Signal Processing, 7(5):895–906, Oct 2013. [9] Oliver Arnold, Fred Richter, Gerhard Fettweis, and Oliver Blume. Power consump- tion modeling of different base station types in heterogeneous cellular networks. In Future Network and Mobile Summit, 2010, pages 1–8. IEEE, 2010. [10] Marvin A. Arostegui, Sukran N. Kadipasaoglu, and Basheer M. Khumawala. An empirical comparison of tabu search, simulated annealing, and genetic algorithms for facilities location problems. International Journal of Production Economics, 103(2):742 – 754, 2006. [11] G. Auer, V . Giannini, C. Desset, I. Godor, P. Skillermark, M. Olsson, M. A. Imran, D. Sabella, M. J. Gonzalez, O. Blume, and A. Fehske. How much energy is needed to run a wireless network? IEEE Wireless Communications, 18(5):40–49, October 2011. 104 [12] Bita Azimdoost, Cedric Westphal, and Hamid R Sadjadpour. Resolution-based content discovery in network of caches: Is the control traffic an issue? IEEE Transactions on Communications, 2017. [13] Ejder Bas ¸tu, Mehdi Bennis, Marios Kountouris, and M´ erouane Debbah. Cache- enabled small cell networks: Modeling and tradeoffs. EURASIP Journal on Wire- less Communications and Networking, 2015(1):1–11, 2015. [14] N. Bhushan, J. Li, D. Malladi, R. Gilmore, D. Brenner, A. Damnjanovic, R. T. Sukhavasi, C. Patel, and S. Geirhofer. Network densification: the dominant theme for wireless evolution into 5G. IEEE Communications Magazine, 52(2):82–89, February 2014. [15] E. Bjrnson, J. Hoydis, M. Kountouris, and M. Debbah. Massive mimo systems with non-ideal hardware: Energy efficiency, estimation, and capacity limits. IEEE Transactions on Information Theory, 60(11):7112–7139, Nov 2014. [16] E. Bjrnson, L. Sanguinetti, J. Hoydis, and M. Debbah. Designing multi-user MIMO for energy efficiency: When is massive MIMO the answer? In 2014 IEEE Wire- less Communications and Networking Conference (WCNC), pages 242–247, April 2014. [17] Charles Blundell, Julien Cornebise, Koray Kavukcuoglu, and Daan Wierstra. Weight uncertainty in neural networks. arXiv preprint arXiv:1505.05424, 2015. [18] H. Bogucka and A. Conti. Degrees of freedom for energy savings in practical adaptive wireless systems. IEEE Communications Magazine, 49(6):38–45, June 2011. [19] Stephen Boyd and Lieven Vandenberghe. Convex optimization. Cambridge univer- sity press, 2004. [20] Russell N Bradt, SM Johnson, and Samuel Karlin. On sequential designs for maximizing the sum of n observations. The Annals of Mathematical Statistics, 27(4):1060–1074, 1956. [21] Matthew J. Chabalko, Mohsen Shahmohammadi, and Alanson P. Sample. Qua- sistatic cavity resonance for ubiquitous wireless power transfer. PLOS ONE, 12(2):1–14, 02 2017. [22] Vinay Chamola and Biplab Sikdar. Solar powered cellular base stations: Cur- rent scenario, issues and proposed solutions. IEEE Communications magazine, 54(5):108–114, 2016. [23] Vinay Chamola, Biplab Sikdar, and Bhaskar Krishnamachari. Delay aware resource management for grid energy savings in green cellular base stations with hybrid power supplies. IEEE Transactions on Communications, 65(3):1092–1104, 2017. 105 [24] E. Chavarria-Reyes, I. F. Akyildiz, and E. Fadel. Energy consumption analysis and minimization in multi-layer heterogeneous wireless systems. IEEE Transactions on Mobile Computing, 14(12):2474–2487, Dec 2015. [25] Yuh-Shyan Chen and Cheng-You Wu. A green handover protocol in two-tier ofdma macrocellfemtocell networks. Mathematical and Computer Modelling, 57(11):2814 – 2831, 2013. Information System Security and Performance Mod- eling and Simulation for Future Mobile Networks. [26] Yi-Han Chiang and Wanjiun Liao. ENCORE: An energy-aware multicell cooper- ation in heterogeneous networks with content caching. In Computer Communica- tions, IEEE INFOCOM 2016-The 35th Annual IEEE International Conference on, pages 1–9. IEEE, 2016. [27] Yi-Han Chiang and Wanjiun Liao. Green multicell cooperation in heterogeneous networks with hybrid energy sources. IEEE Transactions on Wireless Communica- tions, 15(12):7911–7925, 2016. [28] Gencer Cili, Halim Yanikomeroglu, and F Richard Yu. Cell switch off technique combined with coordinated multi-point (CoMP) transmission for energy efficiency in beyond-LTE cellular networks. In Communications (ICC), 2012 IEEE Interna- tional Conference on, pages 5931–5935. IEEE, 2012. [29] P. Demestichas, A. Georgakopoulos, D. Karvounas, K. Tsagkaris, V . Stavroulaki, J. Lu, C. Xiong, and J. Yao. 5g on the horizon: Key challenges for the radio-access network. IEEE Vehicular Technology Magazine, 8(3):47–53, Sep. 2013. [30] Salah-Eddine Elayoubi, Louai Saker, and Tijani Chahed. Optimal control for base station sleep mode in energy efficient radio access networks. In INFOCOM 2011, pages 106–110. IEEE, 2011. [31] Ericsson. Sustainable energy use in mobile communications. 2007. [32] J. Erman, A. Gerber, M. Hajiaghayi, D. Pei, S. Sen, and O. Spatscheck. To cache or not to cache: The 3G case. IEEE Internet Computing, 15(2):27–34, March 2011. [33] L. Falconetti, P. Frenger, H. Kallin, and T. Rimhagen. Energy efficiency in hetero- geneous networks. In 2012 IEEE Online Conference on Green Communications (GreenCom), pages 98–103, Sept 2012. [34] W. Fang, Q. Zhang, Q. Liu, J. Wu, and P. Xia. Fair scheduling in resonant beam charging for iot devices. IEEE Internet of Things Journal, pages 1–1, 2018. [35] D. Feng, L. Lu, Y . Yuan-Wu, G. Y . Li, S. Li, and G. Feng. Device-to-device com- munications in cellular networks. IEEE Communications Magazine, 52(4):49–55, April 2014. 106 [36] X. Ge, S. Tu, G. Mao, C. Wang, and T. Han. 5g ultra-dense cellular networks. IEEE Wireless Communications, 23(1):72–79, February 2016. [37] X. Ge, J. Ye, Y . Yang, and Q. Li. User mobility evaluation for 5G small cell net- works based on individual mobility model. IEEE Journal on Selected Areas in Communications, 34(3):528–541, March 2016. [38] Negin Golrezaei, Karthikeyan Shanmugam, Alexandros G Dimakis, Andreas F Molisch, and Giuseppe Caire. Femtocaching: Wireless video content delivery through distributed caching helpers. In INFOCOM, pages 1107–1115. IEEE, 2012. [39] Alex Graves. Practical variational inference for neural networks. In Advances in Neural Information Processing Systems, pages 2348–2356, 2011. [40] M. Gruber, O. Blume, D. Ferling, D. Zeller, M. A. Imran, and E. C. Strinati. Earth energy aware radio and network technologies. In 2009 IEEE 20th International Symposium on Personal, Indoor and Mobile Radio Communications, pages 1–5, Sep. 2009. [41] F Maxwell Harper and Joseph A Konstan. The movielens datasets: History and context. ACM Transactions on Interactive Intelligent Systems (TiiS), 5(4):19, 2016. [42] X. Hong, J. Wang, C. Wang, and J. Shi. Cognitive radio in 5g: a perspective on energy-spectral efficiency trade-off. IEEE Communications Magazine, 52(7):46– 53, July 2014. [43] K. Huang and V . K. N. Lau. Enabling wireless power transfer in cellular networks: Architecture, modeling and deployment. IEEE Transactions on Wireless Commu- nications, 13(2):902–912, February 2014. [44] I. Humar, X. Ge, L. Xiang, M. Jo, M. Chen, and J. Zhang. Rethinking energy efficiency models of cellular networks with embodied energy. IEEE Network, 25(2):40–49, March 2011. [45] I. Hwang, B. Song, and S. S. Soliman. A holistic view on hyper-dense hetero- geneous and small cell networks. IEEE Communications Magazine, 51(6):20–27, June 2013. [46] Ralf Irmer, Heinz Droste, Patrick Marsch, Michael Grieger, Gerhard Fettweis, Ste- fan Brueck, Hans-Peter Mayer, Lars Thiele, and V olker Jungnickel. Coordinated multipoint: Concepts, performance, and field trial results. IEEE Communications Magazine, 49(2):102–111, 2011. [47] C. Jiang, H. Zhang, Y . Ren, Z. Han, K. Chen, and L. Hanzo. Machine learning paradigms for next-generation wireless networks. IEEE Wireless Communications, 24(2):98–105, April 2017. 107 [48] F. R. Kschischang, B. J. Frey, and H. . Loeliger. Factor graphs and the sum-product algorithm. IEEE Transactions on Information Theory, 47(2):498–519, Feb 2001. [49] Alp Kucukelbir, Dustin Tran, Rajesh Ranganath, Andrew Gelman, and David M Blei. Automatic differentiation variational inference. arXiv preprint arXiv:1603.00788, 2016. [50] E. G. Larsson, O. Edfors, F. Tufvesson, and T. L. Marzetta. Massive MIMO for next generation wireless systems. IEEE Communications Magazine, 52(2):186– 195, February 2014. [51] J. Lee, Y . Kim, H. Lee, B. L. Ng, D. Mazzarese, J. Liu, W. Xiao, and Y . Zhou. Co- ordinated multipoint transmission and reception in LTE-advanced systems. IEEE Communications Magazine, 50(11):44–50, November 2012. [52] Suoheng Li, Jie Xu, Mihaela van der Schaar, and Weiping Li. Popularity-driven content caching. In INFOCOM 2016, pages 1–9. IEEE, 2016. [53] D. Liu, Y . Chen, K. K. Chai, and T. Zhang. Distributed delay-energy aware user association in 3-tier hetnets with hybrid energy sources. In 2014 IEEE Globecom Workshops (GC Wkshps), pages 1109–1114, Dec 2014. [54] D. Liu, L. Wang, Y . Chen, M. Elkashlan, K. Wong, R. Schober, and L. Hanzo. User association in 5g networks: A survey and an outlook. IEEE Communications Surveys Tutorials, 18(2):1018–1044, Secondquarter 2016. [55] D. Liu and C. Yang. Energy efficiency of downlink networks with caching at base stations. IEEE Journal on Selected Areas in Communications, 34(4):907–922, April 2016. [56] Dong Liu and Chenyang Yang. Will caching at base station improve energy effi- ciency of downlink transmission? In GlobalSIP, pages 173–177. IEEE, 2014. [57] J. Liu, B. Bai, J. Zhang, and K. B. Letaief. Content caching at the wireless network edge: A distributed algorithm via belief propagation. In 2016 IEEE International Conference on Communications (ICC), pages 1–6, May 2016. [58] Q. Liu, J. Wu, P. Xia, S. Zhao, W. Chen, Y . Yang, and L. Hanzo. Charging un- plugged: Will distributed laser charging for mobile wireless power transfer work? IEEE Vehicular Technology Magazine, 11(4):36–45, Dec 2016. [59] Y . Ma, H. Chen, Z. Lin, Y . Li, and B. Vucetic. Distributed and optimal resource allo- cation for power beacon-assisted wireless-powered communications. IEEE Trans- actions on Communications, 63(10):3569–3583, Oct 2015. [60] S. Maghsudi and E. Hossain. Multi-armed bandits with application to 5G small cells. IEEE Wireless Communications, 23(3):64–73, June 2016. 108 [61] Patrick Marsch and Gerhard Fettweis. Static clustering for cooperative multi-point (comp) in mobile communications. In Communications (ICC), 2011 IEEE Interna- tional Conference on, pages 1–6. IEEE, 2011. [62] Sabrina M¨ uller, Onur Atan, Mihaela van der Schaar, and Anja Klein. Smart caching in wireless small cell networks via contextual multi-armed bandits. In ICC, pages 1–7. IEEE, 2016. [63] Sabrina M¨ uller, Onur Atan, Mihaela van der Schaar, and Anja Klein. Context- aware proactive content caching with service differentiation in wireless networks. IEEE Transactions on Wireless Communications, 16(2):1024–1036, 2017. [64] Radford M Neal. Bayesian learning for neural networks, volume 118. Springer Science & Business Media, 2012. [65] Michael J Neely. Stochastic network optimization with application to communi- cation and queueing systems. Synthesis Lectures on Communication Networks, 3(1):1–211, 2010. [66] E. Oh, K. Son, and B. Krishnamachari. Dynamic base station switching-on/off strategies for green cellular networks. IEEE Transactions on Wireless Communica- tions, 12(5):2126–2136, May 2013. [67] A. Osseiran, F. Boccardi, V . Braun, K. Kusume, P. Marsch, M. Maternia, O. Que- seth, M. Schellmann, H. Schotten, H. Taoka, H. Tullberg, M. A. Uusitalo, B. Timus, and M. Fallgren. cenarios for 5G mobile and wireless communications: the vision of the METIS project. IEEE Communications Magazine, 52(5):26–35, May 2014. [68] K. Poularakis, G. Iosifidis, and L. Tassiulas. Joint caching and base station ac- tivation for green heterogeneous cellular networks. In 2015 IEEE International Conference on Communications (ICC), pages 3364–3369, June 2015. [69] Konstantinos Poularakis, George Iosifidis, and Leandros Tassiulas. Joint caching and base station activation for green heterogeneous cellular networks. In ICC, pages 3364–3369. IEEE, 2015. [70] Buvaneswari A Ramanan, Lawrence M Drabeck, Mark Haner, Nachi Nithi, Thierry E Klein, and Chitra Sawkar. Cacheability analysis of http traffic in an operational lte network. In IEEE WTS, pages 1–8. IEEE, 2013. [71] Ra´ ul Rojas. Neural networks: a systematic introduction. Springer Science & Busi- ness Media, 2013. [72] Kei Sakaguchi, Gia Khanh Tran, Hidekazu Shimodaira, Shinobu Nanba, Toshiaki Sakurai, Koji Takinami, Isabelle Siaud, Emilio Calvanese Strinati, Antonio Capone, Ingolf Karls, et al. Millimeter-wave evolution for 5G cellular networks. IEICE Transactions on Communications, 98(3):388–402, 2015. 109 [73] Mamoru Sawahashi, Yoshihisa Kishiyama, Akihito Morimoto, Daisuke Nishikawa, and Motohiro Tanno. Coordinated multipoint transmission/reception techniques for lte-advanced [coordinated and distributed mimo]. IEEE Wireless Communications, 17(3), 2010. [74] F. J. F. Silva and D. S. de la Figuera. A capacitated facility location problem with constrained backlogging probabilities. International Journal of Production Research, 45(21):5117–5134, 2007. [75] M. S. Sim, M. Chung, D. Kim, J. Chung, D. K. Kim, and C. Chae. Nonlinear self- interference cancellation for full-duplex radios: From link-level and system-level performance perspectives. IEEE Communications Magazine, 55(9):158–167, Sep. 2017. [76] O. Simeone, A. Maeder, M. Peng, O. Sahin, and W. Yu. Cloud radio access net- work: Virtualizing wireless access for dense heterogeneous systems. Journal of Communications and Networks, 18(2):135–149, April 2016. [77] Y . S. Soh, T. Q. S. Quek, M. Kountouris, and H. Shin. Energy efficient hetero- geneous cellular networks. IEEE Journal on Selected Areas in Communications, 31(5):840–850, May 2013. [78] Richard S Sutton and Andrew G Barto. Reinforcement learning: An introduction, volume 1. MIT press Cambridge, 1998. [79] M. N. Tehrani, M. Uysal, and H. Yanikomeroglu. Device-to-device communica- tion in 5g cellular networks: challenges, solutions, and future directions. IEEE Communications Magazine, 52(5):86–92, May 2014. [80] William R Thompson. On the likelihood that one unknown probability exceeds another in view of the evidence of two samples. Biometrika, 25(3/4):285–294, 1933. [81] R. Torrea-Duran, C. Desset, S. Pollin, and A. Dejonghe. Adaptive energy efficient scheduling algorithm for lte pico base stations. In 2012 Future Network Mobile Summit (FutureNetw), pages 1–8, July 2012. [82] R. Torrea-Duran, C. Desset, S. Pollin, and A. Dejonghe. Adaptive energy efficient scheduling algorithm for lte pico base stations. In 2012 Future Network Mobile Summit (FutureNetw), pages 1–8, July 2012. [83] Gia Khanh Tran, Hidekazu Shimodaira, Roya Ebrahim Rezagah, Kei Sakaguchi, and Kiyomichi Araki. Dynamic cell activation and user association for green 5G heterogeneous cellular networks. In PIMRC, pages 2364–2368. IEEE, 2015. 110 [84] L. Wang, S. Chen, and M. Pedram. Context-driven power management in cache- enabled base stations using a bayesian neural network. In 2017 Eighth International Green and Sustainable Computing Conference (IGSC), pages 1–8, Oct 2017. [85] L. Wang, S. Chen, and M. Pedram. Power management of cache-enabled coopera- tive base stations towards zero grid energy. In 2018 IEEE International Conference on Communications (ICC), pages 1–6, May 2018. [86] M. Xia and S. Aissa. On the efficiency of far-field wireless power transfer. IEEE Transactions on Signal Processing, 63(11):2835–2847, June 2015. [87] Ming Xia, Yasunori Owada, Masugi Inoue, and Hiroaki Harai. Optical and wire- less hybrid access networks: Design and optimization. J. Opt. Commun. Netw., 4(10):749–759, Oct 2012. [88] H. Zhang, C. Jiang, J. Cheng, and V . C. M. Leung. Cooperative interference mitiga- tion and handover management for heterogeneous cloud small cell networks. IEEE Wireless Communications, 22(3):92–99, June 2015. [89] J. Zhang, X. Zhang, M. A. Imran, B. Evans, and W. Wang. Energy efficiency anal- ysis of heterogeneous cache-enabled 5G hyper cellular networks. In GLOBECOM, pages 1–6, Dec 2016. [90] X. Zhang, J. Zhang, W. Wang, Y . Zhang, C. I, Z. Pan, G. Li, and Y . Chen. Macro- assisted data-only carrier for 5G green cellular systems. IEEE Communications Magazine, 53(5):223–231, May 2015. [91] Sheng Zhou, Jie Gong, Zexi Yang, Zhisheng Niu, and Peng Yang. Green mobile access network with dynamic base station energy saving. In ACM MobiCom, 2009. 111

Abstract (if available)

Abstract The incremental improvements on the existing infrastructure fail to meet the explosive data demands of the foreseeable future. The goals of next generation mobile network (5G) are broad and are presumed to include much greater throughput, much lower latency, and lower power consumption. Promising solutions such as heterogeneous network (HetNet), cell sleeping techniques, cache-aided base stations (BSs) and renewable energy power supply have been proposed and deployed to address those challenges. Although the deployment of dense small cell base stations (sBSs) in HetNet can efficiently offload traffic from the existed macro cell base stations (MBSs), the further benefits have not been harnessed well enough and the potential overhead should be carefully alleviated. This work aim to address these challenges by embracing novel mathematical frameworks in control theory, statistical machine learning and probabilistic graph model and advocate for intelligent, energy-efficient and low complexity methodologies for the management of entities (BSs, UEs, power beacon) for next generation network systems. A variational inference (VI) based Bayesian neural network (BNN) is proposed to learn from user's contexts to drive the decision of dynamically switching an sBS between the active mode and the sleep mode to minimize the total energy consumption by solving a multi-armed bandit problem. An efficient online near-optimal control policy is derived with a proved finite optimality gap for cooperative BSs powered solely by renewable energy in HetNet by Lyapunov optimization theory and convex optimization. The problem of user association and dynamic-switching of BSs is jointly considered to minimize energy consumption. In further, an opportunistic joint transmission and charging management framework is proposed to maximize the throughput of a wireless powered IoT network system.

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

SLA-based, energy-efficient resource management in cloud computing systems

PDF

A joint framework of design, control, and applications of energy generation and energy storage systems

PDF

Integration of energy-efficient infrastructures and policies in smart grid

PDF

Advanced cell design and reconfigurable circuits for single flux quantum technology

PDF

Energy-efficient computing: Datacenters, mobile devices, and mobile clouds

PDF

Designing efficient algorithms and developing suitable software tools to support logic synthesis of superconducting single flux quantum circuits

PDF

Development of electronic design automation tools for large-scale single flux quantum circuits

PDF

Towards energy efficient mobile sensing

PDF

Multi-level and energy-aware resource consolidation in a virtualized cloud computing system

PDF

Energy-efficient shutdown of circuit components and computing systems

PDF

Power efficient design of SRAM arrays and optimal design of signal and power distribution networks in VLSI circuits

PDF

Energy proportional computing for multi-core and many-core servers

PDF

Hardware techniques for efficient communication in transactional systems

PDF

Variation-aware circuit and chip level power optimization in digital VLSI systems

PDF

Optimizing power delivery networks in VLSI platforms

PDF

Demand based techniques to improve the energy efficiency of the execution units and the register file in general purpose graphics processing units

PDF

Verification and testing of rapid single-flux-quantum (RSFQ) circuit for certifying logical correctness and performance

PDF

Energy efficient design and provisioning of hardware resources in modern computing systems

PDF

Automatic conversion from flip-flop to 3-phase latch-based designs

PDF

Architectures and algorithms of charge management and thermal control for energy storage systems and mobile devices

Asset Metadata

Creator Wang, Luhao (author)

Core Title Towards green communications: energy efficient solutions for the next generation cellular mobile communication systems

School Viterbi School of Engineering

Degree Doctor of Philosophy

Degree Program Electrical Engineering

Publication Date 04/25/2019

Defense Date 02/28/2019

Publisher University of Southern California (original), University of Southern California. Libraries (digital)

Tag 5G,base station,energy efficient,green communication,OAI-PMH Harvest

Format application/pdf (imt)

Language English

Contributor Electronically uploaded by the author (provenance)

Advisor Pedram, Massoud (committee chair), Gupta, Sandeep (committee member), Nakano, Aiichiro (committee member)

Creator Email luhaowan@usc.edu,luhaowang924@gmail.com

Permanent Link (DOI) https://doi.org/10.25549/usctheses-c89-147083

Unique identifier UC11662694

Identifier etd-WangLuhao-7270.pdf (filename),usctheses-c89-147083 (legacy record id)

Legacy Identifier etd-WangLuhao-7270.pdf

Dmrecord 147083

Document Type Dissertation

Format application/pdf (imt)

Rights Wang, Luhao

Type texts

Source University of Southern California (contributing entity), University of Southern California Dissertations and Theses (collection)

Access Conditions The author retains rights to his/her dissertation, thesis or other graduate work according to U.S. copyright law. Electronic access is being provided by the USC Libraries in agreement with the a...

Repository Name University of Southern California Digital Library

Repository Location USC Digital Library, University of Southern California, University Park Campus MC 2810, 3434 South Grand Avenue, 2nd Floor, Los Angeles, California 90089-2810, USA