2.1.1 Dynamic Planning

Traditionally, the cell size and capacity in network planning are designed based on the peak call traffic load. As discussed in Chapter 1, Section 1.2, the call traffic load exhibits significant spatial and temporal fluctuations. Consequently, the network is over-provisioned at a low call traffic load, which in turn results in energy waste. Switching off some of the available radio resources (e.g. radio transceivers of BSs) at a low call traffic load can yield energy saving and offer acceptable performance. On the contrary, an active BS spends of its total power consumption in processing circuits and air conditioning units (a component represented by the fixed power component in (1.11) and (1.13)) [34]. Consequently, an effective energy-savingapproach at a low call traffic load is to switch off some of the network BSs while simultaneously satisfying the target performance metrics. BS on–off switching according to call traffic load conditions is referred to as dynamic planning [12].

In order to design an effective BS switching mechanism, two issues must be addressed, namely the user association problem and BS operation. The BS on–off switching is coupled with the user association problem. In particular, user association is inevitable to concentrate the call traffic load in a few BSs, and hence to switch off other lightly loaded BSs. Therefore, newly incoming MTs should be associated with a subset of active BSs. In addition, MTs already in service should perform handover when their serving BSs are switching off. Two research directions related to MT association can be identified. The objective of the first direction is to develop new energy-efficient user association mechanisms [34, 35, 63–65], while the second direction aims to derive analytical models to assess the performance of different energy-efficient association mechanisms [40].

In developing an energy-efficient MT association mechanism, two approaches can be adopted to meet the MT target QoS while concentrating the call traffic load in a few BSs. The first approach assumes an objective function that minimizes the networks' energy consumption while accounting for the user target QoS constraints. On the contrary, the second approach aims to balance the trade-off between MTs' flow-level performance (e.g. data rate or delay) and network energy consumption [35]. The latter case assumes a multiobjective optimization problem with a weighting factor. When the weighting factor equals zero, the MT association is determined based only on the MT flow-level performance. As the weighting factor increases, the MT association decision pays more attention to the network power consumption performance. When the weighting factor reaches infinity, the MT is associated to the BS that maximizes the network energy efficiency performance in bits per Joule. Overall, the MT association mechanism can assume a centralized or decentralized architecture [34]. The objective of both architectures is to concentrate the traffic load in a few BSs while satisfying the MTs' target data rate and the bandwidth limitations of BSs. The centralized mechanisms use a central controller that uses global network information related to channel conditions and user requirements to perform an energy-efficient MT association. On the contrary, in a decentralized architecture, an MT locally selects the BS with the highest call traffic load that can serve its target data. The main challenge in designing such a decentralized mechanism is the associated computational complexity imposed by the binary nature of the BS on–off switching decision variables, and hence the mixed-integer nature of the optimization problem. Consequently, greedy algorithms are mainly used to reach a good (sub-optimal) switching decision [34, 35]. In order to design greedy algorithms, a decision criterion should be defined. For instance, a greedy algorithm based on a user–BS distance decision criterion switches off the BSs with the longest user–BS distance to improve the network energy efficiency [63]. The rationale behind this decision criterion is that the longer the user–BS distance, the greater the transmission powerrequired to satisfy the users' target service quality. Another decision criterion is referred to as the network impact, which quantifies the impact of switching off a given BS on the network performance [63]. Switching off a given BS leads to additional load increments into the neighbouring BSs. In addition, switching off a BS can also lead to a positive effect on the neighbouring BSs because of the reduced inter-cell interference. By quantifying the two aforementioned measures, the network impact criterion maps the switching off decision as a BS selection problem, whose objective is to find the BSs that when switched off result in the highest network impact [64]. Furthermore, the coverage holes represent an important problem associated with BS on–off switching. As a result, another BS on–off switching decision metric is related to coverage holes avoidance. In [65], it has been shown that finding the optimal set of BSs that: (i) minimizes the network power consumption and (ii) avoids coverage holes is closely related to the minimum-weight disc cover problem. This problem is known to be NP-hard, and a greedy algorithm is proposed to switch off BSs while maintaining network coverage in polynomial time complexity.

Queueing models are proposed in the literature to assess the performance of different energy-efficient MT association mechanisms [40]. The energy-efficient MT association process in the overlapped coverage of different BSs is modelled as a customer joining a queue with servers, where and denote the number of BSs with overlapped coverage and the maximum number of MTs that can be accommodated in each BS, respectively. For instance, consider a two-BS scenario with three service areas. In service areas 1 and 2, an MT is served by the BS covering that area. In service area 3, an MT can be served by either BS with overlapped coverage. A BS is switched off, and consequently, its corresponding servers are shut down, if no MT is assigned to it. Following such a queueing model, analytical expressions are derived for call-blocking probability, average number of MTs assigned to each BS and average power and energy consumed by the network operator to serve one MT [40]. The model can be further approximated to account for the multiple-BS overlapped coverage case.

On the basis of the MT association phase, the BS operation decision is specified. In particular, BSs with a concentrated call traffic load become active, while lightly loaded BSs are switched off. The BS operation phase deals with three concerns: (i) accommodating future traffic demands, (ii) determining BS wake-up instants for switched-off BSs and (iii) implementing the BS on–off switching decisions. For accommodating future traffic demands, it should be noted that the BS operation decision lasts for a long duration (i.e. several hours), since frequent BS on and off switching is not desirable due to the increased energy consumption in the BS start-up phase [12] and service unavailability for the off cells during the decision computation phase [34]. Consequently, the BS operation decision should account for the future call traffic load by reserving some resources (bandwidth) to accommodate the future demands [34] by exploiting the past call traffic load patterns to estimate the future load [12]. Another approach to estimate the future trafficdemands is based on an online stochastic game [66], where neighbouring BSs communicate with each other to predict their traffic profiles, leading eventually to optimal switching decisions and minimum network energy consumption.

For determining the BS wake-up instants, it should be noted that switching off some cells is acceptable only if the active BSs extend their coverage areas to support the cells with inactive BSs. When the call traffic load of the inactive cells increases beyond the capacity limit of the active BSs, some of the inactive BSs should be switched on. Therefore, besides specifying which BSs to be switched off, another equally important research issue refers to determining the wake-up instants for switched-off BSs. Two BS wake-up schemes are proposed in [67], namely the number - and the vacation time -based schemes, respectively, as shown in Figure 2.1. In the -based scheme, the BS is switched off in an idle condition (i.e. no MT in service) and it wakes up when users arrive at the BS coverage area. On the contrary, for the -based scheme, the BS remains in a sleep state for a specific period of vacation time before waking up. Two versions can be distinguished for the -based scheme, namely the single vacation and multiple vacations. In the single vacation case, the BS remains awake after the vacation period even if there is no call request to serve, while in the multiple vacations case, the BS goes back to sleep if it wakes up and finds no call request to serve. A limitation associated with the -based scheme is that the BS needs to continuously monitor the user request arrivals, which translates into an advantage for the -based scheme. For femto-cell BSs in overlapped coverage with macro-cell BSs, three wake-up modes can be distinguished in the literature, namely BS-controlled, MT-controlled and network-controlled modes [68]. In the BS-controlled mode, the femto-cell BS performs continuous sensing for user activity to wake up, while in the MT-controlled mode, the MT sends wake-up messages for a sleeping femto BS. In the network-controlled mode, the core network controls the femto-BS operation through wake-up messages over the backhaul link. The three different wake-up modes result in different performance in terms of BS and MT energy consumption and signalling overhead. The BS-controlled mode leads to less energy saving for the BS. The MT-controlled mode increases the energy consumption for the MT, while the network-controlled mode incurs additional signalling overhead [68]. Markov decision process (MDP)-based optimal wake-up schemes are presented in [69] for network-operated femto BSs overlapping with a macro BS. In order to wake up the right femto BSs, which serve the extra traffic load and still lead to efficient energy usage, call traffic load and user localization within the macro-cell information are required. In the absence of the traffic localization information, the femto-BS wake-up problem is formulated as a partially observable MDP [69].

The last issue dealt with in the BS operation problem addresses the switching off mode entrance and exit stages, which are two important design stages in implementing the BS operation decision [70]. The switching off mode entrance stage specifies how the transition from the on (active) state to the off (inactive) state is implemented. If a BS is switched off very fast, the corresponding MTs may not be able to successfully execute their handover procedures and their calls eventually will be dropped. This could be due to a strong received signal from the BS that servesthe MT, which prevents the MT from hearing signals from nearby BSs. Hence, if the BS that an MT is associated with is suddenly switched off, the latter will not be able to synchronize and connect to another active BS. Another reason is the maximum number of handovers that can occur simultaneously towards a new BS, due to the limited signalling channel capacity. Hence, a progressive switching off operation, that is, referred to as BS wilting [70], can be used, as shown in Figure 2.2. In BS wilting, the BS transmission power is progressively halved until the BS is switched off. During this process, the MTs associated with the wilting BS initiate a handover process to the neighbouring BSs and the BS switching off procedure is suspended in case of unsuccessful handover of MTs. On the contrary, the switching off mode exit specifies how the transition from the off (inactive) state to the on (active) state is implemented. A BS that is switched on too fast can generate a strong interference to MTs in service. As a result, a progressive switching on process, that is referred to as BS blossoming [70], can be used as shown in Figure 2.2. In BS blossoming, the BS transmission power is progressively doubled until the BS is switched on.

2.1.2 MT Radio Interface Sleep Scheduling

Similar to BS on–off switching (dynamic planning), an MT with a bursty or low traffic load can save energy by switching off its radio interface from time to time. An appropriate on–off switching (sleep) schedule design for the MT radio interface varies based on whether the MT establishes communications on the downlink [57, 71–75], uplink [43] and [76] or both links [77].

For downlink communications, two categories can be identified for the MT radio interface on–off switching mechanism based on whether a traffic-shaping technique is employed or not [57, 71–75]. In the absence of traffic–shaping techniques, the MT radio interface switching off decision (in case of a bursty or low traffic load) is based on the unavailability of data packets for the MT at the serving BS. Consequently, the MT switching on–off (sleep) schedule specifies the switching off intervals and switching on instants for the radio interface based on the data availability. At a switching on instant, the MT checks if there are any packets available for it at the serving BS. In the absence of data packets, the MT enters a switching off interval; otherwise, the MT keeps its radio interface active to receive the available packets. During the MT sleep interval, all incoming data packets are buffered at the BS until the next MT switch on instant. On the one hand, a long sleep interval can enhance the MT energy savings; however, it also increases the packet-buffering delay at the BS until it is received by the MT. Furthermore, buffer overflow at the BS will result in discarding future incoming data packets for the sleep MT. In addition, unnecessarily switching on the MT radio interface to check for data packet availability at the BS buffer leads to MT energy losses. Thus, the main research objective in this case is to design a sleep schedule for the MT radio interface that maximizes its achieved energy saving while reducing the buffering delay of data packets available at the BS. One approach in designing such a schedule is by modelling the MT radio interface as a server that assumes repeated vacations [57, 71], as shown in Figure 2.3. Following this queueing model, analytical expressions can be derived for the expected number of sleep intervals until a data packet is available for the MT at the BS. Using these analytical expressions, myopic optimization problems can be formulated to minimize the energy consumption rate of the MT while achieving an acceptable message response time performance, where the message response time is defined as the time interval from the arrival time of an arbitrary message (data packet) at the BS to the time the message (data packet) leaves the system (BS) after service completion [57]. In addition to myopic optimization techniques, dynamic programming can be used to design a sleep schedule that minimizes a cost function consisting of a weighted sum of the MT energy consumption with radio interface on–off switching and a target performance metric (e.g. the buffering delay at the BS for the MT when its radio interface is switched off) [71]. Besides queueing models coupled with myopic and dynamicoptimization techniques, a Llyod-max algorithm can be used to design a sleep schedule that specifies the switching on instants for the MT radio interface [72].

The main limitation with the works in [57, 71] and [72] is that if the packet inter-arrival duration of the application is too small, the MT will not be able to switch off its radio interface to provide an acceptable QoS performance. Moreover, the MT consumes a significant amount of energy to switch on its radio interface. Every time the MT finds a single data packet available at the BS buffer, an interruption signal is triggered by the MT radio interface to activate the data bus and central processing unit (CPU) of the MT. Consequently, the MT will not be able to enter into a deep sleep state if it experiences frequent interrupts, and hence only a small amount of energy will be saved. Therefore, traffic-shaping techniques are employed to enable a longer idle duration for an MT. In this context, the idle duration denotes the interval during which an MT is not receiving any data packets. The traffic-shaping technique can be implemented by the MT by buffering the incoming data packets at its radio interface for a short period, without activating the data bus and CPU of the MT. Then, the data packets are released as a burst, which reduces the interruption-triggering events, and therefore, more energy is saved [73]. For transmission control protocol (TCP)-based applications, an alternative approach can be adopted by the MT, where the BS is forced to send data packets in bursts, and hence, it enjoys a longer idle duration, by announcing a zero congestion window size. Thus, the data packets are buffered for a longer period at the BS, until an appropriate window size is announced by the MT to allow the BS to release the data packets in bursts [74]. While the aforementioned traffic-shaping research deals with a single-user environment, the main goal in a multiuser scenario is to schedule the sleep intervals of the radio interfaces for different MTs to satisfy their target QoS and achieve energy saving by switching off the radio interfaces of MTs for a long enough duration [75]. An MT stores in its buffer sufficient data to satisfy its QoS and then switches off its radio interface for energy saving while the BS serves another MT. The MT activates its radio interface only when the data available at its buffer are insufficient to satisfy the required QoS.

For uplink communications, in addition to adapting the physical layer parameters as in controlling the MT transmission power and modulation/coding schemes, an MT can further save energy by switching on and off its radio interface. In [76], it is argued that different parameters such as the packet arrival rate and packet delay constraint have a significant impact on the practicality of employing such a switching approach. An on–off switching mechanism can be employed for energy saving at MTs for applications with small packet arrival rates and/or large packet delay constraints. In these scenarios, the research objective is to jointly adapt the power control, modulation and coding schemes (MCS) and switching on and off strategies of the MT radio interface to save energy in accordance with the stochastic traffic and channel conditions (i.e. no a-priori knowledge of traffic arrivals and channel conditions). In such a case, an MT is able to switch off its radio interface and hold data packets in its buffer to transmit them in bursts in better channel conditions. In addition to saving energy, the transmission mechanism should avoid an overflow event at the MT buffer and satisfy the required QoS in terms of data packet delay. An MDP problem can control the data packet transmission throughput (and hence, the amount of buffered data packets at the MT), resulting bit error probability, and MT radio interface state (switch on or off) to balance energy saving with QoS guarantee (i.e. minimizing data packet delay and avoiding buffer overflow) [43].

A general model for MT radio interface sleep scheduling is presented in the context of bidirectional communications in [77]. In this scenario, incoming downlink traffic does not suffer from BS-buffering delays during uplink transmissions, since the MT radio interface is already active. Thus, a finite general Markov background process can be used to model both the uplink activity and downlink traffic to derive analytical expressions for the buffer occupancy and downlink packet delay statistics [77]. These expressions are useful in developing an efficient on–off switching mechanism for the MT radio interface for both uplink and downlink communications.

2.1.3 Discussion

Following the above review, BS on–off switching (dynamic planning) aims to exploit spatial and temporal fluctuations in the call traffic load to achieve energy saving. Consequently, adopting static call traffic models in the switching schedule design (i.e. to determine the switch off and wake-up instant decisions) and/or performance evaluation, as in [65], is not realistic. The call traffic load models should capture the joint spatial and long-term temporal fluctuation behaviours [12, 63]. On the contrary, traffic models [34, 35] that reflect joint spatial and short-term temporal call-level fluctuations are incapable of assessing the daily switching schedule performance due to time-varying traffic demands. In addition, the traffic models that reflect only the long-term (as in [63] and [66]) or short-term (as in [40]) temporal call-level fluctuations fail to exploit the spatial dimension of the problem, and areunrealistic for performance evaluation studies in large-scale networks with multiple BS sites. For BS power consumption models, both static and dynamic components, as in (1.11) and (1.13), should be accounted for, which is the case for the algorithms developed in [12, 35] and [65]. On the contrary, the power consumption models that assume constant transmission power, as in [34, 40, 64] and [66], neglect the transmission power scaling associated with the call traffic load and represent unrealistic models. Overall, the reported solutions in Section 2.1.1 aim to minimize the network energy consumption, which is somehow similar in concept to maximizing the energy consumption gain given in (1.24). However, this expression does not assess the network gain (in terms of transmitted power as in (1.25) or network coverage as in (1.26)) versus the incurred cost (in terms of the network-consumed power). The reported solutions minimize the network energy consumption while satisfying a target performance metric. For BS on–off switching solutions, the trade-offs are based on admission quality requirements (i.e. network coverage and call blocking) as in [12, 34, 40] and [65]. Few works account for mobile user trade-offs [35]. From a practical perspective, a solution should account for both network and mobile user trade-offs to better serve the users who required QoS [63, 64]. Green solutions that adopt BS on–off switching report 25–50% energy savings [14, 78].

Similarly, MTs can save energy by switching off their radio interfaces during idle periods of bursty traffic. Thus, static traffic models that assume a fixed number of backlogged data packets ready for transmission [76] are not realistic to determine the MT idle periods, and therefore, will not be useful in developing practical sleep schedules for the MTs. The practical traffic models should reflect the packet-level short-term temporal fluctuations [43, 57, 71–75] and [77]. While some solutions account for both active and idle power consumption values [57, 71, 72] and [74] and reception power consumption [75] and [77], such solutions do not account for the MT circuit power consumption component. Both transmission and circuit power consumptions should be accounted for [43] and [76]. Yet, these models assume fixed MT circuit power consumption and neglect the dynamic circuit power component as described in (1.22) and (1.23). The reported solutions in Section 2.1.2 minimize the MT energy consumption while accounting for the mobile user trade-offs. However, such a modelling approach overlooks the network capacity limitations, for example, in terms of available bandwidth, which may lead to call blocking. Therefore, the proposed solutions should account for both network and user trade-offs.

2.2.1 Scheduling for Single-Network Access

In this technique, a mobile user receives the required resources from a single wireless access network at a time. In the literature, two system models are adopted for single-network access. The first model deals with a single network that covers a given geographical region, and it is referred to as a homogeneous wireless medium. The second model assumes the availability of multiple networks with overlapped coverage in the geographical region, and it is referred to as a heterogeneous wireless medium.

In the homogeneous wireless medium, the network operator assigns radio resources to MTs in a way that reduces the total power consumption of its BSs. This objective can be achieved by minimizing the BS transmission power while providing acceptable QoS performance for the MTs, a technique that is referred to as the margin-adaptive strategy [10]. One approach to implement the margin-adaptive strategy adopts a score-based scheduler. For instance, in an OFDMA system, the BS calculates a score for every radio resource block to be assigned to MT [79]. Such a score ensures that the BS consumes the least transmission power by allocating the resource block to MT . In addition, the score promotes a fair resource assignment among MTs, since a penalty function is included based on the number of already assigned resource blocks for MT . A low-score reflects a more desirable resource block. Fairness issues are also studied in [80] following a proportional rate constraint that ensures each user eventually receives a specific proportion of the system throughput. Admission control policies are also used to implement a margin-adaptive strategy, where a new session (call) is admitted into the system as long as the sub-frame energy in an OFDMA-based BS is kept below a certain threshold [81]. Furthermore, a margin-adaptive strategy can be implemented following a discrete rate adaptation policy that controls both the transmission rate and power according to the channel conditions to maximize the achieved energy efficiency for a target bit error rate [82]. Similarly, a channel-driven rate and power adaptation strategy can be implemented by jointly adapting the MCS and transmission power to optimize the trade-off between goodput and energy efficiency [83]. Moreover, a margin-adaptive strategy can be implemented via resource scheduling among MTs based on their traffic delay tolerance [38]. Delay-tolerant traffic (e.g. video and data) can be opportunistically served during periods of good channel conditions (i.e. soft real-timeservice). A drawback associated with the margin-adaptive strategy is that it requires CSI knowledge to allocate the transmitted power, which further necessitates the use of pilot symbols. Such pilot symbols will incur some energy consumption. In the literature, two approaches can be used for pilot energy assignment [10], namely the constant single pilot energy and the constant total pilot energy. In the former approach, each pilot maintains the same energy level independent of the number of pilot symbols. Thus, the larger the number of pilot symbols, the more accurate the CSI, and yet higher energy is consumed. The later approach allocates fixed energy to all pilots, which leads to reduced energy per pilot for a larger number of pilots, and to inaccurate CSI.

On the contrary, in a heterogeneous wireless medium, energy can be saved by assigning MTs to the BSs that reduce energy consumption for the set of network operators with BSs that assume overlapped coverage [32]. Moreover, in such a heterogeneous environment, each BS may choose between two modes of operation, namely point-to-point and point-to-multipoint mode. Thus, the problem can be decomposed into two sub-problems, namely the BS selection and BS operation mode selection. While the work in [32] controls the transmission power only through the BS operation mode selection, a joint BS selection and power control mechanism is proposed in [84] to associate MTs to BSs with overlapped coverage with the aim to minimize the BS transmission power to reduce the interference among different communication channels. In addition, data offloading techniques can be adopted to improve energy efficiency in a heterogeneous wireless medium. In particular, through mobility prediction and using the pre-fetching feature, data traffic can be offloaded from cellular networks to WiFi hotspots and femto cells [85]. Consequently, delay-tolerant traffic can be downloaded when mobile users are close to the WiFi access point or femto cell rather than using the macro cell [86]. Overall, data offloading can be either network- or user-driven [87]. Various factors, such as user mobility, backhaul throughput, data size and WiFi and/or femto-cell densities affect the energy efficiency performance when data offloading is adopted [85].

Similarly, MTs can save energy by appropriate uplink radio resource scheduling, based on the network multiple access scheme. In the literature, various energy-efficient mechanisms are proposed for OFDMA-based networks [28, 29, 42] and [46]. The proposed mechanisms mainly improve energy efficiency through sub-carrier allocation, power control and joint sub-carrier allocation and power control [29]. Centralized and decentralized architectures can be adopted to implement the radio resource allocation mechanisms [28, 46]. In a centralized architecture, the BS in each cell jointly performs sub-carrier allocation, modulation order adaptation and power control for the MTs in the uplink. In a distributed mechanism, given a sub-carrier assignment, an MT adjusts its modulation order and transmission power to optimize its own energy efficiency. In a multicell environment, multicell interference should be taken into account while designing an energy- efficient uplink resource allocation scheduling [29, 46]. In addition to sub-carrier allocation and power control, energy efficiency is maximized for OFDMA-basednetworks through dynamic carrier aggregation [42]. Although an MT served by all carrier components will enjoy an enhanced throughput, its energy consumption also increases. In a dynamic carrier aggregation technique, an MT is assigned to the queue of a given carrier component that is referred to as the primary carrier component (PCC). Whenever the queue of a given carrier component is empty, it helps other carrier components through aggregation, and therefore, it is referred to as the supplementary carrier component (SCC). Two mechanisms can be employed for SCC assignment [42]. The first mechanism aggregates all SCCs to support the PCC with the longest queue. The second mechanism orders PCCs according to the queue length, and SCCs are circularly allocated to the ordered PCCs in a round-robin manner.

For time division multiple access (TDMA)-based networks, energy efficiency is maximized for a set of MTs via opportunistic transmission [44]. A scheduler is considered at the BS to select an MT for transmission and determine its transmission rate. The problem complexity is reduced by decomposing it into two tasks. The first task is a user scheduling sub-problem that opportunistically selects an MT for transmission, based on the channel conditions and backlog information. The second task specifies the transmission rate for the selected MT to minimize the transmission power by transmitting data packets in queue such that the average delay constraint is satisfied with equality.

2.2.2 Scheduling for Multi-Homing Access

Currently, the wireless communication medium is a heterogeneous environment with overlapped coverage due to different networks. In this networking environment, MTs are equipped with multiple radio interfaces. Using the multi-homing capability, an MT can maintain multiple simultaneous associations with different wireless access networks. In addition to enhancing the achieved data rate by bandwidth aggregation, the multi-homing service can improve the energy efficiency of network operators and mobile users as MTs experience different channel conditions and bandwidth capabilities over their different radio interfaces.

By supporting multi-homing services, different network operators can reduce the transmission power of their BSs. The reason behind this reduction can be explained using the concept of power–rate curve, which can be graphically divided into two regions [33]. In the first region, the transmission power increases slowly with the growth of the data rate, while in the second region the transmission power increases dramatically with the data rate. Thus, a multi-homing data rate threshold, , can be specified to enable multi-homing transmission if the required data rate is higher than [33]. The data rate multi-homing threshold relies on the ratio of the channel gain between the MT and the BSs of different networks. Moreover, the optimal transmission data rate fromeach BS can be determined to maximize the networks' energy efficiency. In addition, cooperating BSs can control their transmission power by following a semi-Markov decision process (SMDP) to minimize the total power consumption of the BSs under a target QoS constraint at the MTs [36].

Similarly, MTs can improve their energy efficiency via the multi-homing service. In this case, an MT specifies how many and which BSs will be selected for multi-homing, according to the required data rate and the channel parameters of available BSs [23]. In order to reduce the associated complexity, the resource allocation problem is decomposed into two sub-problems. The first sub-problem determines which BSs will be selected for multi-homing, while the second sub-problem specifies the optimal transmission rate for each selected BS. For a constant data rate service, the energy efficiency maximization task is equivalent to the MT total power consumption minimization problem. Different from [23], the authors in [31] deal with energy efficiency maximization for a variable data rate using power allocation in a multi-homing service.

2.2.3 Scheduling with Small-Cells

Small-cell (e.g. pico and femto cells) have a radio coverage in the range of tens to a few hundred metres [88]. Consequently, the division of a macro-cell into several tiers of smaller cells replaces a long transmission range with a short transmission range because of the close proximity between small-cell BSs and MTs [88]. The small-cell power consumption is expected to reach approximately 5 W by 2020 [89]. Thus, an improved energy efficiency can be achieved by small-cell deployment. In [88], an expression is provided for the possible power gain resulting from the macro-cell splitting into smaller cells. It is shown that for an ideal free space propagation channel model, the achieved power gain satisfies , which means that cell splitting should not be implemented in this case. On the contrary, in a non-ideal propagation environment, and it increases with the number of small-cells, that is, the power gain improves with the number of deployed small-cells. However, it should be noted that the BS power consumption model in [88] does not capture the BS-embodied energy as in (1.20). When the BS-embodied energy is accounted for, there is a limit on the number of small-cells that can be deployed to improve energy efficiency. In the literature, different configurations have been adopted for the small-cell deployment, as shown in Figure 2.4. The cell-on-edge deployment mainly distributes the small-cells around the edge of a macro cell to serve the cell-edge users. On the contrary, in the uniform deployment, the small-cells are uniformly distributed across the macro cell. In [89], it is shown that the cell-on-edge deployment leads to a significant reduction in the network energy consumption as compared with the uniformly distributed configuration, due to the lower transmission power for cell edge users.

The main challenge for adopting a cell-splitting approach is the associated inter-tier interference. This is due to the limited available radio resources. As a result, the macro-BS radio resources are shared among the small-cells. Multicell processing can be employed to mitigate the resulting interference [88]. Thus, multiple BSs within a cluster exchange CSI and users' data to serve MTs and eliminate the associated interference. Using the gathered information, beam-forming techniques are applied to minimize the total transmission power while ensuring a certain signal-to-interference plus noise ratio (SINR) for different MTs. Besides multicell processing (and in the presence of both co-tier and cross-tier interference), admission control with QoS guarantee plays a vital role in mitigating interference, where a joint radio resource allocation mechanism can be employed among the multitier networks [41].

2.2.4 Relaying and Device-to-Device Communications

Another approach to reduce the transmission distance and hence achieve energy saving for BSs in the downlink and MTs in the uplink is relaying. In this regard, two types of relays can be employed [7, 14]. The first type is based on fixed relay stations, which are defined as network elements (repeaters) that store and forward the data towards the destination. The second type utilizes MTs as relays and it eliminates the cost of installing fixed relays in the network, yet it increases the system complexity.

Fixed relays are very useful in cell coverage extension and in reducing the power consumption of BSs (in the downlink) and MTs (in the uplink) due to a short transmission range. However, in this regard, the fixed relay stations should not be confused with the small-cell deployment. A small-cell mainly acts as an independent BS that decodes the mobile users' information and passes the decoded information together with the signalling information to the network operator via wired backhaul links [90], as shown in Figure 2.5. On the contrary, the relays mainly forward the user information from MT to BS and vice versa. Two types of fixed relays can be distinguished according to the 3GPP LTE-advanced and IEEE 802.16j standards, as shown in Figure 2.5. A type I relay can help an MT, which is located out of the coverage area of a given BS, to access that BS, while a type II relay can help an MT within the coverage area of a given BS to improve its service quality with reduced power consumption.

Overall, two research directions can be identified in the context of relaying for green networking when fixed relays are employed. The first research direction deals with the optimal placement of fixed relay stations to promote energy efficiency in wireless networks. The optimal relay placement for green networking is mainly affected by several key parameters including the distance between the relay station and the nodes (BSs and MTs), the radio propagation environment and line-of-sight conditions, the relay height and the relay coding scheme (e.g. amplify-and-forward and full/partial decode-and-forward). The authors in [91] propose a geometrical model for energy-efficient relay placement while accounting for the aforementioned factors. In addition, these authors identify the maximum cell coverage of a relay-assisted cell and the average cell energy consumption. It has been shown that (i) for a given relay coding scheme, an optimal relay location exists and the energy efficiency performance rapidly degrades away from this location and a more advanced coding scheme is required to maintain a good performance and (ii) there exists an optimal relay location for which increasing the cell coverage has a minimal impact on the average energy consumed per unit area. In addition to optimizing the relay position, the work in [92] jointly optimizes the relay position and its serving range for energy-efficient operation. The second research direction mainly deals with relay assignment and radio resource allocation for energy-efficient operation. It has been shown that the maximum benefit of the system can be achieved if the single best relay is selected for a specific source–destination pair [93]. Hence, for a set of source–destination pairs, the objective is to assign the optimal relays (and their respective radio resources, e.g. power and sub-carriers) to the source–destination pairs to enhance the network energy efficiency [94].

Installing fixed relays incur additional infrastructure, operational and maintenance costs for the network operators. In addition, each single relay consumes power for data forwarding and hence high power consumption is expected in a wireless network with dense relay deployment [95]. Consequently, employing MTs as mobile relays is a more appealing solution for network operators. In this context, MTs with good channel conditions can forward the data between BSs andother MTs in both directions (uplink and downlink). The main research challenge is how to motivate rationale (selfish) MTs to act as relays while relaying in turn will cost them additional power consumption. One direct approach to stimulate MTs to act as mobile relays is through a payment system. Hence, a buyer–seller market scenario might be adopted between the source nodes (MTs in uplink or BSs in downlink) and the relaying MTs. In order to reach an agreement on the amount of utilities a buyer pays and the set of radio resources the seller offers, a double auction game theory can be used [96].

In addition to using MTs as relay nodes, D2D communications can be employed to reduce the transmission range and hence achieve energy saving in wireless networks. While the MT only forwards the data between source (BS or MT) and destination (MT or BS) nodes in mobile relaying, MTs in close proximity directly communicate with each other. In this context, three types of D2D communications can be distinguished, namely in-band underlay, in-band overlay and out-band communications, respectively. In in-band underlay, both the cellular and D2D communications use the same resources. Hence, the main objective is to allocate resources among cellular users and D2D users in a way such that D2D users do not create interference to the cellular users. In [97–100] and [101], power efficiency is maximized through joint power allocation and mode selection for the MTs. Hence, MTs can choose between cellular and D2D communications, and the mode that maximizes power efficiency for the wireless network is selected. Such a mode selection results in a binary decision, which leads to a mixed-integer program formulation due to the real-valued nature of the power allocation problem and the binary nature of the mode selection problem. Consequently, heuristic algorithms are proposed to reduce the associated computational complexity [100 102]. In in-band overlay D2D communications, dedicated resources are assigned to cellular and D2D users. In [103], the authors proposed a BS-assisted D2D in-band overlay communication algorithm that can enhance the network energy efficiency. In particular, in the peer discovery phase, instead of relying on the MT beacon signals (which incur power consumption), BSs assist the D2D users to identify their peers. Furthermore, radio resources (e.g. power and bandwidth) can be allocated to cellular and D2D users in a more energy-efficient manner and via a less complex approach. In out-band D2D communications, the D2D communications take place over a separate band than the cellular band, for example, WiFi band. For instance, in [104], the authors proposed to form clusters for the cellular users who are in close proximity for WiFi communications. MTs coordinate their D2D communications over the cellular radio interface while exchanging their data over the WiFi direct interface. High energy efficiency is achieved as MTs can select one of the two modes: cellular or D2D communications.

2.2.5 Scheduling with Multiple Energy Sources

Various scheduling techniques have been proposed in the literature to deal with the presence of multiple energy sources [26, 37, 105–114] and [115]. The objective of these works is to simultaneously control transmission power and select the energy source that minimizes the total energy consumption. For network operators, multiple energy sources address the availability of different electricity retailers [26, 105], on-grid and green (renewable) energy [37] and different (complementary) renewable sources [106–113] and [114]. For MTs, multiple energy sources consider the availability of multiple batteries at the MT [115].

In an electricity market liberalization model, electricity retailers compete with each other to achieve the highest individual profits by adjusting the electricity price offered to users in different regions [26]. The electricity prices offered by different retailers change frequently to reflect the variations in the cost of energy supply, a strategy which is referred to as real time pricing. For a set of electricity retailers, a Stackelberg game can be formulated, where each retailer provides the real time price to maximize the own profit, and the network operator determines how much electricity to procure from each retailer to power on its BSs and achieve the lowest call blocking with the least monetary cost [26]. In [105], the optimal amount of energy to be procured from each retailer is specified via evolutionary algorithms (i.e. genetic algorithm and particle swarm optimization), which are shown to outperform the deterministic algorithm [26] because of the random nature of the evolution process. In addition to the presence of multiple electricity retailers, it is argued that the BSs of future cellular networks will be powered by both on-grid and green (renewable) energy (e.g. solar and wind energy) [37]. Hence, hybrid energy systems are expected to power the future BSs, where a combination of renewable and grid energy sources is utilized, as shown in Figure 2.6. Complementary renewable energy sources can be employed as well. If the power grid is absent, that is, the BS is not connected to the power grid (and hence, the controller 2 in Figure 2.6 does not exist), the BS is powered only by the renewable sources. An energy-harvesting battery should be used, as shown in Figure 2.6, to overcome the intermittent nature of renewable energy sources. With such a hybrid energy system, the objective is to optimize energy utilization in such networks by maximizing the green energy utilization and saving of on-grid energy. In this case, network designers are faced with the following two main concerns [37]: (i) how to optimize the usage of green energy at different time slots to accommodate the temporal dynamics of the green (solar) energy generation and the call traffic load and (ii) how to accommodate the spatial dynamics of the call traffic load with the objective of maximizing the green energy utilization by balancing the green energy consumption among BSs through cell size adjustment. While the aforementioned works assume the presence of on-grid energy, the long-term objective is to power BSs in appropriate locations using only a combination of complementary renewable sources (e.g. wind in winter and solar in summer) [106]. Moreover, cooperative techniques enable different BSs (networks) to share (trade) their green power, whenever possible,with each other for a sustainable and energy-efficient network operation [107].

In order to use renewable energy sources, renewable energy generation and storage should be investigated. As renewable energy sources are intermittent, energy storage units are deployed to address this limitation. Thus, the harvested (solar, wind) energy is stored in a battery with finite capacity before it is used in data transmission [108 109]. In this context, the energy replenishment process and the storage constraints of the rechargeable batteries should be taken into account while designing energy-efficient transmission strategies [110]. Two constraints should be considered at the energy-harvesting battery [111]. The first constraint ensures that the energy drawn from the battery is almost equal to the energy stored in the battery, a condition which is referred to as the causality constraint. The second constraint ensures that the energy level at the battery does not exceed a maximum level to avoid battery energy overflow. Consequently, storage sizing is very important to guarantee a sustainable energy at a reduced monetary cost. Moreover, BSs have to adapt their data transmission to the energy available at a particular time instant [112 113]. Therefore, more studies are needed to minimize the overall power consumption of BSs through on–off switching at a low call traffic load or through scheduling and node cooperation [114] at a high call traffic load to reduce the required energy generation and battery storage capacity. A very important aspect of green communications is to consider the environmental dimension of the proposed solution. For selecting an appropriate energy supply (i.e. electricity retailer and/or renewable energy source), it is necessary to guarantee that the associated emission cost is below a target level. The emission cost, in kg/h, related to the BS power consumption is given by Bu et al. [26]

2.1

where and are constants that depend on the pollution level of the electricity retailer.

For MTs that adopt a pulsed discharge profile, the battery is able to recover some charges during the interruptions of the drained current (i.e. no transmission period). Thus, an improved battery performance can be achieved. This phenomenon is referred to as the recovery effect. In order to promote the recovery effect and enhance the battery performance (and hence improve energy efficiency), a package of multiple batteries can be used and a scheduling policy can be developed to efficiently distribute the discharge demand among the multiple batteries connected in parallel [115].

2.2.6 Discussion

The majority of research works that investigate green communication solutions at a high traffic load assume static traffic models for radio resource scheduling and performance evaluation [10, 26, 28, 29, 32, 33, 46, 79, 80, 84, 88] and [105]. Very few works in the literature employ traffic models that reflect long term (as in [81] and [89]) or short term (as in [38, 41] and [42] for call-level and [44] for packet-level) temporal fluctuations. Also, few works assume traffic models that capture spatial fluctuations in traffic load [37] and [36]. Spatial and temporal traffic models should be employed for performance evaluation of green resource-scheduling algorithms. Spatial traffic models are useful in evaluating the algorithm performance in large-scale networks, while temporal models are important to investigate the associated signalling overhead, which may jeopardize the energy-saving benefits, if high overhead is expected. Moreover, many references account only for transmission power consumption [10, 29, 32, 33, 38, 41, 42, 81, 84 88] and [89]. Both transmission and circuit power consumption should be considered [26, 28, 36, 37, 44, 46, 79 80] and [105]. However, the aforementioned models do not account for dynamic circuit power consumption, as depicted by (1.22) and (1.23). In addition, BS transmission power consumption should scale with the traffic load as expressed in (1.11) and (1.13). Also, for small-cell and multi-tier deployment, both the operation and embodied energy should be accounted for as in (1.20) to avoid misleading conclusions. While some works aim to minimize energy consumption, reference [79] aims to maximize an energy consumption gain expression similar to (1.24). Furthermore, the works in [28, 29, 33, 42, 46] and [80] aim to maximize an energy efficiency expression similar to (1.27), (1.28) or (1.29). Such an expression provides a better indication of the performance in terms of the achieved gain (the resulting data rate) versus the incurred cost (the energy consumed). Almost all reported solutions aim to minimize the energy consumption or maximize energy efficiency, while maintaining a satisfactory performance that balances the mobile user operation. The works in [26] and [105] aim to balance the network operator objectives. In practice, an effective solution should account for both network operator and mobile user [41].

Green solutions that adopt small-cell deployment reports up to energysavings when combined with BS on–off switching [07, 14] and [78]. Other green solutions that employ renewable energy to power BSs report up to power savings [78]. In what concerns the green solutions that exploit relay deployments, they report 5–20% energy savings [14]. Finally, green solutions based on D2D communications report an improvement of 20–100% in terms of power efficiency [116].