1 Dynamic Cell Expansion with Self-Organizing Cooperation Weisi Guo, Timothy O’Farrell Department of Electronic and Electrical Engineering University of Sheffield, United Kingdom Email: {w.guo, t.ofarrell}@sheffield.ac.uk Abstract—This paper addresses the challenge of how to reduce the energy consumption of a multi-cell network under a dynamic traffic load. The body of investigation first shows that the energy reduction upper-bound for transmission improving techniques is hardware-limited, and the bound for infrastructure reduction is capacity-limited. The paper proposes a novel cell expansion technique, where the coverage area of cells can expand and contract based on the traffic load. This is accomplished by switching off low load cellsites and compensating for the coverage loss by expanding the neighboring cells through antenna beam tilting. The multi-cell coordination is resolved by using either a centralized controller or a distributed self-organizing-network (SON) algorithm. The analysis demonstrates that the proposed distributed algorithm is able to exploit flexibility and performance uncertainty through reinforced learning and improves on the centralized solution. The combined energy saving benefit of the proposed techniques is up to 50% compared to a reference deployment and 44% compared with alternative state-of-the-art dynamic base-station techniques. I. I NTRODUCTION A. Background and Review In the past decade, the growth in mobile data traffic has also seen a corresponding increase in the energy consumed by cellular networks. Traditionally, operators have met any increase in data traffic by deploying more cells or purchasing increased bandwidth. However, such solutions result in increased energy consumption and operating costs. In order to limit the ecological impact of the wireless communication industry and be financially competitive, there needs to be a change in strategy. Energy efficiency of wireless networks has attracted significant research attention. The characterization of energy- and spectral-efficiency relationships in noise-limited channels was theoretically analyzed in [1]. In terms of transmission techniques: the transmission energy efficiency of single cell MIMO and multi-cell CoMP [2] has been investigated. However, a shared caveat in higher-order MIMO is that the overhead energy consumption scales faster than the capacity improvement [3]. This generally leads to an increase in transmission energy efficiency, but a reduction in overall energy efficiency. Whilst schedulers showed promising transmission energy savings, the overall energy saving when overhead is accounted for, is very low (∼ 5%). In terms of architecture research, fixed wireless relays [4] demonstrate increased energy- and cost-efficiency [5] compared to increasing cell density. For dynamic architectures that varies with the offered traffic rate, passive sleep mode and off-loading to access-points has been considered [6]. Techniques that reduce the number of active cell sectors and antennas were considered in [6], [7], [8]. In order to maintain coverage and reduce the number of active cells, changing the coverage pattern of cells based on traffic load was proposed in [9], [10]. To the best of our knowledge, how cells in this dynamic configuration can compensate for potential coverage holes, whilst minimizing energy consumption, haven’t been investigated. B. Proposed Solution The rationale for sleep mode is that reducing the transmit power of a cell can only save a certain amount of transmission energy, whereas putting an entire cell-site in sleep mode can save most of its total energy expenditure. An illustration of a classical cell-site layout is shown in Fig. 1a, and the cell expansion scheme is shown in Fig. 1b and c. The technique dynamically reduces the amount of active wireless infrastructure as a function of the traffic load, and can dramatically reduce the energy consumption of the network. The challenge of multi-cell coordination (centralized and distributed) has been considered in [9], [10], but to the best of our knowledge, existing research has not tackled the fundamental challenges of: • How to resolve contention between cells that can both enter sleep mode? • When a cell enters sleep mode, its original coverage area suffers both strong interference and an increased pathloss degradation to compensating cells. How can this be resolved? Furthermore, existing solutions have lacked an integrated approach to research. In order to achieve a realistic implementation of sleep-mode and cell-expansion on sectorized basestations, a combination of techniques need to be considered. The novel contribution of our research is that it resolves the aforementioned challenges by employing an integrated combination of the following: dynamic antenna-beam tilting, frequency reuse, distributed inter-cell coordination and cooperative transmission. The benefit of this approach is that it is able to expand the cell coverage with greater spectral efficiency and achieve this with a low-complexity distributed algorithm. Furthermore, the paper compares this solution with centralized coordination and alternative low energy research techniques. 2 Active BS Active BS θref Sleeping BS Compensating BS θinner θouter θouter θinner Outer Cell (800 MHz) Compensation Zone Normal or Inner Cell (2.6 GHz) (a)Reference Deployment: 3-Sector Cell-Sites with 2x2 SFBC MIMO Fig. 1. Illustration of Deployments: a) Reference Deployment, b) Proposed Deployment, and c) Proposed Deployment in Cell Expansion Mode. II. S YSTEM M ODEL A. Cell-Site Model The paper considers an OFDMA based Long-TermEvolution (LTE) multiple-access system. The paper defines a cell-site or base-station (BS) as an entity with many cell sectors, where each sector is effectively an independent cell from the user’s perspective. A single homogeneous layer of NBS BSs are deployed, each with Nv vertical sets of Nh horizontal sectors, and each sector has Na antennas. This yields a total of Ns = Nv Nh Na antennas per BS, and a total of NBS Ns antennas in the multi-BS radio-access-network (RAN). In this paper, we consider the following types of network architectures with the same number of antennas: • • (c) Central Cell-Site Switched Off, Neighbouring Cells Expand (b) Proposed Deployment: 6-Sector Cell-Sites with 1x2 MIMO Reference: 3-Sector BS with 2x2 Space-FrequencyBlock-Coding (SFBC) MIMO, as recommended by 3GPP standards [11] and used to benchmark performance [12]. There are Ns = 6 antennas per BS and an illustration of the deployment is shown in Fig. 1a. Proposed: 6-Sector BS with 1x2 SIMO with MaximumRatio-Combining (MRC), which has the same peak power consumption as the reference deployment. There are Ns = 6 antennas per BS and an illustration of the deployment is shown in Fig. 1b. The benefit of this deployment is that it has the flexibility of expanding and contracting the cell coverage by adaptively tilting the antenna beam of the outer and inner sectors, as shown in Fig. 1c. B. Downlink Throughput Model For a certain user i that is of a distance di,k from the serving-cell k, the downlink received SINR (γi,k ) is calculated as a function of the BS transmit power (P ), BS antenna gain (A), and AWGN power (n), and fading gain (H): γi,k = −αi Hi Kdi,k Ai P , PNBS Ns −α n + j=1,j6=k Hj Kdi,j j Aj P (1) where the user experiences interference from up to NBS Ns other co-frequency cells. The parameters and their values are defined in Table. I. The generic sectorized BS antenna pattern employed in the cells is given as [11]: (Ab −min[12( Ai (ϑi,k ) = 10 |ϑi,k −θk | 2 ) ,Am ])/10 θ3dB , (2) where ϑi,k is the angle between the user i and the horizontal plane of the BS k, and θk is the antenna down-tilt angle of BS k. The other antenna parameters are: θ3dB = 75◦ for azimuth and θ3dB = 20◦ for elevation, Am = 25dBi, and the antenna bore-sight gain is Ab = 17.6dBi. The paper defines each BS as being able to employ a certain operational strategy (sk,t ∈ S). This strategy can refer to sleep mode, antenna tilt, or any host of transmission techniques. The precise details in the context of this paper is defined later in Section VI. C. Traffic and Load The traffic demanded by each user is assumed to be full buffer and a round-robin scheduler is employed. The downlink throughput is obtained by using the SINR (1) and the appropriate LTE adaptive modulation and coding scheme. The aggregate throughput achieved in a cell k is the sum of the throughput achieved by all its users at time t: RBS,k,t,s = NUE X i=1 RUE,i,k,t,s , (3) 3 TABLE I S YSTEM PARAMETERS Parameter Operating Frequency BS Coverage Radius Interference Model Pathloss Exponent Pathloss Constant AWGN power Multipath Fading Shadow Fading Var. UE antenna Height BS antenna Height BS antenna pattern BS transmission Propagation Model Scheduler BS Transmit Power BS Radio-head Eff. BS Overhead Power Backhaul Power Symbol rcell α K n H σs2 A λ P µ POH PBH the load, the more efficient the power amplifier. Existing literature has largely assumed that radio-head efficiency is constant with the load, which means the radio-head power consumption scales linearly with load. However, research in [13] has shown that µ can vary between 50% to 10% for a micro-BS. By fitting an expression to the data in [13], the p value of µ is approximately given by µ(Lk,t,s ) ≈ µpeak Lk,t,s , where µpeak is the efficiency achieved at maximum load (Lk,t,s = 1). Value 2600 MHz 500 m 19 BS Wrap Around 3.67 4.6 × 10−4 6 × 10−17 W Rayleigh 9 dB 1.5 m 20m (2) 2x2, 1x2 MIMO 3GPP WINNER [11] Round Robin 20W 10-50% [13] 60W 50W where the value of each user’s throughput (RUE,i,k,t,s ) is a function of the achieved SINR (γ) and the scheduler function (round robin in this case). This is obtained from the previously mentioned simulation. The paper defines the traffic rate (Rtraffic,k,t , bits/s) offered to each cell as the aggregate traffic intensity demanded by users (circuit- and packet-switched packets). This varies with the cell concerned k, and the time t, which is taken in 15 minute samples for simulation results presented later in the paper. Therefore, at any particular instance, a traffic rate is offered to the BS, and the load (L) experienced by the BS with strategy s can be defined as: Lk,t,s = Rtraffic,k,t + max[(Lk,t−1,s ) − 1, 0], RBS,k,t,s (4) where in order to not incur outage, Lk,t,s ≤ 1. If an outage occurs, the remaining unsatisfied load (max[(Lk,t−1 )−1, 0]) is added to the load of the next transmission time interval (TTI). D. Power Consumption Model A general power consumption model for a BS with Ns antennas is comprised of a load-dependent and a loadindependent function [3]: P Lk,t,s + POH ) + PBH µ P p ≈ Ns ( Lk,t,s + POH ) + PBH . µpeak PBS,k,t,s = Ns ( (5) In the load dependent part, P is the transmit power and µ is the radio-head efficiency. In the load-independent part, POH is the over-head power (base-band and cooling) and PBH is the backhaul power. Data on the power consumption values can be found in Table. I. As the load Lk,t,s in the BS varies, the radiohead efficiency µ also varies. Generally speaking, the greater III. E NERGY S AVING B OUNDS In this section, the paper considers the energy saving bounds, which serves as motivation for sleep mode control. The paper considers upper-bounds achieved if the capacity of the RAN’s BSs have been improved (or equivalently if the offered traffic rate has been decreased). There are 2 separate paths to energy reduction: • Reduce transmission power and save transmit energy. • Switch-off BSs and compensate for coverage loss through cell-expansion. The energy saved by the test system, in comparison with the reference is defined as the energy-reduction-gain (ERG): PT PNBS k=1 PBS,test,k,t,s (Ltest,k,t,s ) . (6) ERG =1 − Pt=1 PNBS T t=1 k=1 PBS,ref.,k,t,s (Lref.,k,t,s ) Assuming the throughput in the test system has been improved by a factor of ρs (due to strategy s), compared to the reference system. The resulting energy saved compared to the reference system is given in Lemma 1 of the Appendix as: ERGhardware → 1 , 1+Ω Ω = µpeak POH + P PBH Ns (7) for: ρs → inf . It can be seen that the parameter Ω is strictly positive and is a ratio of hardware power consumption parameters. Typically the value of Ω = 1.7, with figures taken from Table. I. The resulting ERG’s upper-bound is 30-40% and the saving can be described as being hardware-limited. The paper now considers an ideal case, whereby the amount of active infrastructure can flexibly vary with the traffic load. This concept can be achieved through sleep-mode operation and the performance is improved with the proposed cellexpansion technique. The resulting ERG given in Lemma 1 of the Appendix is: ERGcapacity → 1, for: ρs → inf . (8) It can be seen that the ERG is entirely a function of the parameter ρs , and can theoretically achieve 100%. In reality, the energy-consumption of BSs in off- or sleep-mode is nonzero, and 100% energy reduction can not be achieved. As the parameter ρs is an improvement in BS capacity, the energy saving can be described as being capacity-limited. IV. C ELL E XPANSION : C ELL L AYOUT A. Cell-Site Structure and Frequency Bands The rationale of cell-expansion as a low energy solution is that it switches off a percentage of the BSs, as well as 4 to Fig. 1, our simulation results found that for a homogeneous set of BSs, a shallow tilt of θnormal,k = 0 − 1◦ yielded a small improvement in the cell-edge throughput, whilst the greatest mean throughput was achieved with a down-tilt of θnormal,k = 8◦ . By applying adaptive antenna beam-tilting to the proposed inner-outer BS setup, it was found that the inner cell should have a tilt of θinner,k = 12◦ and the outer cell a tilt of θouter,k = 3◦ . When the coverage of cell expands, this was found to change to θinner,k = 9◦ and θouter,k = 0◦ . This yielded a significantly improved throughput profile compared to alternatives. 60 Ref. Cell, 2.6GHz (8◦) Ref. Cell, 2.6GHz (0◦) Inner Cell (2.6GHz, 12◦) & Outer Cell (2.6GHz, 3◦) Mean Cell Throughput, Mbit/s 50 Inner Cell (2.6GHz, 12◦) & Outer Cell (0.8GHz, 3◦) 40 Normal Coverage Region Compensation Region Normal Coverage Region 30 20 BS in Sleep Mode 10 0 BS in Compensation Mode BS in Compensation Mode 0 500 1000 1500 Distance, m Fig. 2. BS Configuration and Optimal Antenna Down-Tilt Angles for Cell Expansion. All antenna down-tilt angles are optimized for maximum cell throughput. reducing the interference of the network in that process. This way, a capacity-limited ERG can be achieved. The BS can operate in the following modes: 1) Normal: in normal mode, the BS adjusts its antenna down-tilt as to maximize the throughput of users within its traditional coverage area. 2) Sleep: in sleep mode, the BS is switched off and its users are passed to neighboring BSs, which must expand to compensate for the coverage loss. 3) Expand: in expansion mode, the BS expands by tilting its outer cell’s antenna towards the contracted cell area. The inner cell’s down-tilt is also adjusted to maximize the throughput of users within its traditional coverage area. As shown in Fig. 1b, the mechanism requires inner and outer cells to be implemented. This is achieved by deploying an additional vertical set of sectors and adopting different antenna down-tilt angles. The outer sectors transmit with in the 800MHz band and the inner sectors transmit in the 2.6GHz band. The 800Mhz channel allows the pathloss attenuation to be reduced by approximately 10dB, which leads to an improved signal strength under expanded cell coverage. This is preferred to increasing the transmit power, which increases energy consumption. The results in Fig. 2 show that by employing an inner and outer sector with 2.6GHz and 800MHz, the achievable throughput is improved for both the normal cell coverage region and the compensation region. B. Dynamic Antenna Beam-Tilting Existing work [14] has shown that dynamically tilting the antenna can improve the throughput of the cell. The optimal down-tilt for the reference deployment was found to be between 6-8◦ , depending on the antenna pattern employed. For antenna arrays, the vertical pattern is typically very nonuniform, and therefore optimization is performed via bruteforce search within a reasonable range of (2-20◦ ). Referring V. C ENTRALIZED C OORDINATION A. BS Groupings As mentioned previously, the actions of one BS, directly affects the performance of all other BSs, i.e., if a BS enters sleep mode, the interference it causes to other BSs is reduced, increasing their throughput. By allowing a fixed set of BSs to take known actions, the network performance changes in a way that is predictable. In centralized coordination, the aim is to achieve an approximately deterministic performance, by restricting which BSs can go into sleep mode to a limited set of patterns. How this pattern is determined at any particular instance, is determined by a central Sleep Mode Management (SMM) controller. This sub-section explains how the BSs in a RAN are grouped together to form the previously mentioned patterns. Each BS is associated to a set of other BSs via a logical grouping identity. As shown in Fig. 3a and b, each BS has two identities: Ifull = {A, B, C} and Ipart = {D, E}. The identities are pre-determined at the network-planning phase. Two centralized sleep-mode strategies (Sc ) are proposed: • Full Compensation Strategy (1/3), Sfull : each sleeping BS’s previous coverage region is compensated by 3 neighboring BSs’s expanded outer sectors (Fig. 3a). Up to 1/3 of BSs in a network can be in sleep mode (upperbound energy saving of 33% compared to reference). Therefore, within the Full Strategy Sfull , there are 3 different sub-strategies to choose from, each related to a group of BSs. • Partial Compensation Strategy (1/2), Spartial : each sleeping BS’s previous coverage region is compensated by 2 neighboring BSs’s outer sectors (Fig. 3b). Up to 1/2 of BSs can be in sleep mode (upper-bound energy saving of 50% compared to reference). Therefore, within the Partial Strategy Spartial , there are 2 different sub-strategies to choose from, each related to a group of BSs. As previously defined, each strategy refers to a category or grouping of BSs. Each BS within that category is seen as independent as the inter-BS distance is large. Therefore, each BS inside a category can decide on its own, whether it can enter sleep mode based on the predicted traffic load. That prediction process is explained in the sections below. B. Contention Between Groupings A dilemma arises when multiple strategies can be implemented, and contention occurs between compensating BS 5 Compensation Zone Compensation Zone Compensation Zone D B B 1/3 Rexpand Rexpand D E 1/2 2 Rsleep Rsleep 1/2 C 1/3 1 D E Rnormal B 2 1 E Rsleep 1 2-BS Partnership Rexpand A C 1 D C 2 D D (a)Centralized Structure: Full Compensation Pattern (1/3) BS Type A in sleep, Type B in compensation, and Type C in normal mode (b) Centralized Structure: Partial Compensation Pattern (1/2) BS Type E in sleep, and Type D in compensation mode (c) Distributed Structure: 2-BS Coordination (1/2) BS Pair-Up: Type 2 in sleep, and Type 1 in compensation mode Fig. 3. Cell Expansion Strategies for: a) Centralized Full Compensation Strategy (regular,1/3); b) Centralized Partial Compensation Strategy (regular,1/2); c) De-centralized Partner Selection Strategy (irregular). TABLE II C ENTRALIZED C OORDINATION DATA FOR SMM: BS T HROUGHPUT FOR VARIOUS OPERATION MODES . Operational Mode Normal Mode (Rnormal ) Full Expansion (Rexp,1/3 ) Full Sleep (Rsleep,1/3 ) Partial Expansion (Rexp,1/2 ) Partial Sleep (Rsleep,1/2 ) Theory 33.3 Mbit/s 16.6 Mbit/s 10.0 Mbit/s 22.2 Mbit/s 8.1 Mbit/s Simulation 30.2 Mbit/s 15.1 Mbit/s 12.3 Mbit/s 20.0 Mbit/s 7.3 Mbit/s groups that wish to enter sleep mode. In order to account for this potential conflict, a network-wide co-ordination must be performed so that the SMM employs: the strategy (Sc ) that allows the highest energy saving, whilst the offered traffic load is still met. This is determined by the Minimum Energy Traffic Constrained (METC) algorithm presented in (9) below and in Fig. 4a. The METC criteria is given as follows: min[ Sc NBS X PBS,k (Lk,t,s )t], subject to: Lk,t,s ≤ 1, ∀k, (9) k=1 which means the algorithm selects the strategy Sc that minimizes the instantaneously energy expenditure, subject to the traffic condition is satisfied for all BSs at any time t. The expected or predicted performance for each strategy is obtained using Table II. The cell-edge throughput per BS (R) are averaged over all cells and time samples. The simulation results are from the Monte Carlo simulator, and the theoretical results are from a stochastic geometry framework given in the Appendix, whereby the interference-limited capacity at a distance r0 for a cell density of Λ is: Z +∞ (10) R̄s = exp − Λπr02 Q(ζ, α) dζ. 0 As explained in Fig. 4a, the centralized coordination process is as follows: 1) Each BS uploads its current load condition to the SMM; 2) Each BS is grouped with corresponding BSs according to its logical identity; 3) For each strategy (Sc ), predict the resulting energy and load performance based on Table II; 4) Find lowest energy strategy using the METC criteria in (9); 5) Feedback operational mode commands to BSs. The energy saving results for the centralized coordination in comparison with other energy saving solutions is presented in Section VII. In terms of algorithm complexity and mutual information exchanged, each BS has to upload its load information to the SMM every time step t. Typically traffic is measured on 15 minute basis, and this seems a reasonable time step for sleep mode as well. The complexity is in the order of NBS , with a multiplication factor of the number of strategies. For an urban environment RAN, this can be up to 100 BSs with 5 fixed BS grouping combinations (strategies), and this is clearly a highcomplexity scheme. The remaining coordination challenge is: can a lower-energy and lower-complexity solution be devised, based on distributed coordination and machine-learning? VI. D ISTRIBUTED C OORDINATION A. Partner Selection In this section, the paper proposes a distributed cell expansion algorithm that allows the BSs to select partners and coordinate cell expansion between themselves, without knowledge of the modeling environment. This is illustrated in Fig. 3c. However, there is an added challenge: without knowledge of what action other BSs outside the partnership might take, the expected throughput is undetermined. Depending on how many neighbor BSs are in sleep mode, the distributed decision process can cause unexpected performances. Similar to the centralized coordination, BSs need to be grouped to create 6 Load information given to BS Upload Load information to SMM Is Load value lower than sleep mode threshold?** *BS identities are predefined and 2 types exist (full and partial). Find resulting load and energy saving for each group of BSs in sleep mode (strategy) ** **Use look-up table that predicts the resulting traffic load, for both fulland partial-expansion strategies. NO All BSs are in active mode YES NO Is Load value lower than expansion mode threshold?** Request Sleep Mode YES Is there collision between the BS requests? Request Expansion Mode NO METC Algorithm: Do any of the request combinations meet the traffic load requirements? Select highest energy saving strategy and group of BSs to sleep. YES Select highest energy saving BSs to sleep. a) Centralized SMM Controller • BSs achieve their requests NO All BSs in partnership are in active mode **Can also use machinelearning to decide on action. b) Distributed SMM Controller Decision Model for Sleep Mode via: a) Centralized Cell Expansion; b) Distributed Cell Expansion. a sleep-mode and expansion-mode association between BSs. The paper considers a smaller group of 2-3 BSs, and proposes two different forms of partner selection: • Not Available Mode YES YES Fig. 4. NO Exchange information with partner BS* Closed Loop: Load Threshold Feedback Group BSs based on identity* METC Algorithm: Do any of the grouping s meet the traffic requirements? *BS partners are selected from nearest neighbour BS list. See section on Partner Selection for more details. Load information given to BS Random Neighbor (RN): each BS establishes a temporary partnership with a neighboring BS. Each BS picks its neighbor in an ad-hoc manner based on whether it will enter sleep mode or compensation mode, and which neighbors are available. There is no guarantee that a successful partnership can be negotiated. Fixed Neighbors (FN): each BS establishes a fixed relationship with 2 neighboring BSs, to form a triangular 3BS relationship. Coordination is achieved between the 3BSs to decide which one of them enters sleep, expansion and normal mode. Due to the fixed nature, there is guarantee that a partnership can be found. B. Distributed Decision Process 3) Partners check if their mutual requests are compatible, i.e., sleep-mode and expansion-mode is compatible, whereas sleep-mode and not-available is not compatible. If compatibility is met, the BSs have their requested actions mutually accepted; 4) If the compatibility is not met and more than one strategy is possible, for each strategy (Sc ), predict the resulting energy and load performance based on previous experiences. The paper now formally defines the various parameters used in the machine-learning process and analyze the benefits of having open- and closed-loop threshold feedback. The paper formally defines the following: • • • Once a partnership has been formed, the transition rules between operational modes can be defined (as shown in Fig. 4b: 1) Each BS makes an action decision based on the current traffic load and a decision threshold. The resulting action can be to request entering sleep-mode, make itself available for expansion-mode or to signal being notavailable for expansion; 2) Each BS then communicates this information to its partner BSs; • • • Environment: the network containing all BSs and subject to a certain offered traffic rate that has a uniform temporal and spatial distribution. Agents: the BSs, where the observed BS is k ∈ NBS . State: the current load L experienced by the agent concerned. Action: each agent can take an action a ∈ A to enter an operational mode (normal, expansion or sleep). As a consequence of this action, the BS will reach a new state (load, L). Strategy: the strategy to take υ ∈ Sd that affects the probability of taking the request action a ∈ A to reach a sleep mode state. Feedback: by taking an action that leads to a state, the BS has an interference impact on the network containing 7 100 70 Aggressive 60 50 40 Sleep Mode Saving 30 Conservative 10 90 ERG - Open Loop, =0.6 ERG - Closed Loop Outage Pb. - Closed Loop Exploration (Δ = 15) 80 Exploitation (Δ = 2) ERG - Open Loop, =0.6 70 Performance, % 80 70 Performance, % Performance, % ERG - Closed Loop Outage Pb. - Closed Loop 90 80 20 100 100 ERG - Open Loop Outage Pb. - Open Loop 90 60 50 ERG 40 Outage Prob. ERG 60 Outage Prob. 50 40 30 30 20 20 10 10 Radiohead Saving 0 0 0.2 0.4 0.6 0.8 1 Strategy, 1.2 1.4 1.6 0 0 500 1000 1500 0 500 1000 1500 Iterations(t) Iterations(t) (b) Closed-Loop with Random Greedy Algorithm with Exploration factor Δ = 2 and 15. (a) Open-Loop Distributed Coordination with Fixed Strategy, υ Fig. 5. a) Open-Loop Distributed Coordination with Fixed Strategy; b) Closed-Loop Distributed Coordination with Random Greedy Algorithm with Exploration factor ∆ = 2 and 15; simulation results only. • other BSs. Reward: the power saved (compared to peak) in the observed BS, at time t with state s: <k,t,s = 1 − • PBS,k,t,s (Lk,t,s ) . PBS,k,t,s (Lk,t,s = 1) (11) As shown in (5), the potential power saving when not in sleep mode is in the radio-head, and when in sleep mode is the total power. Punishment: the outage probability, at time t with state s: Rtraffic,k,t − RBS,k,t,s ℘k,t,s = . (12) Rtraffic,k,t Observation: the BS observes the reward and punishment of the previous action, as well as the resulting load on the observed BS at time t. Therefore, the observed BS trades-off the reward (energy saved) of entering sleep mode and the punishment (outage) of over-estimating the ability to meet the offered traffic rate. Furthermore, there is also a tradeoff between exploiting what the BS thinks is the best strategy based on existing knowledge and exploring the alternative strategies. • C. Open Loop Performance In open loop performance, the strategy is fixed for all BSs (υ), which means the probability that triggers a sleep mode request is fixed at any particular time instance (quasi-static traffic model). The results presented in Fig. 5a show that: • An over-aggressive BS (high υ) can enter sleep mode only to discover that it can not meet the offered traffic rate and cause an outage ot > 1. • An over-conservative BS (low υ) never enters sleep mode and saves only the radio-head energy (10%), but can generally guarantee an arbitrarily low outage rate. For a target outage rate of 5%, a maximum of 16% energy can be saved with a strategy of υ = 0.6. The advantage with open loop strategy is that only a single network-wide parameter (υ) needs to be tuned to optimize performance during a single time period. The disadvantage is that the network can not exploit local spatial effects and create local threshold parameters. D. Closed Loop Performance with Reinforced Learning In closed loop performance, individual BSs keep track of previous strategies’ performance results and picks a threshold based on this memory information. The previous reward (<k,t,s ) and punishment (℘) performances for a given strategy s over a period of T is: E[<k,s ] = T 1X <k,t,s , T t E[℘k,s ] = T 1X ℘k,t,s . T t (13) For small-finite-state problems, Markov Decision Processes (MDP) techniques such as Q-learning can be applied. Even when the state transition probabilities are unknown (as is the case here), they can be discovered using exploration. However, the challenge addressed in this paper concerns an infinite-state problem, where the BS can be in any state (any load) and an action can potentially lead to any other state. The Q-learning algorithm would take too long to discover the MDP. Instead, the agent can weigh the discovery process using Q-values (rewards) based on a soft-max distribution. This paper considers a Random Greedy Strategy with Boltzmann Exploration, similar to automata learning: • the agent (BS) always selects the estimated best strategy based on the expected reward and punishment from previous iterations. However, there is a probability (prandom ) that it will instead select a random alternative strategy that is not the estimated best strategy. • under the Boltzmann Exploration strategy, the strategy selection is based on a weighted probability biased in favor of likelihood to yield high rewards: prandom,k,t = P e E[<k,s ] ∆ k,s0 ∆ E[< s0 ∈Sd e ] (14) for: E[℘k,s ] ≤ outage threshold, subject to meeting the outage threshold and s0 represents all other strategies (s0 6= s) in the available set of 8 Antenna Reduction: number of active antennas reduce with traffic load [6], [7]. The results in Fig. 6 show that for an offered urban traffic rate of 10-120 Mbit/s/km2 , the reference power reduction technique can only save up to 27% energy. By employing antenna reduction, the energy saving is up to 33% compared to the peak consumption. By employing cell expansion (centralized with SMM), the energy saving is up to 64% compared to peak value. By assuming a uniform variation in the offered traffic rate intensity, the antenna reduction technique improves on the existing reference by up to 17% (mean 9%). The cell expansion technique improves on the reference by up to 50% (mean 19%), and up to 44% (mean 11%) over the antenna reduction technique. By assuming a realistic variation in the offered traffic rate intensity, the antenna reduction technique improves on the existing reference by up to 18% (mean 6%). The cell expansion technique improves on the reference by up to 50% (mean 10%), and up to 44% (mean 5%) over the antenna reduction technique. • 2800 2600 Power Reduction RAN Power Consumption, W/km 2 2400 2200 2000 MIMO Reduction 1800 1600 Sector Reduction Full Expansion, 1/3 1400 Reference (Sim.) Antenna Reduction (Sim.) Cell Expansion (Sim.) Reference (Theory) Antenna Reduction (Theory) Cell Expansion (Theory) 1200 1000 800 600 10 Partial Expansion, 1/2 20 30 40 50 60 70 80 90 100 110 120 RAN Offered Load, Mbit/s/km2 Fig. 6. Power-Capacity-Tradeoff for different deployment schemes under a varying traffic load. Simulation Results as symbols and Theory as lines. strategies s0 ∈ Sd . The parameter ∆ adjusts the level of exploration: small ∆ favors exploitation and large ∆ favors exploration. Generally, a large ∆ can guarantee asymptotic optimality. The results in Fig. 5b and c show that there is a tradeoff between exploiting what is already known (rapid convergence) and exploration (better asymptotic behavior). For a target outage rate of 5%, an asymptotic average of 12% energy can be saved with exploitive behavior (D = 2), with a convergence time of ∼200 iterations. This performance is worse than the optimal open-loop performance, where 16% ERG can be achieved with a 5% outage rate and a strategy of υ = 0.6. With greater exploration behavior (D = 15), an asymptotic average of 24% energy can be saved, with a convergence time of ∼600 iterations. However, the level of convergence (variance) is much greater than the exploitation case. The optimization of convergence speed and optimality is beyond the scope of this paper. The results have demonstrated that distributed coordination can in fact use the uncertainty in performance to its advantage, provided that a reinforced learning algorithm is employed and the tradeoff between learning and exploitation is adjusted to benefit asymptotic performance without leading to an unbearable convergence rate. This can be further analyzed in future work. In Section VII, the paper presents the energy saving results for the distributed solutions with different partner selection schemes and compares its performance with the centralized solution and other energy saving solutions. VII. C ELL E XPANSION R ESULTS A. Baseline Comparison The paper begins by examining how the RAN energy consumption scale down at low loads, given that an initial deployment is made to satisfy a high offered load (L = 1). The investigation considers the following methods for comparison: • Reference: transmit power reduces with traffic load (5); B. Centralized vs. Distributed Results Generally speaking, compared to the centralized algorithm, the lower-complexity distributed algorithm can achieve a comparable low-load saving (48%) and a higher high-load saving (20%). However, the distributed coordination requires tuning the learning rate. As mentioned previously in Section VI, the paper proposed two different forms of distributed cellexpansion: Random Neighbor (RN) and Fixed Neighbors (FN) partner selection. From the results in Fig. 7, it can be seen that the energy saving at low loads for distributed-RN is lower than the centralized solution. The energy saving compared to the reference is up to 39% (mean 21%). From the results, it can be seen that whilst the energy saving is between the centralized and distributed-RN solutions at high loads, the distributed-FN solution has a similar performance to the centralized solution at low loads. C. Cooperative Transmission A key challenge is how to improve the received SINR in the compensation region, which is limited by interference and increased propagation. For a user with an SINR of γi,k , it can either receive a single direct transmission or repeated transmissions from the NCoop,i,k compensating BSs. The loss 1 . The resulting SINR gain (G+ in bandwidth is NCoop,i,k i,k ) must be: G+ i,k > (1 + γi,k )NCoop,i,k − 1 . γi,k (15) If expression (15) is expressed in the context of the user positions: • Exponent Relationship: For users close to a serving cell and experiencing a high SINR, the cooperation gain NCoop,i,k −1 required would be: G+ . i,k > γi,k • Linear Relationship: For users in the compensation zone and far from a serving cell, the SINR is typically low. Using the binomial expansion approximation, the cooperation gain required would be: G+ i,k > NCoop,i,k . 9 load (Ltest,k,t,s = 2800 2600 ERG = 1 − RAN Power Consumption, W/km 2 2400 2200 1600 Distributed Advantage Irregular Coverage 1200 Reference Cell Expansion (Centralized) Cell Expansion (Centralized, CoMP) Cell Expansion (Distributed - RN) Cell Expansion (Distributed - FN) 1000 800 600 10 20 30 40 50 60 70 80 90 100 Ns ( µPpeak POH + CoMP Gain 1400 Ns ( µPpeak is: q Lref.,k,t,s ρs + POH ) + PBH p Lref.,k,t,s + POH ) + PBH (16) 1 → 1+Ω 2000 1800 Lref.,k,t,s ) ρs 110 120 RAN Offered Load, Mbit/s/km2 Fig. 7. Power-Capacity-Tradeoff for different schemes: reference, centralized cell-expansion, cell-expansion with CoMP, distributed cell-expansion. Simulation Results only. The results found show that up to 0.5 bit/s/Hz and 50% spectral efficiency gain can be achieved in certain regions of the switched off cell. However, cooperation in regions of high SINR can lead to 60% less in spectral efficiency. Therefore, a repetition cooperation scheme based on user positioning can be employed, similar to those devised in [2]. The results are presented in Fig. 7 and up to 28% spectral efficiency improvement can be achieved for full expansion strategy. VIII. C ONCLUSIONS This paper has considered the challenge of how to scale energy consumption with spatial and temporal variations in the traffic. The paper first demonstrated that the energy saving upper-bound of transmission based techniques is hardwarelimited, whereas sleep mode and deployment techniques is capacity-limited. The proposed cell expansion technique allows a higher number of base-stations to be in sleep mode. This is achieved by expanding the coverage of neighbouring base-stations. The results show that the centralized coordination algorithm can achieve a deterministic coverage pattern and a strong low-load energy saving (50%) and no highload saving (0%). Compared to the centralized algorithm, the lower-complexity distributed algorithm can achieve a comparable low-load saving (48%) and a higher high-load saving (20%). However, the distributed coordination requires tuning the learning rate. Furthermore, integration with cooperative transmission can further improve the baseline expansion performance by 28%. A PPENDIX A. Lemma 1: Energy Saving Bounds The hardware-limited energy reduction gain (ERG) of a reference system with load (Lref.,k,t,s ) and a test system with PBH Ns where Ω = µpeak . The conditions are for an initial P L = load of Lref.,k,t,s = 1 and an ideal load reduction to ref.,k,t,s ρs 0. The capacity-limited energy reduction gain (ERG) of a s reference system with (Ns ) and a test system with ( N ρs ) antennas is: p Ns P Lref.,k,t,s + POH ) + PBH ρs ( µpeak p ERG = 1 − P Ns ( µpeak Lref.,k,t,s + POH ) + PBH (17) 1 →1− , ρs for both systems being fully loaded (Lref.,k,t,s = 1), and assuming that sleeping basestations consume close to no energy. B. Lemma 2: Theoretical Capacity for Different BS Modes The paper employs recent developments in stochastic geometry to provide an approximate theoretical capacity performance for different cell expansion operational modes under centralized coordination. Stochastic geometry allows the interference-limited capacity to be found for a multi-cell network [15]. The major short-fall with the analysis is that no antennas are considered. Another key difference is that a random cell deployment topology is considered, as opposed to hexagonal deployment. The cell-edge capacity is defined as the capacity achieved by a single user location that is at a distance r0 from the 2 serving cell. For a cell density of Λ = 1/πrcell , the average interference-limited (n = 0) cell-edge capacity is [16]: Z +∞ R̄s = P(Rs > ζ)dζ Z0 +∞ = exp(−βr0α n(2ζ − 1) − Λπr02 Q(ζ, α))dζ (18) Z0 +∞ = exp(−Λπr02 Q(ζ, α))dζ for: n → 0, 0 where the Q-function is defined as follows: The Q(ζ, α) function is given by: Z +∞ (2ζ − 1)−2/α Q(ζ, α) = du 1 + uα/2 (2ζ −1)−2/α p π 1 − arctan( √ ) for: α = 4, = 2ζ − 1 2 2ζ − 1 (19) where α is the pathloss distance exponent. The integrals can be solved numerically using the Gauss-Korond technique. In the normal cell operational mode, the worst cell-edge user is at a distance r0 = rcell from the serving cell. The 10 achievable aggregate cell-edge capacity of a cell-site with Ns sectors each with Bcell bandwidth is: Z +∞ Rnormal = Ns Bcell exp(−Q(ζ, α))dζ = 33.3Mbits/s 0 (20) where we consider 10MHz band per sector with Ns = 3 outer sectors, the edge-throughput per BS is 33.3 Mbits/s. In the full compensation (1/3) cell operational mode, the worst cell-edge user is at a distance r0 = 2rcell from the serving cell (Fig. 3b). That is to say, a serving cell has to serve as far as the location of a sleeping neighbouring cellsite. The achievable aggregate cell-edge capacity of a cell-site with Ns sectors each with Bcell bandwidth is: Z +∞ Rsleep,1/3 = Ns Bcell exp(−4Q(ζ, α))dζ = 10.0Mbits/s. 0 (21) The compensating or expanding cell-sites have in turn, lost 3 sectors and therefore suffer a capacity loss of 12 . Given the the result in (20), the expanding cell’s capacity is therefore Rexp,1/3 = 16.6 Mbits/s. In the partial compensation (1/2) cell operational mode, √ the worst cell-edge user is at a distance r0 = 5rcell from the serving cell (Fig. 3c). The achievable aggregate celledge capacity of a cell-site with Ns sectors each with Bcell bandwidth is: Z +∞ Rsleep,1/2 = Ns Bcell exp(−5Q(ζ, α))dζ = 8.1Mbits/s. 0 (22) The compensating or expanding cell-sites have in turn, lost 1 sector and therefore suffer a capacity loss of 13 . Given the the result in (20), the expanding cell’s capacity is therefore Rexp,1/2 = 22.2 Mbits/s. Acknowledgement The work reported in this paper has formed part of the Green Radio Research Programme of the Mobile VCE. Fully detailed technical reports on this research are available to industrial members of the Mobile VCE. www.mobilevce.com R EFERENCES [1] Y. Chen, S. Zhang, S. Xu, and G. Li, “Fundamental trade-offs on green wireless networks,” in Communications Magazine, IEEE, vol. 49, no. 6, Jun. 2011. [2] T. A. Le, S. Nasseri, A. Zarrebini-Esfahani, M. Nakhai, and A. Mills, “Power-efficient downlink transmission in multicell networks with limited wireless backhaul,” in Wireless Communications, IEEE, Oct. 2011, pp. 82–88. [3] G. Auer, V. Giannini, I. Godor, P. Skillermark, M. Olsson, M. Imran, D. Sabella, M. Gonzalez, C. Desset, and O. Blume, “Cellular energy efficiency evaluation framework,” in Vehicular Technology Conference (VTC Spring), 2011 IEEE, May 2011, pp. 1–6. [4] W. Guo and T. O’Farrell, “Relay deployment in cellular networks: Planning and optimization,” in Selected Areas in Communications (JSAC), IEEE Journal on, vol. 30, Nov. 2012. [5] A. Furuskar, M. Almgren, and K. Johansson, “An infrastructure cost evaluation of single- and multi-access networks with heterogeneous traffic density,” in Vehicular Technology Conference, IEEE, May 2005. [6] O. Holland, V. Friderikos, and A. H. Aghvami, “Energy efficient cross band spectrum management for mobile operators,” in Global Communications Conference (Globecom), IEEE, Dec. 2010, pp. 2020–2040. [7] J. Christofferson, “Energy Efficiency by Cell Reconfiguration: MIMO to non-MIMO and 3-Cell Sites to Omni,” in Green Wireless Communications and Networks (GreenNet), IEEE Workshop on, May 2010, pp. 1–5. [8] W. Guo, J. Rigelsford, K. Ford, and T. O’Farrell, “Dynamic basestation antenna design for low energy networks,” in Progress in Electromagnetic Research C (PIER), vol. 31, Jun. 2012. [9] Z. Niu, Y. Wu, J. Gong, and Z. Yang, “Cell zooming for cost-efficient green cellular networks,” in Communications Magazine, IEEE, Nov. 2010, pp. 74–79. [10] X. Weng, D. Cao, and Z. Niu, “Energy-Efficient Cellular Network Planning under Insufficient Cell Zooming,” in Vehicular Technology Conference (VTC Spring), IEEE, May 2011, pp. 1–5. [11] 3GPP, “Further Advancements for E-UTRA Physical Layer Aspects (Rel.9),” 3GPP TR36.814v9, Technical Report, Mar. 2010. [12] Ericsson, “Summary of Downlink Performance Evaluation,” 3GPP TSG RAN R1-072444, Technical Report, May 2007. [13] K. Mimis, K. Morris, and J. McGeehan, “A 2GHz GaN class J Power Amplifier for Basestation Applications,” in Radio and Wireless (RWS), IEEE Symp. for, Phoenix, USA, Jan. 2011, pp. 5–8. [14] J. Niemela, T. Isotalo, and J. Lempiainen, “Optimum antenna downtilt angles for macrocellular WCDMA network,” in EURASIP Journal of Wireless Communications Networking, 2005, pp. 816–827. [15] M. Haenggi, J. G. Andrews, F. Baccelli, O. Dousse, and M. Franceschetti, “Stochastic geometry and random graphs for the analysis and design of wireless networks,” in Selected Areas in Communications (JSAC), IEEE Journal on, vol. 28, Sep. 2009, pp. 1029–1046. [16] S. Wang, W. Guo, and T. O’Farrell, “Two Tier Networks with Frequency Selective Surface,” in IEEE International Conference on High Performance Computing and Communications (HPCC), Jun. 2012. Weisi Guo received his M.Eng., M.A. and Ph.D. degrees from the University of Cambridge. He is currently an Assistant Professor at the University of Warwick and is the author of the VCEsim LTE System Simulator. His research interests are in the areas of self-organization, energy-efficiency, and multi-user cooperative wireless networks. Tim O’Farrell holds a Chair in Wireless Communication at the University of Sheffield, UK. He is the Academic Coordinator of the MVCE Green Radio Project. His research encompass resource management and physical layer techniques for wireless communication systems. He has led over 18 research projects and published over 200 technical papers including 8 granted patents.