Energy Efficiency

In order to study the EE performance of HetNets with massive MIMO, the total power consumption needs to be captured as well. The power consumption model at the MBS in each channel use is expressed as [28, 31]

P_{M BS} = 1 ω₁

 Pm

σ_am(1 − σ_{f eed}) + Pc+ P_bb

+ P_bh, (6.29)

where ω₁ = (1 − σ_DC)(1 − σ_{M S})(1 − σ_cool) and σ_am is the power amplifier efficiency. P_c is the circuit power consumption, i.e., Pc= N (pdac+ pmix+ pf ilt) + psyn, where pdac, pmix, pf ilt, psyn, σ_{f eed}, σ_DC, σ_{M S}, σ_cool are defined in Section 2.1.3 of Chapter I, respectively. We assume that P_bh is the backhaul power consumption for signalling exchange between a MBS and its SBSs, and as well as core network. Moreover, it has been shown in [28] that the loss factor of active cooling σ_cool and the feeder loss σ_{f eed} for the SBS can be typically negligible. From (6.29), the power consumption of the single-antenna SBS is defined as

PSBS = 1 ω₂

 P2

σ_am + p_dac+ pmix+ p_{f ilt}+ psyn+ P_bb



, (6.30)

where ω₂= (1 − σ_DC)(1 − σ_{M S}).

Without loss of generality, we assume that the system has unit bandwidth in order to simplify

Chapter 6: Massive MIMO Heterogeneous Networks

notation [31]. The theoretical analysis is studied at the typical user which is located at the origin [126]. The probabilities that the typical user is served by MBS0, {MBS0, SBS0}, and SBS₀ are defined as ∆_m, ∆_c, and ∆_s, respectively. For given K users associated with each MBS, the number of users served by only MBS0 is given by K1= _∆^K∆m

m+∆c, while the number of users served by both MBS₀ and SBS₀ is given by K₂ = _∆^K∆c

m+∆c. In every macrocell, the average number of SBSs located in the outer region is defined as ^λ2Pr[o∈A_λ ^out]

1 = ^λ2(∆_λ^c+∆s)

1 [127, 136].

Since each SBS is assumed to serve a user, then the number of users served by the SBSs is defined as K₃ = ^λ2(∆_λ^c+∆s)

1 − K₂ in every macrocell. Hence, the area spectral efficiency (ASE) (in bits/s/Hz/m²) of the whole network is obtained as

CASE ' λ₁(K1τm+ K2τc+ K3τs). (6.31)

We also assume that the MBS and SBS contribute main power consumption in the downlink [31, 69]. Thus, we formulate an optimization problem for maximizing EE as [26, 130, 154]

maximize

Pm,N,λ2

ηEE = CASE

P_AEC, (6.32)

subject to (C₁) : K ≤ N ≤ N_max, (C₂) : 0 < P_m ≤ P_m^max, (C₃) : λ^min₂ ≤ λ₂ ≤ λ^max₂ , (C₄) : C_ASE ≥ C_ASE^min , (C5) : τi≥ τ_i^min, ∀i ∈ {m, c, s}

where PAEC = λ1PM BS+λ2(∆c+∆s)PSBS is the area energy consumption (AEC) (in Joule/s/m²), C_ASE is defined in (6.31), P_m is the maximum MBS transmit power, τ_i^min is the minimum rate requirement, Nmaxis the available number of MBS antennas. C_ASE^min is the minimum requirement of the ASE, λ^min₂ and λ^max₂ are the minimum and the maximum SBS densities, respectively.

The minimum SBS density should be λ^min₂ = λ₁(K₂+ 1)/(∆_c+ ∆_s) due to K₃ ≥ 1. Next, we present an alternative method to solve problem (6.32). Firstly, we define the optimal value of

Chapter 6: Massive MIMO Heterogeneous Networks

Table 6.2: Bisection search algorithm for obtaining N^∗ 1. Input: Given N_min = K, N_max= N_avai, and accuracy .

2. Initialize: η^min_EE = ηEE(Nmin) and η_EE^max= ηEE(Nmax).

3. while |η_EE^max− η_EE^min| >  or (N_max− N_min) > 1

4. Ntemp = (Nmax+ Nmin)/2 and η_EE^temp= ηEE(Ntemp).

5. if (6.33) is feasible and ηEE(Ntemp+ 1) ≥ ηEE(Ntemp) do N_min ← N_temp and η_EE^min← η_EE^temp.

6. else Nmax← N_temp and η_EE^max← η^temp_EE . 7. return N^∗= (Nmax+ Nmin)/2.

8. Output: N^∗∗= min(N^∗, N_avai) and η_EE^opt = η_EE(N^∗∗).

each parameter (N, P_m or λ₂) for maximizing the EE when the other parameters are fixed as the following.

6.3.1 Optimal Number of MBS Antennas

The total power consumption of massive MIMO systems increases with N , which can sig-nificantly affect the EE. Therefore, it is important to determine the optimal number of MBS antennas that maximizes the EE. For given Pm and λ2, the problem of EE optimization (6.32) can be rewritten as

maximize

N η_EE subject to (C1) and (C5). (6.33) To solve problem (6.33), we adopt the bisection search algorithm (BSA). First, we initialize Nmin and Nmax, then we can determine η_EE^min = ηEE(Nmin) and η_EE^max = ηEE(Nmax), respec-tively. We define N_temp= (N_max+ N_min)/2 at every iteration, then we obtain η_EE(N_temp+ 1) and ηEE(Ntemp). If the conditions in (6.33) are satisfied, and if ηEE(Ntemp+ 1) ≥ ηEE(Ntemp), the algorithm updates N_min = N_temp, otherwise N_max = N_temp. The algorithm will be ter-minated whenever |η_EE^max− η_EE^min| ≤  or (N_max− N_min) = 1 is satisfied. Following that, the maximum overall EE is η_EE^opt = ηEE(N^∗∗) at the optimum value of the number of MBS anten-nas N^∗∗= min(N^∗, N_avai), where N^∗ = (N_max+ N_min)/2. The above steps of the BSA, named as algorithm 1, are summarized in Table 6.2. In the worst case, the computational complexity

Chapter 6: Massive MIMO Heterogeneous Networks

of the algorithm 1 is on the order of O(log₂(N − K) + 1) to find the optimal number of MBS antennas. The additional comparison in the algorithm is due to the fact that the algorithm needs to compare the two consecutive values in the array, as compared with the conventional bisection search algorithm.

6.3.2 Optimal MBS Transmit Power

For given N and λ₂, a solution for maximizing the EE means to wisely select a good transmit power level to use. The problem of EE optimization with respect to P_m is rewritten as

maximize

Pm

η_EE subject to (C₂) and (C₅). (6.34)

Following the similar approach, we modify the bisection search algorithm to solve problem (6.34) by initializing P_m^min = 0 dBm. All other steps are almost the same. When we use a step of 0.5 dBm for EE comparisons, in the worst case, the complexity of the modified algorithm is on the order of O(log₂(2P_m^max) + 1) to find the optimal values of P_m.

6.3.3 Optimal SBS Density

In the network with a certain MBS density, it is feasible to improve the area spectral efficiency performance by adopting high small-cell density [155]. However, the interference from the second tier and its power consumption will be increased. The goal of the following optimization problem is to maximize the EE while satisfying the demand of sum rate for the network with respect to SBS density. Thus, it can be formulated as

maximize

λ2

η_EE subject to (C₃), (C₄) and (C₅). (6.35)

Chapter 6: Massive MIMO Heterogeneous Networks

Table 6.3: EE maximization with the AO algorithm 1. Initialize: Randomly feasible sets (N, P_m, λ₂) and given accuracy  2. Repeat

3. Obtain N by using algorithm 1,

4. Obtain P_m by using modified algorithm 1, based on N defined in the step 3 and given λ₂, 5. Obtain λ2 by using modified algorithm 1, based on Pm and N defined in steps 3 and 4.

6. Until |η_EE(i + 1) − η_EE(i)| ≤  (or convergence) 7. Output: (N^opt, P_m^opt, λ^opt₂ )

To solve this problem, we also modify the algorithm 1. Particularly, we define κ as the small step for EE comparisons. If the conditions in (6.35) are satisfied, and if η_EE(λ^temp₂ +κ) ≥ η_EE(λ^temp₂ ), where λ^temp₂ = (λ^max₂ + λ^min₂ )/2, the algorithm updates λ^min₂ = λ^temp₂ , otherwise λ^max₂ = λ^temp₂ . The modified algorithm will return the optimal value of λ₂when |η_EE^max(λ^max₂ ) − η^min_EE(λ^min₂ )| ≤ .

We define J = d(λ^max₂ − λ^min₂ )/κe as the number of elements in the array. In the worst case, the complexity of the modified algorithm is on the order of O(log₂(J ) + 1) to define the optimal λ₂.

6.3.4 Alternating Optimization Algorithm

In the previous section, the optimal values of N, P_m and λ₂ for maximizing the EE are obtained independently for given parameters. However, the global maximal value of the EE can only be achieved by joint optimization. In the following, we adopt a standard AO algorithm as shown in Table 6.3 to achieve an optimum EE. This method is a practical solution because it has low complexity and it has been widely used in literature [26,112]. According to [113, Proposition 4] and [114, Lemma 2], the AO algorithm is locally convergent. In our EE maximization problem, the algorithm optimizes these parameters (N, Pm, λ2) sequentially until |ηEE(i + 1) − η_EE(i)| ≤ . It has been shown that the EE metric is a quasi-concave function of N in [31]

and [134], transmit power in [26] and [156], and SBS density in [157] for given constraints, respectively. Therefore, the achievable value of EE in the AO algorithm is non-decreasing in

Chapter 6: Massive MIMO Heterogeneous Networks every step, such as

η_EE(N⁽ⁱ⁾, P_m⁽ⁱ⁾, λ⁽ⁱ⁾₂ )

(e5)

≤ η_EE(N⁽ⁱ⁺¹⁾, P_m⁽ⁱ⁾, λ⁽ⁱ⁾₂ ) ≤ · · · ≤ η_EE(N⁽ⁱ⁺¹⁾, P_m⁽ⁱ⁺¹⁾, λ⁽ⁱ⁺¹⁾₂ ), (6.36)

where (e₅) is due to the fact that η_EE(N⁽ⁱ⁺¹⁾, P_m⁽ⁱ⁾, λ₂) = max

N η_EE(N⁽ⁱ⁾, P_m⁽ⁱ⁾, λ⁽ⁱ⁾₂ ) is the op-timal solution of N for the sub-problem (6.33). Moreover, the EE metric has a finite upper bound. Hence, the local convergence of the algorithm is guaranteed for any initially feasi-ble set (N, P_m, λ₂). To avoid local convergence, we randomly generate several initial sets and choose the one which results in the overall maximum EE. An output of the chosen ini-tial set via the algorithm is considered as an optimal solution, such as (N^∗∗, P_m^∗∗, λ^∗∗₂ ). Note that N^∗∗ is a real-valued number. To define an integer-valued solution of N^opt, we only need to take floor and ceiling operations on the real-valued solution N^∗∗ for achieving the max-imal EE (6.32). By defining Pm^opt = P_m^∗∗ and λ^opt₂ = λ^∗∗₂ , the final solution is given by (N^opt, Pm^opt, λ^opt₂ ). The actual convergence speed of the algorithm depends on specific value of accuracy  [114]. As compared with the Brute-force ES algorithm, it has been shown in Section 6.5 that the AO algorithm performs well and it is almost able to obtain the global optimum EE. In the worst case, the total computational complexity of the AO algorithm is on the order of d1d2(O(log₂(N − K) + 1) + O(log₂(2P_m^max) + 1) + O(log₂(J ) + 1)), where d1 is the number of initial sets and d2 is the number of iterations, to achieve an optimum EE.