2. System Model
As shown in
Figure 1, a multi-user two-tier heterogeneous network scenario is considered in the paper, which includes a single massive MIMO enabled macro-base station (MBS), multiple single-antenna small base stations (SBSs) and single-antenna users. The MBS and SBSs differ in transmit power, size, density and the number of antennas. Let
be the set of BSs, where index 0 denotes the MBS and the others are SBSs.
is the set of users. It is assumed that all users distribute uniformly within the coverage area of the MBS and each user is associated with only one BS at a time.
2.1. Ergodic Rate and Power Consumption
It is assumed the MBS employs large scale antenna array, and the number of active antennas is represented by . Due to the channel hardening effect introduced by massive MIMO, the channel fluctuations will be averaged. In addition, the user association is always assumed to be carried out in a very large time scale relative to the change of channel. Thus, we can use the ergodic rate as the performance metric for the network.
It is assumed the network is operating in time-division duplexing (TDD) mode and all the BSs share the perfect channel state information (CSI). Moreover, we assumed each BS share the same frequency band and all users associated with the same BS share the orthogonal frequency band. The MBS can serve multiple users on a given resource block, and
denotes the maximum number of downlink transmission streams. Then, with linear zero-forcing (ZF) precoding employed for eliminating the intra-cell interference, the downlink ergodic rate from BS
j to user
k can be lower bounded by Equation (
1) [
33,
34].
Here, represents the ergodic rate of user k when associated with BS j. denotes the transmit power from BS j. denotes the channel fading coefficient from BS j to user k. We note that the assumption of the perfect CSI can be relaxed by the equation . and denote the estimated channel variable and the estimated error variable. The estimated error reduces the network energy efficiency. However, we only analyze the relationship between the network energy efficiency and other influence factors, such as the spectrum efficiency, the number of antennas installed on the MBS and so on. Thus, the assumption of the perfect CSI has little effect on the analysis in the paper. is the variance of the additive white Gaussian noise (AWGN) with zero mean.
In the wireless downlink transmission system, the power consumed when the BS provides the service to the user is mainly considered, and the power consumption generated by the user receiving the signal can always be negligible. The power consumption at the BS mainly includes wireless transmission power consumption and static circuit power. Besides, it is necessary to consider circuit power of each antenna due to the installation of large-scale antenna matrix at the MBS. The transmit power of the MBS and SBSs can be calculated as Equation (
2) [
35].
Here,
denotes the circuit power per antenna,
denotes the power amplifier efficiency of the BS
j, and
denotes the static circuit power term independent of the antenna number. The total system power consumption can be expressed as Equation (
3).
2.2. Problem Formulation
In this paper, we consider the network EE as the objective function, and study the joint optimization problem of power control and user association in the scenario of massive MIMO enabled HetNets. The network EE is defined as shown in Equation (
4).
Here,
denotes whether the user
k is associated with the BS
j. If the user
k is associated with the BS
j,
, otherwise
.
,
. We consider the proportional fairness and define
as the proportional fair rate [
36]. Therefore, the EE defined in this paper can be understood as the network proportional fair throughput in the case of energy consumption per unit spectrum. Then, the energy efficient joint power control and user association optimization problem can be formulated as shown by Problem P0.
Here, C1 represents the maximum transmit power constraint of each BS. C2 shows the principle of user association in the network, that is, each user can be associated with only a single base station. C3, which constrains the variable of each user’s association, is a binary variable.
3. Energy Efficient Joint Power Control and User Association Optimization Algorithm
Problem P0 is a fractional and mixed integer nonlinear programming problem (FMINLP). Firstly, the objective function can be transformed into an integral expression with weight coefficient based on the Dinkelbach’s theorem and an iterative algorithm giving priority to the EE is proposed correspondingly. Then, the joint optimization problem is decomposed into power control and user association optimization subproblems. Both subproblems can be transformed into convex optimization problems based on the approximately iterative method and Lagrange’s decomposition dual method, respectively. Finally, the joint optimization algorithm is proposed through the interior point method and the sub-gradient method.
3.1. Optimal EE Iterative Algorithm Based on Dinkelbach’s Theorem
Considering load optimization of each BS and the simplification of the problem, Problem P0 can be firstly transformed into Problem P1 as follows.
Here,
represents the load of the BS
j in the network, that is, the number of users associated with the BS
j. The vector
denotes the load of the network. According to Lemma 1, it is obvious that Problem P1 can be transformed into Problem P2, with the objective function transforming from a fractional expression into integral one.
Here, denotes the weight coefficient and the range of the weight coefficient is . The weight coefficient can denote the degree of preference for energy consumption in the objective function of Problem P2.
Lemma 1. If and only if there is an optimal parameter for Problem P2, such that holds and Problem P1 obtains the optimal EE value .
Based on the Dinkelbach’s theorem [
37], Lemma 1 is proved in
Appendix A. In fact, Lemma 1 can be easily considered as the specific example of Dinkelbach’s theorem in this paper. Problem P1 is the fractional optimization problem where the functional is given by the ratio of two integrals. With Dinkelbach’s approach, the fractional optimization problem is transformed into an equivalent parametric family
of the optimization problem, where the ratio disappears and the functional is given by the weighted difference of the numerator and the denominator of the ratio. Finally, the optimization iteration algorithm can be designed as follows to solve Problem P1.
Algorithm 1 Optimal EE iterative algorithm |
Input: the maximum iteration number , the maximum tolerance error ; Output: the optimal solution ; Initialize: the weight coefficient of the first iteration , iteration number , the variable vector while and do Update ; Calculate , by solving Problem P2 (see Algorithm 2); Update ; end while Return and ;
|
3.2. Joint Optimization Algorithm
To solve Problem P2, we reformulate the problem by separating the individual variables
and
to decompose the joint optimization problem into two dependent Sub-Problems P3 and P4: the Lower level problem (Power control) and Master problem (User association optimization).
Here, . We can obtain the joint power control and user association optimization algorithm as follows. In addition, the convergence of the algorithm can be seen in the simulation.
Algorithm 2 Joint optimization algorithm |
Initialize: the weight coefficient , iteration number , the maximum tolerance error , the actual error , each BS’s transmit power while do Update ; According to , calculate the optimal user association vector by solving Problem P4; According to , calculate the optimal user association vector by solving the problem P3; Calculate ; end while
|
3.2.1. The Lower Level Problem: Power Control
Problem P3 is a non-convex optimization problem. In this paper, we approximate the objective function into a convex function in each iterative step. Then, Problem P3 can be transformed into a convex optimization problem. The transformation process is shown as follows.
Lemma 2. There exists , where . If only if , then the inequality can take the equal sign. The proof can be seen in Reference [38]. According to Lemma 2, the equation is obviously available, where .
Since the
log-sum-exp function in the objective function of Problem P5 is proven to be a concave function [
39], the non-convex optimization Problem P3 can be transformed into Problem P5. Note that Problem P5 is a convex optimization and can be solved by the interior point method.
Here,
. To sum up, the power control algorithm based on approximately iterative method (Algorithm 3) can be designed as follows in this paper. In Algorithm 3, the objective function increases monotonously with the iteration number increasing and finally converges to a fixed value, which can be seen in the simulation later. The gap between the solution of the proposed algorithm and the optimal solution are described in [
40,
41].
Algorithm 3 Power control algorithm based on approximately iterative method |
Initialize: iteration number , the maximum tolerance error , the flag of convergence , each BS’s transmit power while do Update ; According to , calculate and ; Calculate the optimal solution of Problem P5 by the interior point method; According to , calculate and ; Calculate ; Update ; end while
|
3.2.2. The Master Problem: User Association Optimization Problem
Problem P4 is a mixed-integer nonlinear programming (MINLP) problem, which can be converted into a convex optimization problem with the help of the Lagrange’s dual decomposition method. The Lagrange function of Problem P4 can be expressed as Equation (
21).
Correspondingly, the dual Problem P6 can be expressed as Equation (
22).
Then, Problem P6 can be decomposed into Equations (23) and (24).
In this paper, the binary variable can be replaced by a continuous variable in the interval [0,1] [
42], which is the main source of the gap between the proposed algorithm and the optimal algorithm. We analyze the gap later in the paper. Note that Problem P6 is a convex optimization problem and can be solved by the sub-gradient method. According to the KKT conditions, the optimal solution of user association and each BS’s load can be obtained by deriving the partial derivative
and
as shown in Equations (25) and (26).
In addition, parameters
and
can be calculated as shown in Equations (27) and (28).
Here, denotes . and represent iterative step size of and , respectively.
Algorithm 4 User association algorithm based on Lagrange’s dual method |
Initialize: iteration number , the maximum tolerance error , the flag of convergence , each BS’s transmit power while do Update ; According to Equations (25) and (26), update and ; According to Equations (27) and (28), update and ; Calculate ; end while
|
As described above, there may be a gap between the solution of the proposed algorithm and the optimal one, due to the conversion process of the variable from 0–1 integer to the interval [0,1]. Next, the gap analysis of the proposed algorithm is shown as follows.
Suppose
and
are the solution of the proposed algorithm and the global optimal solution, respectively.
and
are the values of corresponding objective functions, respectively. It can be proven that there will be
, where
. The proof can be seen in
Appendix B.
3.3. Complexity Analysis
The asymptotic complexity of the proposed algorithm is analyzed in this section. In Algorithm 1, the calculation of each user’s association variable and each BS’s transmit power entails operations. Correspondingly, the computation in Algorithms 2–4 also calls operations. Suppose Algorithms 1–4 need , , and iterations to converge. The total complexity of the proposed algorithm is thus . Compared with the exhaustive search for the joint optimization problem, which has a worst case complexity of , the proposed algorithm has a much lower complexity. Moreover, , , and are small enough, as can be seen in the simulation.
4. Simulation and Analysis
In this paper, the scenario of massive MIMO enabled HetNets is considered, where the coverage radius of MBSs and SBSs are 500 m and 50 m, respectively. The network simulation parameters are shown in
Table 1 [
13]. Assumed that all the SBSs and users distribute evenly within the coverage area of the center MBS and each SBS is at least 40 m away from the center MBS. The number of antennas installed on the macro-BS and the maximum data streams supported by the macro-BS are supposed to be 100 and 20, respectively. In addition, the maximum transmit power of the MBS and each SBS are 43 dBm and 23 dBm, respectively. All the simulation results are obtained by 5000 Monte Carlo runs.
Due to the characteristic of channel hardening in the massive MIMO system, the effects of fast fading can be eliminated. Then, we can suppose the channel fading only includes path loss, shadow fading and frequency selective Rayleigh fading. The path loss model can be found in
Model A.2.1.1.2-3 for outdoor RRH or hotspot area model 1 [
43], as shown in
Table 2, where the unit of
d is km. The Rayleigh fading channel gains are modeled as i.i.d. unit-mean exponentially distributed random variables. The variance of the lognormal shadowing from the associated BS to each user is considered to be 10 dB.
Figure 2 shows the convergence in term of different initial values of each BS’s transmit power for Algorithm 3 versus the number of iterations, where
and
, respectively. The simulation setting of the number of users and the number of antennas installed on the MBS are 30 and 100, respectively. The ordinate in
Figure 2 represents the change number of the convergence flag
in Algorithm 3. Obviously, we can find Algorithm 3 converges after a limited number of iterations irrespective of different initial values of each BS’s transmit power. This result, together with the previous analysis, ensures that the proposed Algorithm 3 is applicable in the massive MIMO enabled HetNets.
Figure 3 and
Figure 4 show the convergence in terms of different initial values of users for Algorithms 1 and 2, respectively. The simulation setting of
and the number of antennas installed on the MBS are
and 100, respectively. The ordinate in
Figure 3 and
Figure 4 represent the change number of the convergence flag
in Algorithm 2 and the weight coefficient
in Algorithm 1, respectively. Obviously, we can find Algorithms 1 and 2 converge after a limited number of iterations irrespective of different initial values of users. This result, together with the previous analysis, ensures that the proposed Algorithms 1 and 2 are applicable in the massive MIMO enabled HetNets.
Figure 5 shows the relationship between the network EE and the weight coefficient
in Algorithm 2, where the numbers of users
K has different values. The simulation setting of the number of antennas
M and the initial value of each BS’s transmit power
are 100 and
, respectively. The vertical axis of
Figure 5 is the value of the network EE calculated by Algorithm 2 with different values of given weight coefficient
. We can find two interesting points in
Figure 5. The first point is that the network EE is a concave function on the weight coefficient
. In addition, it can be concluded that the network EE has no change irrespective of the different numbers of users. Then, we analyze and discuss the two points as follows. Firstly, the weight coefficient can denote the degree of preference for energy consumption. With the weight coefficient increasing, the energy consumption increases and the network capacity reduces, correspondingly. Then, the network EE is a concave function on
. Secondly, the energy consumption of users can be almost neglected in the wireless downlink transmission system. In the OFDMA system, the distribution and the number of users have little effects on the spectrum efficiency after enough number of Monte Carlo run. Based on the above discussion, it can be concluded that the number of users has a negligible impact on the network EE.
Figure 6 shows the relationship between the network EE and spectrum efficiency where the number of antennas installed on the MBS is different. The simulation setting of the number of users
K and the initial value of each BS’s transmit power
are 30 and
, respectively. The horizontal and vertical axes in
Figure 6 represent the spectral efficiency and EE values of the network calculated by Algorithm 2 with different values of given weight coefficient
, respectively. It can be obviously seen in
Figure 6 that the network EE is a concave function on the spectrum efficiency. In addition, the maximum value of the network EE and spectral efficiency increase and reduce with the number of antennas installed on the MBS increasing from 50 to 200. This point is interesting and we have some discussion about it. On the one hand, the number of available spatial channels increases with the more antennas existing in the network. The network obtains greater throughput and spectrum efficiency. On the other hand, the energy consumption of the network also increases as the number of antennas increases. To some extent, the network EE reduces, correspondingly.
Figure 7 shows the network EE calculated by different algorithms when the number of antennas installed on the MBS increases from 10 to 170. The simulation setting of the number of users
K and the initial value of each BS’s transmit power
are 30 and
, respectively. We can find two interesting points in
Figure 7. Firstly, it can be obviously found that the network EE is a concave function on the number of antennas installed on the MBS. This point can be proven as described above. The second point is that the performance of the network EE obtained by the proposed Algorithm 1 is superior to other single optimization algorithms, which also demonstrates the available effectiveness of the proposed Algorithm 1 to some extent.
5. Conclusions and Future Work
In the scenario of massive MIMO enabled HetNets, we propose an optimal iterative EE algorithm and joint optimization algorithm. From the simulation results and relevant discussion about the performance of proposed algorithms, we conclude the network EE is a concave function on the weight coefficient, the spectrum efficiency and the number of antennas installed on the MBS. That is, we sacrifice some part of the performance of spectrum efficiency with maximizing the network EE. Furthermore, we also propose the joint user association and power control algorithm for the development of new energy reduction policies in massive MIMO enabled HetNets and optimize the number of antennas installed on the MBS to increase the network EE. In addition, we analyze the validity and effectiveness of all the proposed algorithms. In future works, the joint number of antennas, user association, and power control and channel allocation optimization will be considered in the massive MIMO enabled HetNets. That is, to further increase the network EE, we will also optimize the number of antennas and channel allocation in future work, which is the main source of increasing the network EE.