1. Introduction
Wireless data traffic and the demand for bringing a higher data rate to a growing number of users has been increasing with each passing year and, in order to provide seamless connectivity, future generation networks will have to rely on denser deployment of infrastructure, reducing the inter- and intra-cell interference, simple signal processing, and reduction in the transmitted power along with improved energy and spectral efficiency [
1,
2]. In the conventional techniques, communication between the base station and users has happened in separate time-frequency resources by orthogonalizing the channel, but it results in interference when the number of users increases, because, in order to make sure the higher data rates, several users have to operate in the same time and frequency resources [
3,
4] and we have to use complex signal processing techniques like dirty paper coding and maximum likelihood multiuser detection [
5] in order to mitigate the interference [
6,
7]. The initial focus of the researchers was on Multiple-Input Multiple-Output (MIMO) technologies because they provide a substantial gain in area and spectral efficiency [
8,
9]. It has been seen that the deployment of a large antenna array at the base station (BS) results in substantial reduction in the intra cell interferences along with simple signal processing [
10], which in turn have shifted the focus of researchers towards Massive MIMO.
In Massive MIMO, hundreds of antennas are deployed at the BS serving a comparatively lower number of single antenna users which results in higher though put for each user along with increased energy efficiency due to focusing of energy on the intended users and with simple signal processing [
11,
12]. The energy efficiency of a system is defined as the sum-rate (the spectral efficiency) divided by the transmitted and consumed power and it is an important parameter for communication systems [
13] because carbon emission out of the communication devices has become a vital environmental and economic issue [
14,
15]. The initial conception regarding the energy efficiency of Massive MIMO was that it was directly proportional to the number of antennas at the BS but in practical situations when the number of antennas is increased, the power consumption in the circuit also is increased and this cannot be ignored when we are designing the actual and practical systems. Various circuit power consumption models have been proposed and examined in the case of MIMO systems [
16,
17,
18,
19,
20,
21,
22].
In [
23], the authors have estimated the optimal number of antennas and users based on the capacity maximization, but they have not considered the overhead of signaling factor which is used for channel acquisition. In [
24], the authors have considered the overhead signaling factor and improved the energy efficiency of Massive MIMO by calculating the optimal number of transmitters and users under the perfect hardware conditions, but the numerical algorithm that they have proposed is only applicable under perfect channel conditions. This research is further extended in [
25] in which the authors have calculated the optimal energy efficiency along with optimal number of transmitters and users under both the perfect and imperfect channel conditions, but the transmitted power starts becoming reduced when the area of coverage gets increased which is not accurate because in order to cover more area, more transmitted power is required. In [
26], effects of nonlinear amplifiers on the spectral characterization of transmitted signals have been studied in the case of Massive MIMO. Effects of nonlinear amplifiers can be reduced by designing the precoders for low Peak to Average Power Ratio [
27,
28]. In [
29], the authors have calculated the energy efficiency of massive MIMO by considering the effects of nonlinear amplifiers and other hardware imperfections under the perfect channel situations, but they have taken the circuit power consumption as a fixed quantity which is not correct because of the dependence of circuit power consumption on the number of transceiver chains and coherent participation of all BS antennas [
30,
31].
In this paper, we have maximized the energy efficiency of massive MIMO and calculated the optimal number of antennas and users along with optimal transmitted power and their corresponding achievable spectral efficiency under both the perfect and imperfect channel situations. Different from the existing studies [
23,
24,
25,
26,
27,
28,
29,
30,
31], we have taken the overhead signaling factor into account and included the effects of nonlinear amplifiers in each transmitter branch under both the perfect and imperfect channel conditions and with proper modelling of circuit power consumptions. To the best of our knowledge, not much research has been done on the energy efficient designing of Massive MIMO by considering the effects of nonlinear amplifiers under the imperfect channel conditions and with proper modelling of circuit power consumptions. Moreover, we have calculated the optimal number of antennas and users along with optimal transmitted power and their corresponding achievable spectral efficiency under both the perfect and imperfect channel situations. Effects of nonlinear amplifiers on the energy efficiency of Massive MIMO are investigated by calculating the energy efficiency at different nonlinear power amplifier efficiencies and distortion loses under both the perfect and imperfect channel conditions. We have proposed an alternative optimization method that works for both perfect and imperfect channel conditions without much complexity and provides the optimal parameters by converging quickly. The contributions and novelties of this article are summarized as follows:
- (1)
The energy efficient design of Massive MIMO along with the effects of nonlinear amplifiers under the perfect and imperfect channel conditions, and by using the realistic power consumption model, is first proposed and formulated.
- (2)
Mathematical expressions of the spectral efficiency and energy efficiency are derived by considering the effects of nonlinear amplifiers in each transmitter branch under the perfect and imperfect channel conditions.
- (3)
A numerical approach is proposed to optimize the energy efficiency and calculation of optimal parameters. Simulation results are provided to support the mathematical modelling and investigate the relevant trend.
The remainder of the paper is organized as follows: In
Section 2 we have discussed the frame structure and working of Massive MIMO, modeled the transmission and reception of signals and derived the achievable rates of Massive MIMO by considering the effects of nonlinear amplifiers under both perfect and imperfect channel conditions. In
Section 3, we have modeled the power consumptions of Massive MIMO starting from transmitter end to user terminal. In
Section 4, we have defined the problem definition and energy efficiency of Massive MIMO under the perfect and imperfect channel situations. In
Section 5, we have modeled the power amplifiers and in
Section 6, we have proposed a numerical algorithm in order to solve the optimization problems discussed in
Section 4.
Section 7 presents simulation results and discussions, and in
Section 8 we conclude and summarize all the discussions.
Notations: , and show the inverse, Hermitian and transpose operator respectively, means the expectation operation, and denote the logarithm of x to base and respectively, denotes the set of positive integers, and shows the differentiation.
2. Frame Structure and Achievable Rates of Massive MIMO
In Massive MIMO, base station and users have to send training signals known to both transmitters and receivers in order to achieve channel estimation. Accurate and timely acquisition of channel state information (CSI) is very important because Massive MIMO relies on the frequency response of propagation channel. Time Division Duplex (TDD) operation is preferable in the case of massive MIMO because the overhead factor of channel estimation is not dependent on the number of antennas M as compared to FDD operation where overhead factor is so large due to its dependence on the number of antennas. However, few techniques have been proposed and suggested for having the FDD operation in the case of Massive MIMO [
32,
33,
34,
35].
Figure 1 illustrates the frame structure of Massive MIMO in the case of TDD protocol. An uplink and downlink channel are reciprocal to each other in TDD operation and use the same frequency spectrum during the uplink and downlink communications at different time slots. During the uplink operation, each user needs to send training signals or orthogonal pilots to the base station in order to estimate the CSI at the base station for
channel uses followed by the transmission of data from all
K users to BS in the same time-frequency resources for
channel uses as shown in
Figure 1. BS uses the linear precoding to retrieve the signals transmitted from all
K users together with channel estimation. In the downlink, BS uses the estimated channel in order to transmit the required signals to the intended users for
channel uses. Number of transmitters
M and users
K are required to be same during the uplink and downlink operation in the case of TDD protocol.
2.1. Achievable Rates of Massive MIMO under Perfect CSI
Consider the data symbols
transmitted by the base station antennas intended for the
number of users as shown in
Figure 2 then the transmitted vector
can be written as:
where
is a linear precoding matrix and can be expressed as:
where
is a
beam forming vector and can be described as:
and
, where
denotes the power allocation for all users as shown in
Figure 2.
According to Bussgang’s theorem [
36], we can decompose the output of an amplifier as a sum of two uncorrelated components (input signal and the distortion). Let
be the distortion caused by the nonlinear amplifier as shown in
Figure 2 then the signal received at the
user can be expressed as:
The second term in the above equation is due to interference among data symbols and is the Additive White Gaussian Noise (AWGN) having zero mean and unity variance.
Let
is the correlation of
on the interference term and
are the power losses due to nonlinear amplifier, then Equation (3) can be written as:
where
can also be seen as the effect of a nonlinear amplifier to the amplitude of the intended signal which can be termed as ‘clipping’ and in practical situations this contribution is negative, i.e.,
The corresponding clipping power
at the
user terminal can be written as:
The variance of the distortion at the
user terminal due to the nonlinear amplifier can be written as:
In order to have the equal rate for all the users, power allocations need to be done in a clever way and by employing a technique from [
24]; it can be written as:
where
is the received signal to noise ratio and it is considered as an optimization parameter because to optimizing
is equivalent to optimizing
. Since we know that
ZF suppresses the interference, the interference term will be zero:
By using Equations (7) and (8), Equation (4) can be written as:
Additionally, the corresponding signal to noise ratio for the
user (
) can be computed as:
The corresponding achievable rates for the
user can be defined as:
By considering the over-head factor, achievable rate for the
user can be expressed as:
where the factor
accounts for the pilot over-head in each coherence block
U and
is the total relative pilot length.
2.2. Achievable Rates of Massive MIMO under Imperfect CSI
In this subsection, we have calculated the achievable rate of Massive MIMO under imperfect channel conditions. Perfect channel conditions mean that the BS knows all the frequency components of the channel which results in improvement in the performance of the system. In practical situations, due to infinite precision of the electronic instruments and instantaneous nature of the transmission, achieving a perfect CSI is almost impossible. Imperfect CSI causes the inevitable interference among the users which in turn affects the performance of the system. We have assumed that the average attenuation (
) between the users and base station antennas is inversely proportional to transmission power of each user and for the
user it will be (
). As explained in
Section 2, the transmission is divided into two phases, i.e., pilot transmission followed by data transmission.
During the pilot transmission phase, variance of the estimated channel by using MMSE estimator can be written as [
37,
38,
39]:
During the data transmission phase, achievable rates for the
user by assuming the
ZF and treating the estimated channel as true channel, considering the effects of a nonlinear amplifier under imperfect channel conditions, can be written as:
where
is the same as that of
and, similarly, achievable rates for the
user by considering the pilot overhead can be expressed as:
3. Modeling of Power Consumptions
In this section we have modeled the power consumptions of Massive MIMO. The total power consumptions in the circuit of Massive MIMO can be composed into two parts:
where
is the total power consumed by the power amplifiers and can be illustrated as [
14]:
where
is the path loss factor and when the required SNR will be fixed then this factor would be very important in order to calculate the total power consumption of the power amplifiers.
is the efficiency of the power amplifier and is explained in detail in the power amplifier modeling section (
Section 5) and
is the bandwidth.
is the total circuit power consumptions of Massive MIMO, i.e., power consumed in the transmitter and receiver chains, oscillator and filter power consumption, power required for the coding and decoding of the desired signals, power required for the channel estimation and linear processing. So, we need to model all the required or consumed power in the above mentioned processes.
Power consumed at the transmitter and receiver chain can be illustrated as:
where
is the total power consumption at the transmitter and receiver chains,
is the power consumption at the transmitter chain,
is the power consumption at the receiver chain, i.e., power consumed at the filters, converters and mixers, and
is the oscillator power in order to synchronize the frequencies.
The power required for the coding and decoding of the desired signal can be demonstrated as:
where
and
denote the corresponding power consumption during coding and decoding.
As explained in
Section 2, Massive MIMO relies on CSI of the channel, i.e., BS and users have to send training or pilot signals during the uplink and downlink of the channel in order to get the frequency response of the channel during the coherence time. Power consumption during this process can be written as [
40]:
where
and
are the computation efficiencies at the transmitter and receiver end.
Consumption of power during linear processing by assuming
ZF has been explained in [
24] and can be written as:
So the total circuit power consumption of Massive MIMO by using Equations (18)–(21) can be expressed as:
where
is the fixed power required for site cooling and the total power consumptions Equation (16) of Massive MIMO by using Equation (17), Equation (22) can be illustrated as:
Total power consumptions can be written in more simplified and concentrated way:
with the following substitutions:
5. Modeling of Nonlinear Amplifiers
Conventionally used amplifiers in the case of MIMO systems are the multi-transistor amplifiers such as Doherty amplifiers. Doherty amplifier splits the input signal into two parts and then amplifies them in two different amplifiers (peaking amplifier and carrier amplifier) and then the outputs of these two amplifiers are summarized to get the desired output. The Doherty amplifier provides higher efficiency and is well suited for the signals which have the higher peak to average power ratios. However, the issues of using the Doherty amplifiers are their higher cost and complexity. Due to these drawbacks, they are not feasible to use in the case of Massive MIMO because of the large number of BS antennas.
We need to have the simple design and cost efficient power amplifiers in the case of Massive MIMO like class A, class B or class C. In this article, we have considered the most basic class B amplifiers and the power efficiency of such kind of amplifier is given as [
41]:
where
is the AM–AM conversion of the power amplifier. Various models have been proposed and suggested for modeling of power amplifiers in the literature like Solid State Power Amplifier (SSPA) model, Travelling Wave Tube Amplifier (TWTA) model, RAPP model and ERF model. Out of them, the most commonly used is the RAPP model where AM–PM conversion is assumed to be negligible and AM–AM conversion is given by [
41]:
where
p controls the smoothness of the curve and in order to keep the total power
P,
and
are assumed to be [
42]:
where
is the compensation factor for the power loses. In the next section, we have developed an algorithm in order to solve the optimization problems of
and
.
6. Problem Solution and Numerical Algorithm
In this section, we have designed an algorithm to solve the optimization problems
and
. It is difficult to solve the optimization problem of
and
due to mixed nature of their corresponding objective functions with respect to
M,
K and
p. Consider the following substitutions in order to simplify the objective functions of
and
:
where
can be explicated as the number of active users,
can be explicated as the number of active antennas per user and
along with multiplication of some constant factor as described in Equation (17) can be explicated as the total power of power amplifiers. The simplified objective functions of
under perfect channel conditions can be written as:
with the following modified optimization problem:
Similarly, objective function of energy efficiency under imperfect channel conditions following the above mentioned substitutions can be written as:
with the following modified optimization problem:
Objective functions of optimization problems
and
follows a quasi-concave response because they are first increasing and then deceasing in each dimension while the other dimensions are fixed and their second order derivatives are less than zero. The proof of the quasi-concave nature of objective functions (
and
) have been shown in the
Appendix A and
Appendix B respectively.
According to
Appendix A and
Appendix B, objective functions
and
undergo a peak point at the unique zero crossing of
and
in each dimension while the other dimensions are fixed. The following flow chart summarizes the above mentioned discussions and shows the simulation steps (
Figure 3).
7. Simulations and Numerical Results
In this section, we have performed simulations to test the mathematical and numerical algorithm discussed in the earlier sections. Realistic simulation parameters have been chosen for simulations as shown in the
Table 1.
Figure 4 shows the amount of power lost due to clipping at different efficiencies of power amplifier with respect to different back-offs, calculated by using Equations (5), (27) and (28).
Number of transmitters and receivers are set to be 120 and 20 and it can be seen from
Figure 4 that the power losses due to the consequences of clipping are less than −0.3 dB.
Similarly,
Figure 5 shows the losses due to distortion at different efficiencies of power amplifier with respect to different back-offs. The number of transmitters and receivers and path loss exponent are set to be same in
Figure 4 and
Figure 5.
Figure 6 shows the optimal number of transmitters at different area of coverage ranges from 100 m to 500 m by setting different circuit power consumption levels under both the perfect and imperfect channel conditions As can be seen from
Figure 6, when the coverage area increases, the optimal number of transmitters increases, respectively, in order to cover that area and when the channel condition is imperfect then more numbers of transmitters are required, whereas when the power consumptions of the circuit are less, optimal numbers of transmitters required for the system are less and vice versa.
Similarly,
Figure 7 shows the optimal number of users at different area of coverage ranges from 100 m to 500 m at different circuit power consumption levels under both the perfect and imperfect channel situations. As can be seen from the
Figure 7, more users can be accommodated at a higher area of coverage.
Figure 8 shows the optimal transmitted or PA power at different area of coverage ranges from 100 m to 500 m by setting different circuit power consumption levels under both the perfect and imperfect channel conditions.
As can be seen from
Figure 8, more transmitted power is required in order to cover more distance and imperfect channel condition results in more transmitted power with the corresponding area throughput that maximizes the energy efficiency of Massive MIMO shown in
Figure 9.
Figure 10 shows the optimal energy efficiency and it can be seen from
Figure 10 that less power consumptions of the circuit results in more achievable energy efficiency and under imperfect channel conditions energy efficiency is reduced because the system need to transmit more transmitted or PA power in order to mitigate the negative effects of imperfect channel conditions.
Figure 11 and
Figure 12 show the 3D representation of energy efficiency along with all the optimal parameters in which maximum distance is set to be 300 m and power consumption parameters are set to be
,
= 1,
= 1 and
= 2 under both the perfect and imperfect channel conditions respectively. The optimal parameters come out to be
M = 216,
K = 112,
= 141.4 W, EE = 19.5 Mbit/Joule whereas energy efficient area throughput to be 11.9 Gbits/Km
2 in the case of perfect channel conditions as shown in
Figure 11 and when the channel conditions are not perfectly known then the optimal parameters comes out to be
M = 241,
K = 127,
= 245 W, EE = 16.1 Mbit/Joule and area through put = 11.2 Gbits/Km
2 as shown in
Figure 12.
Figure 13 shows the convergence of energy efficiency with respect to the number of iterations by using the numerical algorithm (discussed in
Section 6) at various distances under the perfect and imperfect channel conditions. The computation complexity of the proposed algorithm at each iteration can be written as:
where
represents the required computation complexity during the computation of
, and
and
represent the required computation complexity during the computation of
and
respectively at each iteration. As can be seen from the
Figure 13, the energy efficiency converges completely at the sixth iteration, thus the overall computation complexity of the proposed algorithm can be written as
. The power consumptions parameters in
Figure 13 are set to be
,
= 0.5,
= 0.5 and
= 1.
Figure 14 shows the impacts of power amplifier efficiencies on the energy efficiency of Massive MIMO and it can be seen easily that when the power amplifiers are operating at higher efficiency, energy efficiency is maximum and vice versa under both perfect and imperfect channel conditions. The power consumptions parameters are set to be
,
= 0.5,
= 0.5 and
= 1 for simulations in
Figure 14.