2016 24th Iranian Conference on Electrical Engineering (ICEE)

Power minimization in 5G Heterogeneous Cellular

Networks
Mohammad Moltafet Paeiz Azmi Nader Mokari
Faculty of Electrical and Faculty of Electrical and Faculty of Electrical and
Computer Engineering Computer Engineering Computer Engineering
Tarbiat Modares University Tarbiat Modares University Tarbiat Modares University
Tehran, Iran Tehran, Iran Tehran, Iran
Email: m.moltafet@tmu.ac.ir Email: pazmi@tmu.ac.ir Email: nader.nmy@gmail.com

AbstractIn this paper, the minimizing transmit power in 5th (SIC) on the receiver side to separate the multi-user signal.
generation (5G) is investigated. To minimize transmit power, NOMA with SC and SIC improves the capacity region of the
different appropriate schemes such as heterogeneous cellular down-link AWGN channel compared to orthogonal multiple
network (HetNet) and non-orthogonal multiple access (NOMA)
have been proposed in 5G. In NOMA scheme more than one access (OMA) schemes [2]. Resource allocation is very im-
user can be allocated in the same sub-carrier. This scheme uses portant and critical in HetNet and NOMA-based system to
superposition coding (SC) on transmitter side to multiplexing improve the system performance. In [3], the authors proposed
the multiple users, and successive interference cancellation (SIC) an algorithm to minimize total power consumption. A low-
on the receiver side to separating users signal. Obtaining joint complexity algorithm to minimize the down-link transmitted
sub-carrier and power allocation in NOMA-based HetNet is the
aim of this work. In this regard, the proposed objective function power in (multiple input multiple output-orthogonal multiple
is to minimize the total transmitted power in all cells subject access) MIMO-OFDMA systems has been proposed in [4].
to the constraint related to NOMA scheme in addition to users The authors proposed a scheme to dynamically turn off some
minimum rate requirement constraint. An iterative algorithm BSs when the respective cells have low trafc load in [5].
is applied to solve our problem. Since the problem is non- In [6], a resource allocation algorithm to energy saving in
convex and incorporate both integer and continuous variables, it
cannot be solved directly. Therefore, we decompose it into two heterogeneous networks has been investigated. The system
subproblems: sub-carrier and power allocation. In each iteration, level performance of a OFDM-NOMA system considering
the sub-carrier allocation is updated by applying integer linear fractional transmit power allocation has been studied in [7].
programing (ILP) and, the power allocation is updated by The main contributions of this paper are summarized as
applying successive convex approximation for low complexity follows:
(SCALE). Simulations results show that the proposed algorithm
in NOMA-based system performs better than the OFDMA-based
system. We investigate about resource allocation for down-link of
Index TermsHetNet, NOMA, power minimization. a NOMA-based HetNet system. The motivation behind
this consideration is that HetNet and NOMA improve SE
I. INTRODUCTION and ES, therefore, they can be a reasonable choice for
The demanding for various high data-rate multi-media 5G wireless network to reduce energy consumption.
wireless services is increasing dramatically. The result of We propose a novel resource allocation problem which
this growth is great use of energy [1]. The environmental aims to minimize the total transmitted power for the
effects of this increasing energy consumption has received down-link of a NOMA-based HetNet system. the con-
much attention. Therefore, the next generation communication straints in our problem are users minimum rate require-
systems has to provide higher data rate with considering power ment and constraint related to SC and SIC. We optimize
and bandwidth limitation. To this end, different appropriate the total transmit power in term of sub-carrier and power
schemes such as heterogeneous cellular network (HetNet) and allocation.
non-orthogonal multiple access (NOMA) have been proposed To solve the resource allocation problem an iterative
in 5th generation (5G) wireless communications. HetNet has algorithm is applied. In each iteration power allocation
advantages from the standpoint of spectral efciency (SE) and sub-carrier assigned are updated separately. To update
and energy saving (ES). This advantages come from the fact sub-carrier allocation, integer linear programing (ILP)
that, low power access nodes in HetNet, enhance SE by with xed power allocation is used, and to update power
increasing area spectrum reuse and improve ES by decreasing allocation successive convex approximation with low
the distance between end users and access node. Also, NOMA complexity (SCALE) with xed sub-carrier assignment
is a proposed multiple access (MA) scheme which, uses super- is applied. Simulation results show that by applying our
position coding (SC) on the transmitter side to multiplexing the resource allocation algorithm, near-optimal performance
multiple users and, applies successive interference cancellation can be achieved, and it is outperforms OFDMA-based

978-1-4673-8789-7/16/$31.00 2016
c IEEE
234
2016 24th Iranian Conference on Electrical Engineering (ICEE)

system. Also, rk,f (P, ) denotes the sum-rate that user k can achieve
The rest of this paper is organized as follows. In Section in BS f , which evaluated by
II, we give system model and problem formulation for the N

n n
considered system. In Section III, the algorithm development is rk,f (P, ) = k,n rk,f (pnf ) . (4)
studied. Simulation results are presented in Section IV. Finally, n=1
concluding remarks are drawn in Section V. B. Problem Formulation
II. SYSTEM MODEL AND PROBLEM FORMULATION Our resource allocation problem is formulated as
 F
N 
K 
We consider a multiuser NOMA-based HetNet system n
where consist of F + 1 base stations (BSs)(one macro base min k,f pnk,f , (5a)
,P
k=1 n=1 f =0
station (MBS) and F small base stations (SBSs)). We denote
F  F
N 
by Kf the total number of users in BS f where f =0 Kf = n n
K. The total bandwidth is shared between MBS and SBSs s.t. : k,f rk,f (, P) Rkmin , k, (5b)
n=1 f =0
through OFDMA, which is divided into N sub-carriers. Based
on NOMA technique one sub-carrier can beassigned to more
k1
  
n n
than one user. The system parameter is dened as: hnk,f is i,f pni,f k,f pnk,f , k, n, f, (5c)
i=1
the channel between user k and BS f over sub-carrier n, K
pnk,f is the transmit power of BS f to user k over sub-carrier 
n
n k,f LT , n, f, (5d)
n, the binary variable k,f {0, 1} indicates the sub-carrier
n k=1
allocation of user k in BS f , e.g., k,f = 0 if the sub-carrier F
n 
n is not allocated to user k in BS f , otherwise k,f = 1, n
k,f 1, k, n, (5e)
f
pmax denotes the total power available to BS f , Nf is the f =0
number of sub-carriers that use in BS f , ln is the number pnk,f 0, k, n, f, (5f)
of users over sub-carrier n and LT denotes the maximum n
 
k,f 0, 1 , k, n, f, (5g)
number of users that allocate in each sub-carrier. We also use
P = [pnK,f ] K, n, f and = [K,f
n
] K, n, f . We cumulate where (5b) demonstrates the minimum requirement rate for
K
{Pk,f }k=1 into a vector as pf = [p1,f , pn2,f , ..., pnKf ,f ].
n f n n each user, (5c) denotes constraint related to SC an SIC, (5d)
We suppose that each sub-carrier can be used in one BS. demonstrates that at most LT users can allocate in each sub-
Hence, cells do not interfere with each other. carrier, (5e) demonstrates that each sub-carrier can be used at
most in one BS. This problem is a nonlinear program incor-
A. Signal Model porating both integer and continuous variables. Furthermore,
In each small-cell we consider NOMA scheme such that constraint (5b) is not convex. Therefore, problem (5) might
channels are sorted as |hn1,f |2 |hn2,f |2 |hnKf ,f |2 . We have multiple local optima. Therefore, we can not use the
Kf  n n available convex optimization method directly [8]. To reduce
suppose that the BS f sends i=1 pi,f si,f over sub-carrier n
n computational complexity we redene k,f in constraint (5g)
n, where si,f is the message for the ith user over sub-carrier n
as a continues variable between zero and one (0 k,f 1).
n. The received signal at k th user is expressed as
The new denition of is interpreted as a portion of time that
Kf 
 sub-carrier n is assigned to user k in BS f [9], [10]. Then,
n
yk,f = hnk,f pni,f sni,f + k,f
n
, (1) our problem can be rewritten as
i=1 K  F
N 

n
where k,fn
denotes the additive Gaussian noise. min k,f pnk,f , (6a)
,P
Based on NOMA concept, for sorted channel as |hn1,f |2 k=1 n=1 f =0
|hn2,f |2 |hnKf ,f |2 , we let pn1,f pn2,f pnKf ,f . s.t. : (6b) (5f), (6b)
In NOMA-based system user k would be able to remove the n
k,f [0, 1], k, n, f. (6c)
signals of the users with lower order, and users with higher
order ({1, . . . , k 1}) will be acted as noise by user k. After III. SOLUTION OF THE RESOURCE ALLOCATION PROBLEM
applying SIC the SINR of user k over sub-carrier n in BS f An iterative algorithm is applied which allocates the power
is evaluated by and sub-carrier of each user alternatively [11], [12]. The
|hnk,f |2 pnk,f algorithm of power allocation and sub-carrier assigned is
n
k,f =  k1
. (2) summarized in Table I.
|hnk,f |2 i=1 pni,f + E[|k,f
n |2 ]
Our algorithm starts with nding the initial power (P(0))
n
rk,f denotes the rate that user k can achieve over sub-carrier and sub-carrier ((0)). The iteration number is s, and at
n in BS f , which evaluated by the beginning of each iteration, the optimal P is found
by solving problem (6) with xed sub-carrier ( = (s))
n n
rk,f = log(1 + k,f ). (3) Then, the optimal (s) for problem (6) is computed with

235
2016 24th Iranian Conference on Electrical Engineering (ICEE)

TABLE I: ITERATIVE RESOURCE ALLOCATION

C. Power Allocation
1: Initialization: Set s=0, compute (0) and P(0) with the initialization Power allocation problem with xed sub-carrier assignment
method. can be expressed as
2: Set = (s) and replace the non-convex constraint (6b)
with its convex approximation,
3: Solve problem (5) and assign its solution in to P(s + 1),  N 
K  F
n
4: Compute (s + 1) by solving (5) with xed power, min k,f pnk,f , (8a)
5: If ||vec(P(s)) vec(P(s 1))|| P
m=1 n=1 f =0
stop,
else  F
N 
n n
set s = s + 1 and go back to step 2. s.t. : k,f rk,f (, P) Rkmin , k, (8b)
n=1 f =0
k1
  
n n
i,f pni,f k,f pnk,f , k, n, f, (8c)
considering P = P(s + 1). To approximate the non-convex i=1
power allocation problem by convex problem for STEP 2, pnk,f 0, k, n, f. (8d)
SCALE method is applied. Sub-carrier allocation problem for
STEP 4, is an integer linear programing (ILP) problem that
can be solved by using available optimization software for
This problem is non-convex. In order to tackle this issue, we
convex optimization problems, e.g., CVX [13]. The iteration
use the successive convex approximation (SCA) approach [14],
is continued until vec(P(s)) vec(P(s 1)) , where
to replace the non-convex constraint with the approximated
0 <  1. Note that, operator vec(.) stacking the columns
convex. In order to apply SCA approach, at rst we dene the
of a matrix into a vector and . shows the Euclidean norm
following inequality
of a vector.

log(w) + log(1 + w), (9)

A. Initialization Method
where
To start the iterative algorithm the initial vectors of P and
are needed. To compute P(0) the uniform power allocation is W0 W0
pfmax = , = log(1 + W0 ) log(W0 ).
applied. Hence, if k,nf
= 1, we have pfk,n = otherwise, W0 + 1 W0 + 1
N f ln (10)
pfk,n = 0. To demonstrate (0), we nd the appropriate vector By applying (9), the constraint (8b) is replaced by
that minimize total transmit power with P = P(0).
N 
 F
n n n
k,f log(k,f ) + k,f Rkmin , (11)
B. Sub-carrier Allocation n=1 f =0

The problem of nding the optimal sub-carrier allocation n n

with xed power is expressed as where k,f and k,f are xed. Our optimization problem is
still non-convex, because the constraint (8b) is not convex.
K  F
N  By utilizing transformation pnk,f = exp( pnk,f ), the standard

min n
m,f pnk,f , (7a) concave maximization problem in the new variables pnk,f can
be achieved. The new optimization is expressed as
k=1 n=1 f =0
N 
 F
n n K 
 N 
F
s.t. : k,f rk,f (, P) Rkmin , k, (7b) n
n=1 f =0
min k,f pnk,f ),
exp( (12a)
P
k=1 n=1 f =0
k1
  
n n N 
 F
i,f pni,f k,f pnk,f , k, n, f, (7c) n n n
i=1
s.t. : k,f log(k,f ) + k,f Rkmin , (12b)
K n=1 f =0

n
k,f LT , n, f, (7d) k1
  
n n
k=1 i,f pni,f ) k,f
exp( pnk,f ), k, n, f,
exp(
F i=1

n
k,f 1, k, n, (7e) (12c)
f =0 pnk,f ) 0, k, n, f.
exp( (12d)
n
k,f [0, 1], k, n, f. (7f)

This problem can be solved by applying ILP. This is a standard form of convex minimization. To show

236
2016 24th Iranian Conference on Electrical Engineering (ICEE)

TABLE II: POWER ALLOCATION ALGORITHM TABLE III: AlGORITHM TO FIND STATIONARY POINT

1: Initialization: Set m = 0, and initialize P(s)0 = P(s), 1: Initialization: Set u = 0, nd 0f and k,n,f
0 with applying (20),
= 0 and = 1, 2: Compute p by using (17),
2: Find P(s)m , by solving problem (12), 3: Update and by using (18) and (19) respectively,
3: By using W0 = m,f n (P(s)m ) update and ,
4: If || u u1 ||  and ||u u1 || 
4: If m = M or convergence stop,
stop, else
else set u = u + 1 and go back to STEP 2.
set m = m + 1 and go back to STEP 2,
5: Output:
P(s + 1) = P(s)M .
After simplifying (16), pnk,f is given by
n
K n
convexity of constraint (12b), we rewrite it as k,f i=k+1 i,f

 n +
N 
F pk,f
 pnk,f = .
n
k,f (log(|hnk,f |2 ) + exp(
pnk,f ) log(|hnk,f |2 (13) K  n
i i,f n
i,f 
i=k+1 n 2 1
n=1 f =0 i1 E[|k,f | ]
k1
 pnj,f )j,f
j=1 exp(
n +
n 2
|hk,f |
exp(exp pni,f )) + E[|k,f
n 2 n
| ]) + k,f . (17)
i=1
To update dual variables, we apply sub gradient method as
Each term in (13) is convex (log-sum-exp is convex [15]), follow
consequently, (13) is concave. In order to determine a power N 
F
allocation better than P(s), we apply an iterative power allo- 
u+1 = [u 1 ( n
k,f n
rk,f Rkmin )]+ , (18)
cation algorithm. Our iterative resource allocation algorithm n=1 f =0
is summarized in Table II. Since the initial choice of the
constants and is not critical, we start with = 0 and  k1
 
= 1. The iteration number is m, and P(s + 1) is the output u+1 = [ u 2 (k,f
n
pnk,f n
i,f pni,f )]+ , (19)
of this algorithm. The Lagrangian function of optimization i=1

problem (12), with variables , is expressed as where []+ = max(, 0), the iteration number is u, and 1
 N 
K  F K
 and 2 are step-size for updating algorithm. We summarize
, ) =
L(P, n
k,f pnk,f ) +
exp( k (14) the algorithm to solve (12) in Table III. To initialize variables
k=1 n=1 f =0 k=1 f and k,n,f , the uniform method is used as follow
N 
 F K 
 N 
F
1 1
( n
k,f n
log(k,f n
) + k,f Rkmin ) + 0f = , 0 = . (20)
Kf k,n,f (F + 1)Nf ln
n=1 f =0 k=1 n=1 f =0
 k1
 
n n n
k,f (k,f pnk,f )
exp( i,f pni,f )).
exp( IV. SIMMULATION RESULTS
i=1
In this section, we evaluate the performance of the proposed
To solve the dual maximization problem stationary point of
scheme, by doing simulation under various system parameters.
(14) with respect to p should be found. In this regard, the
The diameter of MBS and SBSs are considered 1 Km and 20
related equation is dened as
m, respectively. The small-cells and users are randomly located
, ))
d(L(P, inside MBSs coverage area. The simulation parameters are
= 0. (15)
pnk,f
d considered as: N = 64, L = 3, K0 = 5, Kf = 3 for
f F , hnk,f = xnk,f (dnk,f ) where is the path loss exponent
Therefore, we have
which is considered = 2, xnk,f representing the Rayleigh
, )) K
d(L(P, n n fading, and d is the distance between the the base stations and
= 0 exp( p k,f ) +
k k,f + (16) theirs users, e.g., dnk,f denotes the distance between user k
dpnk,f
i=k+1
 n and BS f . The performance of proposed algorithm in NOMA-
n n
exp( p k,f ) based system is compared to OMA-based system and global
i i,f i,f n 2
i1 n ) n +
E[|k,f | ] optimum solution. To nd global optimum solution we use
j=1 exp(
p j,f j,f n 2
|hk,f | exhaustive search algorithm that shown in Table IV. G is the
K step size for searching in exhaustive search algorithm, also
n
k,f  n
+   i,f = 0. is considered 1000 in our simulation. Fig. 1 shows the total
2 exp(pnk,f ) i=k+1 2 exp(pnk,f ) transmit power versus the users required rate. Generally, with
increasing the users required rate, the total transmit power is
increased for both NOMA-based and OFDMA-based system.

237
2016 24th Iranian Conference on Electrical Engineering (ICEE)

4 12
NOMAProposed algorithm NOMAProposed algorithm
NOMAExhaustive search NOMAExhaustive search
3.5 OFDMAExhaustive search 10 OFDMAExhaustive search
Total Transmit Power (Watt)

Total Transmit Power (Watt)

3 8
Users minimum
rate requirement
= 3 (bps/Hz)
2.5 6

2 4

1.5 2 Users minimum

rate requirement
= 1 (bps/Hz)
1 0
1 1.5 2 2.5 3 3.5 4 4.5 5 3 6 9 12 15
Minimum rate requirement (bps/Hz) Number of users)

Fig. 1: Total transmit power versus the users minimum rate require- Fig. 2: Total transmit power versus the number of users
ment

HetNet system the system performance in terms of transmit

This is because by increasing users required rate, users signal power is increased by approximately 22%, compared to OMA.
to noise ratio (SNR) should be increased. Therefore, in xed
noise variance transmitted power should be increased. In Fig. R EFERENCES
2, total transmitted power versus the number of users with [1] G. Fettweis and E. Zimmermann, ICT energy consumption trends and
users required rate 1 (bps/Hz) and 3 (bps/Hz) is shown. We challenges, in The 11th International Symposium on Wireless Personal
Multimedia Communications (WPMC 2008)., 2008, pp. 1-3.
compare the transmit power of proposed NOMA-based system [2] D. Tse and P. Viswanath. Fundamentals of Wireless Communication,
with the OMA-based system. Also, we compare our algorithm Cambridge University Press, May. 2005.
with global optimum solution. As the gures show, proposed [3] Y.-B. Lin, T.-H. Chiu and Y. T. Su, Optimal and Near-Optimal Resource
Allocation Algorithms for OF-DMA Networks, IEEE Trans. Wireless
algorithm is near optimal and system performance in terms of commun. , Vol. 8, no. 8, pp. 4066-4077, Aug. 2009.
transmit power is improved by approximately 22% compared [4] N. UL Hassan and M. Assaad, Low Complexity Margin Adaptive
to OMA, also there is by approximately 7% gap between Resource Allocation in Downlink MIMO-OFDMA system, IEEE Trans.
Wireless Commun., Vol. 8, no. 7, pp. 3365-3371, jul. 2009.
proposed algorithm and global optimal solution. [5] J. Kim, P.-Y. Kong, N.-O. Song, J.-K. K. Rhee and S. A.-Araji,
MDPbased dynamic base station management for power conservation
V. CONCLUSION in self-organizing networks, in Proc. IEEE Wireless Communications
and Networking Conference (WCNC)., April 2014.
In this paper, we have investigated the power minimization [6] Xun. Sun , Shaowei. Wang, Resource Allocation Scheme for Energy
problem in NOMA-based HetNet. The optimization objective Saving in Heterogeneous NetworKs, IEEE Trans. Wireless Commun.,
is to minimize the total transmitted power subject to the 2015,doi:10.1109/TWC.2015.2420558.
[7] Y. Saito, A. Benjebbour, Y. Kishiyama, and T. Nakamura, System-
constraint related to NOMA scheme in addition to users level performance evaluation of downlink non-orthogonal multiple access
minimum rate requirement constraint. We have applied an (NOMA), in proc. IEEE PIMRC., pp. 611-615, Sep. 2013.
iterative algorithm to solve this problem. The main optimiza- [8] D. P. Bertsekas, Nonlinear Programming, 2nd ed. Singapore: Athena
Scientic, 2003.
tion problem is decomposed into two subproblems: 1) Power [9] W. Dang, M. Tao, H. Mu, and J. Huang, Subcarrier-pair based resource
allocation. 2) Sub-carrier assignment. In each iteration power allocation for cooperative multi-relay OFDM systems, IEEE Trans.
allocation is updated by applying SCALE algorithm and, sub- Wireless Commun., vol. 9, no. 5, pp. 1640-1649, May. 2010.
[10] Y. Cheong, R. Cheng, K. Lataief, and R. Murch, Multiuser OFDM
carrier assignment is updated by using ILP. As the results with adaptive subcarrier, bit, and power allocation, IEEE J. Sel. Areas
show, by applying the proposed algorithm in NOMA-based Commun., vol. 17, no. 10, pp. 1747-1758, Oct. 1999.
[11] T. Wang and L. Vandendorpe, Iterative resource allocation for maxi-
mizing weighted sum min-rate in downlink cellular OFDMA systems,
TABLE IV: EXHAUSTIVE SEARCH ALGORITHM IEEE Trans. Signal Process., vol. 59, no. 1, pp. 223-234, 2011.
[12] K. Son, S. Lee, Y. Yi, and S. Chong, REFIM: A practical interference
pfmax management in heterogeneous wireless access networks, IEEE J. Sel.
1: Initialization: Set G = 1000 to produce p = as the step
G Areas Commun., vol. 29, no. 6, pp. 1260-1272, Jun. 2011.
size for searching, [13] I. CVX Research, CVX: Matlab software for disciplined convex
2: Find all possible solutions of (5) as programming, version 2.0, http://cvxr.com/cvx, Aug. 2012.
ppossible = [pn
f,k ] k, n, f, [14] B. R. Marks and G. P. Wright, A general inner approximation algorithm
n ]
possible = [f,k k, n, f, for nonconvex mathematical programs, Operations Research., vol. 26,
where no. 4, p. 681-683, 1978.
f
pnf,k = 0, p , 2p , ..., pmax k, n, f ,
[15] S. Boyd and L, Vandenberghe, Convex Optimization Cambridge. U.K.:
n
f,k = 0, 1 k, n, f , Cambridge Univ. Press, 2004.
4: Among the feasible solutions, choose the best corresponding
to the maximum throughput.

238