Accepted Manuscript

A new dynamic firefly algorithm for demand estimation of water

resources

Hui Wang, Wenjun Wang, Zhihua Cui, Xinyu Zhou, Jia Zhao, Ya Li

PII: S0020-0255(18)30053-7
DOI: 10.1016/j.ins.2018.01.041
Reference: INS 13395

To appear in: Information Sciences

Received date: 3 November 2017

Revised date: 26 December 2017
Accepted date: 21 January 2018

Please cite this article as: Hui Wang, Wenjun Wang, Zhihua Cui, Xinyu Zhou, Jia Zhao, Ya Li, A new
dynamic firefly algorithm for demand estimation of water resources, Information Sciences (2018), doi:
10.1016/j.ins.2018.01.041

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service
to our customers we are providing this early version of the manuscript. The manuscript will undergo
copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please
note that during the production process errors may be discovered which could affect the content, and
all legal disclaimers that apply to the journal pertain.
 ACCEPTED MANUSCRIPT

A new dynamic firefly algorithm for demand estimation of

water resources

Hui Wanga,b , Wenjun Wangc , Zhihua Cuid , Xinyu Zhoue , Jia Zhaoa,b , Ya Lif,∗

T
a Jiangxi Province Key Laboratory of Water Information Cooperative Sensing and Intelligent Processing,
Nanchang Institute of Technology, Nanchang 330099, China
b School of Information Engineering, Nanchang Institute of Technology, Nanchang 330099, China

IP
c School of Business Administration, Nanchang Institute of Technology, Nanchang 330099, China
d School of Computer Science and Technology, Taiyuan University of Science and Technology, Taiyuan

030024, China

CR
e College of Computer and Information Engineering, Jiangxi Normal University, Nanchang 330022, China
f School of Computer Science and Educational Software, Guangzhou University, Guangzhou 510006, China

Abstract
US
Firefly algorithm (FA) is an effective optimization technique based on swarm intel-
AN
ligence, which has been successfully applied to various practical engineering problems.
In this paper, a new dynamic FA (called NDFA) is proposed for demand estimation of
water resources in Nanchang city of China. First, a dynamic parameter strategy is used
M

to avoid manually adjusting the step factor. Second, three estimation models in differ-
ent forms (linear, exponential and hybrid) are developed in terms of the historical water
ED

use and local economic structure. Third, normalization method is utilized to eliminate
the influences of different units of data. In the experiments, water use in Nanchang
city from 2003 to 2015 is considered as a case study. The data from 2003 to 2012 are
PT

used for finding the optimal weights of the models, and the rest of data (2013–2015)
are applied to test the models. Computational results show that all five FA variants
can achieve promising solutions. The proposed NDFA obtains better performance than
CE

four other FA variants, and its prediction accuracy is up to 97.91%. Finally, the water
demand in Nanchang city from 2017 to 2020 is predicted.
AC

Keywords: Firefly algorithm (FA), Swarm Intelligence, Dynamic parameter, Water

∗ Corresponding
author
Email addresses: [email protected] (Hui Wang), [email protected] (Wenjun Wang),
[email protected] (Zhihua Cui), [email protected] (Xinyu Zhou), [email protected] (Jia
Zhao), [email protected] (Ya Li)

Preprint submitted to Information Sciences January 23, 2018

 ACCEPTED MANUSCRIPT

demand estimation, Water demand prediction, Optimization

1. Introduction

Water is an important natural resource for human beings. With fast development of

T
society, water consumption (especially for freshwater) dramatically increases. How-

IP
ever, most water resources on the Earth is salt water and only 3% is freshwater. There-
fore, optimal and reasonable allocation of water resources is the key to sustainable

CR
utilization of water resources [13].
Water demand estimation is the first step for optimal allocation of water resources,
and it aims to infer the water demand in the future according to the historical water

US
use, current situation, and environment changes. In general, water demand is related to
several factors, such as population, water price, weather, environment, and economy.
AN
Due to some uncertain factors, the estimation error exists. How to exactly estimate the
water demand is worthy to be investigated.
Traditional estimation methods for water demand includes quota method [3], time
M

series [9], regression analysis [6], gray predication [24], and artificial neural network
(ANN) [1]. For most methods, how to choose weighting parameters is a difficult task
because of the random behaviors of water consumptions. In the past several years,
ED

some intelligent algorithms have been used to estimate the water demand. Bai et al. [5]
proposed a multi-scale method for urban water demand estimation. In the approach,
PT

an adaptive chaotic particle swarm optimization (PSO) was used to search the optimal
weighting factors of the relevance vector regression model. In [30], harmony search
(HS) was applied to short term water demand estimation, in which HS aims to search
CE

the parameter of a double seasonal ARIMA model. In [31], a hybrid model based on
soft computing techniques was used to improve the demand estimation of irrigation
water. Fuzzy logic and genetic algorithm (GA) were combined with computational
AC

neural network (CNN). Experimental results show that the hybrid model outperforms
the single CNN model. Romano and Kapelan [32] combined evolutionary algorithms
(EAs) and ANN to construct a smart estimation model. Reported results show that the
mean error is about 5%.

2
 ACCEPTED MANUSCRIPT

Swarm intelligence is a hot research filed in optimization community. In recent

years, many efficient optimization algorithms inspired by swarm intelligence have been
proposed, such as particle swarm optimization (PSO) [23, 39], ant colony optimization
(ACO) [14], artificial bee colony (ABC) [11, 20], bat algorithm (BA) [7, 8], cuckoo

T
search (CS) [12, 49], and firefly algorithm (FA) [45]. FA is a new optimization tech-

IP
nique, which simulates the mating behaviors of flashing fireflies. Some recent studies
show that FA obtains promising performance on many benchmark functions and real-

CR
world problems [15]. In this paper, we present an application of FA on demand estima-
tion of water resources in Nanchang city of China. To reduce the dependency of FA on
its parameters, a new dynamic FA (NDFA) is proposed. In NDFA, the step factor is not

US
fixed and it can be dynamically updated during the search process. Three estimation
models based on linear, exponential and hybrid forms are developed according to the
historical water use and local economic structure. Compare to the literature [38], the
AN
effect of the dynamic step factor is analyzed. Moreover, our approach NDFA is com-
pared with four other FA variants and the future water demand from 2017 to 2020 is
predicted.
M

The remainder of this paper is organized as follows: In Section 2, the standard FA

and its brief review are presented. Estimation models are developed in Section 3. Our
ED

approach NDFA is proposed in Section 4. Simulation results and discussions are given
in Section 5. Finally, the work is concluded in Section 6.
PT

2. Firefly algorithm

Like PSO, FA is also a population-based random search algorithm. Each individual

CE

(firefly) in the population represents a candidate solution. The search of FA simulates

the mating behavior of flashing fireflies. When a firefly is attracted by other brighter
ones, it can move toward other new positions and find potential solutions. To construct
AC

the search model, Yang [45] proposed three assumptions: 1) one firefly is attracted to
other all brighter ones; 2) the attractiveness is determined by the brightness; and 3) the
brightness is affected by the given objective function. The assumptions mean that a
brighter firefly has a better fitness value.

3
 ACCEPTED MANUSCRIPT

Let Xi = (xi1 , xi2 , . . . , xiD ) be the ith firefly in the population, where i = 1, 2, . . . , N,
N represents the population size, and D is the dimensional size. For any two different
fireflies Xi and X j (i , j), their attractiveness can be computed as follows [46].
−γri2j
β(ri j ) = β0 e (1)

T
where β0 is the attractiveness for r = 0, γ is the coefficient for the light absorption, and

IP
ri j is the distance between Xi and X j . The ri j is calculated by [46].

CR
v
u
t D 
X 2
ri j = k Xi − X j k = xid − x jd (2)
d=1
For two fireflies with different brightness, the less brighter firefly will move to the

US
brighter one. Assume that X j is brighter (better) than Xi . Then, Xi will move to X j due
to the attraction. In the standard FA, this movement is defined as follows [46].
  1
!
AN
−γr2
xid (t + 1) = xid (t) + β0 e i j x jd (t) − xid (t) + α rand − (3)
2
where xid and x jd are the dth dimensions of Xi and X j , respectively, rand is a random
number with the range [0, 1], and α ∈ [0, 1] is call step factor. By the suggestions
M

of [47], it is a good choice to replace α by α · S d , where S d is the length scale for the
dth dimension.
ED

The pseudo code of FA is shown in Algorithm 1, where FEs is the number of fitness
evaluations, and MaxFEs is the maximum number of fitness evaluations, Gmax is the
maximum number of generations, and t represents the generation index. As seen, FA
PT

has a simple algorithm framework and can be easily implemented by any programming
language. In this paper, finding the optimal weights of estimation models is considered
CE

as a minimization problem. For maximization problems, please change f (X j ) < f (Xi )

to f (X j ) > f (Xi ) in Step 7.
Since the development of FA, it has attracted much attention, and different FA vari-
AC

ants have been proposed. Fister et al. [16] proposed a memetic FA (MFA), which
modifies the updating strategies of β and α. In [19], different chaotic maps were intro-
duced to update β and γ. Wang et al. [42] analyzed the convergence behaviors of FA
and proposed an adaptive method to update the step factor α. Verma et al. [35] pro-
posed opposition and dimensional based FA (ODFA), which employs two strategies

4
 ACCEPTED MANUSCRIPT

Algorithm 1: The Standard FA

1 Randomly initialize N fireflies (solutions) {X1 (t), X2 (t), . . . , XN (t)} and set t = 0;
2 Evaluate the fitness values of all N initial fireflies;
3 FEs = N;
4 while t ≤ Gmax && FEs ≤ MaxFEs do

T
5 for i = 1 to N do
6 for j = 1 to N do

IP
7 if f (X j (t)) < f (Xi (t)) then
8 Move Xi (t) toward X j (t) according to Eq. 3;
9 Compute the fitness value of Xi (t);

CR
10 FEs++;
11 end
12 end
13 end
14

15

16
end
t++;

Output the best solution;

US
AN
including opposition-based learning (OBL) and dimensional based strategy. Results
show ODFA performs better than FA, PSO, ACO, and differential evolution. Nasiri and
M

Meybodi [29] introduced a history-driven speciation-based firefly algorithm (HdSFA)

to solve dynamic optimization problems. In the standard FA, each firefly in the pop-
ulation is attracted by other all better fireflies. This attraction model is called full at-
ED

traction. Wang et al. [40] proposed a random attraction model, in which each firefly is
compared with another randomly selected firefly. Then, each firefly moves one time at
most. Results show that the random attraction can help FA obtain good performance.
PT

More attractions among fireflies may result in high computational complexity, while
less attractions may lead to premature convergence. In [41], a neighborhood attraction
CE

is proposed. Compared to the random attraction, the neighborhood attraction is a mid-

dle phase between random attraction and full attraction. Lv and Zhao [27] employed
Gaussian disturbance and local search to improved the performance of FA. Results
AC

show that the new FA is superior to five other FA variants.

The research of complex networks is an active area [44]. Amiri et al. [2] used a
multi-objective FA for community detection in complex networks. Simulation results
show that FA can provide useful paradigm for discovering overlapping community

5
 ACCEPTED MANUSCRIPT

Table 1: Historical water use in Nanchang city from 2003 to 2015 (108 m3 ).
Total Industrial Agricultural Residential Ecological
Year
water use water use water use water use water use

2003 24.21 9.81 11.55 2.53 0.32

T
2004 26.22 8.72 14.47 2.75 0.28

2005 28.14 8.30 16.92 2.60 0.32

IP
2006 27.71 8.11 16.73 2.52 0.35

2007 32.55 7.51 21.27 2.92 0.85

CR
2008 30.42 6.90 19.73 2.94 0.85

2009 33.42 6.57 20.15 3.21 3.49

2010 30.87 7.51 17.37 3.49 2.50

2011

2012

2013
31.26

28.82

32.62
8.97

9.20

9.35
US 17.70

14.68

18.23
4.03

4.36

4.45
0.56

0.58

0.59
AN
2014 31.42 8.92 17.35 4.54 0.61

2015 30.64 9.17 16.21 4.64 0.62

Average 29.87 8.39 (28%) 17.10 (57%) 3.46 (12%) 0.92 (3%)
M

structures robustly. Stock price prediction is regarded as one of the most attractive and
meaningful research issues in financial market [36]. Kazem et al. [22] applied support
ED

vector regression with chaos-based FA to forecast the stock price. Cloud computing is a
new computing method based on internet, which can share computational resources to
other computers [25, 26]. In [17], FA was used to maximize the usage rate of resource
PT

in cloud servers. Marichelvam and Geetha [28] presented a hybrid discrete FA (HDFA)
to solve flow shop scheduling problem with minimizing total flow time. Experiments
CE

on some randomly generated test problems show the effectiveness of HDFA.

3. Estimation models
AC

In this paper, we present an application of FA to estimate the water demand in

Nanchang city of China. The water demand is related to the social and economic
conditions. Table 1 shows the historical water use in Nanchang city from 2003 to
2015 [33, 34]. It can be seen that the water use of Nanchang is distributed in three de-

6
 ACCEPTED MANUSCRIPT

partments, agriculture, industry and residents. The average proportion of agricultural

water use is up to 57%. It demonstrates that agriculture is the main department of water
use. Industrial and residential water use also takes large proportions. The ecological
water use is only 3%. To have a clear illustration, Fig. 1 presents the distribution of

T
average water use in different departments. So, three factors related to agricultural, in-

IP
dustrial, and residential water use are utilized to construct the estimation model, while
the ecological factor is ignored. .

CR
From the above analysis, we use gross agricultural production, gross industrial pro-
duction, and population to associate with agricultural, industrial, and residential water
use, respectively. Table 2 lists the total water use, population, gross industrial produc-

US
tion, and gross agricultural production in Nanchang city from 2003 to 2015 [33, 34].
By the suggestions of [4], both linear and exponential forms of models for water de-
mand estimation are defined as follows.
AN
Linear estimation model:

Yl = x1 · W1 + x2 · W2 + x3 · W3 + x4 (4)
M

Exponential estimation model:

Ye = x1 · W1x2 + x3 · W2x4 + x5 · W3x6 + x7 (5)

ED

where W1 , W2 , and W3 are the population, gross industrial production, and gross agri-
cultural production, respectively, and xi are the corresponding weights.
PT

In this paper, we propose a hybrid estimation model, which is a middle phase be-
tween linear and exponential models. The new model is defined by
CE

Yh = x1 · Yl + (1 − x1 ) · Ye (6)

where Yl and Ye are linear and exponential models, respectively, and x1 is the weighting
AC

factor. Based on Eqs. 4 and 5, the hybrid estimation model can be rewritten as follows.

Yh = x1 · (x2 · W1 + x3 · W2 + x4 · W3 + x5 )
  (7)
+ (1 − x1 ) x6 · W1x7 + x8 · W2x9 + x10 · W3x11 + x12

7
 ACCEPTED MANUSCRIPT

Table 2: The total water use, population, gross industrial production, and gross agricultural production in
Nanchang city from 2003 to 2015.
Total water Gross industrial Gross agricultural
Year Population
8 3 8
use (10 m ) production (10 yuan) production (108 yuan)

T
2003 24.21 4437476 250.95 51.29

2004 26.22 4469671 306.08 99.11

IP
2005 28.14 4500672 374.93 115.76

2006 27.71 4530776 448.15 124.58

CR
2007 32.55 4563025 532.75 142.84

2008 30.42 4597936 676.61 171.14

2009 33.42 4632067 753.20 187.20

2010

2011

2012
30.87

31.26

28.82
5042567

5088996

5131564
US 952.75

1223.72

1290.93
204.66

229.70

249.35
AN
2013 32.62 5184231 1398.63 266.12

2014 31.42 5240179 1500.70 283.63

2015 30.64 5302914 1619.50 296.92

M

4. Proposed approach

4.1. Dynamic parameter strategy

ED

Like PSO, the performance of FA is sensitive to its control parameters. Different

parameter settings can make large differences in optimization performance. To tackle
PT

this issue, some improved parameter strategies were proposed. In [16], Fister Jr. et
al. presented a new FA variant called memetic FA (MFA). In MFA, the step factor
α is automatically adjusted with increasing of generations. Experimental results show
CE

MFA is significantly better than FA on several classical benchmark functions. Gandomi

et al. [19] used different chaotic maps to update the parameters γ and β0 . Simulation
AC

results demonstrate that the Gauss map may be a good choice. Yu et al. [48] pointed out
that a large α can explore new search space, and a small α is beneficial for local search.
In [48], a dynamic model was designed to adjust step factor α. Cheung et al. [10]
presented an adaptive FA (AdaFa), which defined an adaptive distance based absorption
coefficient to update γ. In addition, five different parameter strategies were used to

8
 ACCEPTED MANUSCRIPT

Table 3: Results for different k values.

k = 0.2 k = 0.4 k = 0.6 k = 1.0

Mean Mean Mean Mean

Sphere 1.73E–40 1.64E–35 9.53E–24 5.71E–14

T
Schwefel 2.22 1.07E–20 6.92E–14 2.34E–07 8.98E–05

Rosenbrock 2.69E+01 2.88E+01 2.94E+01 2.92E+01

IP
Rastrigin 4.65E+01 4.97E+01 5.23E+01 5.46E+01

Ackley 3.61E–14 1.83E–14 7.21E–13 4.47E–08

CR
Penalized 1.06E–22 9.54E–14 2.63E–10 4.92E–06

adjust α. Besides the above work, there are some other parameter strategies [18, 21].

US
In our previous study [42], we analyzed the relationship between the step factor α
and convergence. If FA is convergent, α should satisfy the following condition.
AN
lim α = 0 (8)
t→∞

According to Eq. 8, we recently designed a new dynamic parameter strategy [43].

Compared to the standard FA, the step factor α is not fixed, and it can be automatically
M

adjusted as follows. !
t
α(t + 1) = α(t) · exp −k · (9)
Gmax
ED

where k ∈ (0, 1] is called decreasing rate, which can adjust the decreasing speed of α.
In some recent literature, α was limited in the range [0, 1]. So, the initial α(0) is set to
0.5, which is the midpoint of the range.
PT

To investigate the proposed dynamic parameter model, Fig. 2 presents the changes
of the parameter α with growth of generations. At the beginning, α is equal to 0.5. As
CE

the generation increases, the value of α gradually decreases. When t is near to 800, α
tends to zero. It is obvious that our dynamic parameter model satisfies the convergence
condition Eq. 8.
AC

The decreasing rate k may affect the performance of the parameter model, differ-
ent k values are tested on six well-known benchmark functions. In this experiment,
MaxFEs and N are equal to 2.0E+05 and 20, respectively. The parameter k is set
to 0.2, 0.4, 0.6, and 1.0, respectively. Other parameters are described in Section 5.1.

9
 ACCEPTED MANUSCRIPT

Table 3 presents the computational results for different k values, where “Mean” repre-
sents mean best fitness value. The function names are given in the first column, and
the detailed definitions of these functions can be found in the literature [41]. From the
results, it can be seen that k = 0.2 achieves better solutions than other k values on most

T
test functions. So, k = 0.2 is used in the following experiments.

IP
4.2. Normalization

CR
In this paper, the historical data from 2003 to 2015 listed in Table 2 are used for
training and testing the estimation models for water demand. To eliminate the influ-
ences of different units of data, the normalization method is used. In Table 2, the total

US
water use, population, gross industrial production, and gross agricultural production
are normalized as follows.
AN
W − Wmin
W∗ = (10)
Wmax − Wmin
where W ∗ is the normalized value, W is the value to be normalized, Wmin and Wmax are
the minimum and maximal values for the corresponding variable, respectively.
M

4.3. Fitness evaluation function

ED

The data from 2003 to 2012 are used to optimize the weighting factors of the es-
timation models, and the rest of data (2013–2015) are applied to test the models. To
evaluate the quality of obtained weighting factors, sum of squared errors (SSE) is em-
PT

ployed to construct the fitness evaluation function.

m 
X 2
CE

f (X) = Y pre − Yact (11)

i=1

where Yact and Y pre are the actual and predicted water demand, respectively, and m is
the number of training samples.
AC

4.4. Framework of NDFA

In this paper, we propose a new dynamic FA (called NDFA) to estimate the water
demand of Nanchang city. In NDFA, the step factor α is dynamically adjusted during

10
 ACCEPTED MANUSCRIPT

Algorithm 2: Proposed Approach (NDFA)

1 Randomly generate N initial fireflies (solutions) {X1 (t), X2 (t), . . . , XN (t)} and set t = 0;
2 Evaluate the fitness values of all N initial fireflies according to Eq. 11;
3 FEs = N;
4 while t ≤ Gmax && FEs ≤ MaxFEs do

T
5 Update α according to Eq. 9;
6 for i = 1 to N do

IP
7 for j = 1 to N do
8 if f (X j (t)) < f (Xi (t)) then
9 Move Xi (t) toward X j (t) according to Eq. 3;

CR
10 Compute the fitness value of Xi (t) according to Eq. 11;
11 FEs++;
12 end
13 end
14

15

16

17
end
end
t++;

Output the best solution;

US
AN

the search process. By analyzing the historical water use of Nanchang, three estima-
tion models in different forms are developed. Moreover, the normalization method is
M

utilized to eliminate the effects of different units of test data. The framework of NDFA
is described in Algorithm 2.
ED

5. Simulation experiments
PT

5.1. Experimental setup

In the experiments, our approach NDFA is applied to estimate the water demand in
Nanchang city. The performance of NDFA is compared with four other FA variants.
CE

The involved FAs are listed as below.

• The standard FA [45].

AC

• Memetic FA (MFA) [16].

• Variable step size FA (VSSFA) [48].

• FA with adaptive parameters (ApFA) [42].

11
 ACCEPTED MANUSCRIPT

Table 4: Results for the linear estimation model.

Algorithm Best MRE Mean MRE Std Worst MRE

FA 4.94% 4.97% 3.60E–04 5.05%

MFA 4.92% 4.96% 2.05E–04 4.98%

T
VSSFA 4.93% 4.97% 2.42E–04 5.01%

ApFA 4.90% 4.96% 2.38E–04 4.97%

IP
NDFA 4.89% 4.95% 2.33E–04 4.96%

CR
• Our approach NDFA.

For the above five algorithms, the parameters N and MaxFEs are equal to 30 and

US
1.0E+05, respectively. In the standard FA, α and β0 are equal to 0.5 and 1.0, respec-
tively. For MFA, the initial α(0), γ, β0 , and βmin are set to 0.5, 1.0, 1.0, and 0.2,
respectively. Both β0 and γ are set to 1.0 for VSSFA [48]. For ApFA, α(0) and β0 (0)
AN
are equal to 0.5 and 1.0, respectively [42]. In NDFA, the initial α(0), β0 , and γ are
1 1
equal to 0.5, 1.0, and Γ2
, respectively. In FA, ApFA, and NDFA, γ is set to Γ2
, where Γ
is the length of search range.
M

Data from 2003 to 2012 listed in Table 2 are used to optimize the weighting factors
of the estimation models, and the rest data (2013–2015) are applied to test the models.
ED

For each model, each algorithm is run 20 times and mean results are recorded. In the
experiments, we use relative error (RE) and mean relative error (MRE) to measure the
performance of FA.
Y pre − Yact
PT

RE = (12)
Yact
n
1 X Y pre (i) − Yact (i)
MRE = · (13)
CE

n i=1 Yact (i)

where Y pre (i) and Yact (i) are the predicted and actual water demand on the ith test data,
respectively, and n is the number of test data.
AC

5.2. Results

Tables 4, 5, and 6 present the computational results for the linear, exponential,
and hybrid estimation models, respectively, where “MRE” represents the mean relative
error and “Std” indicates the standard deviation. As seen, all five FA variants can

12
 ACCEPTED MANUSCRIPT

Table 5: Results for the exponential estimation model.

Algorithm Best MRE Mean MRE Std Worst MRE

FA 2.42% 2.53% 1.95E–03 3.00%

MFA 2.36% 2.40% 3.70E–04 2.47%

T
VSSFA 2.32% 2.55% 1.65E–03 2.80%

ApFA 2.29% 2.35% 2.60E–04 2.37%

IP
NDFA 2.27% 2.35% 3.46E–04 2.38%

CR
Table 6: Results for the hybrid estimation model.
Algorithm Best MRE Mean MRE Std Worst MRE

FA 2.25% 3.55% 1.39E–02 5.76%

MFA

VSSFA

ApFA
2.22%

2.24%

2.09%
US
2.82%

2.97%

2.81%
7.04E–03

1.32E–02

7.51E–03
4.25%

6.21%

4.36%
AN
NDFA 2.09% 2.79% 6.98E–03 4.06%

achieve good results for forecasting the water demand from 2013–2015. The best MRE
is 2.09% and the worst one is only 6.21%. It means that the prediction accuracy is
M

between 93.79% and 97.91%.

For linear and hybrid models, the mean MRE obtained by NDFA is better than
ED

other algorithms. Both ApFA and NDFA have the same mean MRE on the exponential
model. For the best MRE, NDFA achieves the best results on all models (ApFA and
NDFA have the same best MRE on the hybrid model).
PT

According to the obtained mean MRE, the performance of the three estimation
models can be sorted into the following order: exponential, hybrid, and linear models.
CE

The results confirm our design idea, which aims to provide a middle phase between the
linear and exponential models. For all FA variants, the mean MRE on the hybrid model
is better than the linear one, but worse than the exponential one. For the hybrid model,
AC

the obtained best MRE is better than other two models. It demonstrates that combining
the linear and exponential models can achieve a better prediction accuracy.
Tables 7, 8, and 9 show the actual water use and the best predicted water demand
for the linear, exponential, and hybrid estimation models, respectively, where “RE” in-

13
 ACCEPTED MANUSCRIPT

Table 7: The actual water use (108 m3 ) and the best predicted water demand (108 m3 ) for the linear estimation
model.
Year Actual FA MFA VSSFA ApFA NDFA

Pre RE Pre RE Pre RE Pre RE Pre RE

T
2013 32.62 32.28 1.04% 32.27 1.07% 32.27 1.06% 32.25 1.12% 32.25 1.14%

2014 31.42 32.91 4.75% 32.90 4.71% 32.90 4.73% 32.88 4.66% 32.88 4.64%

IP
2015 30.64 33.40 9.02% 33.39 8.99% 33.40 9.00% 33.37 8.92% 33.37 8.90%

Average 4.94% 4.92% 4.93% 4.90% 4.89%

CR
Table 8: The actual water use (108 m3 ) and the best predicted water demand (108 m3 ) for the exponential
estimation model.
Year

2013
Actual

32.62
Pre

31.24
FA

RE

4.23%
Pre

31.21
MFA

US RE

4.32%
VSSFA

Pre

31.38
RE

3.79%
Pre

31.41
ApFA

RE

3.70%
Pre

31.11
NDFA

RE

4.62%
AN
2014 31.42 31.46 0.14% 31.42 0.01% 31.53 0.36% 31.47 0.16% 31.20 0.69%

2015 30.64 31.53 2.90% 31.48 2.75% 31.50 2.82% 31.56 3.01% 31.10 1.50%

Average 2.42% 2.36% 2.32% 2.29% 2.27%

M

Table 9: The actual water use (108 m3 ) and the best predicted water demand (108 m3 ) for the hybrid estimation
model.
Year Actual FA MFA VSSFA ApFA NDFA
ED

Pre RE Pre RE Pre RE Pre RE Pre RE

2013 32.62 31.11 4.41% 31.21 4.31% 31.28 4.11% 31.34 3.93% 31.35 3.90%

2014 31.42 31.20 0.51% 31.38 0.13% 31.33 0.27% 31.39 0.08% 31.42 0.00%
PT

2015 30.64 31.10 1.82% 31.32 2.23% 31.35 2.32% 31.33 2.25% 31.36 2.36%

Average 2.25% 2.22% 2.23% 2.09% 2.09%

CE

dicates the relative error. For year 2013, the linear model achieves the best prediction
accuracy, while the other two models obtain larger RE values. For years 2014 and
2015, the exponential and hybrid models are better than the linear one. Fig. 3 presents
AC

the comparison between actual and estimated values for water demand. As shown, all
FA variants can achieve a good matching for the actual water use.
In order to estimate the water demand from 2017 to 2020, we need to determine
the annual average growth rates of population, gross industrial production, and gross

14
 ACCEPTED MANUSCRIPT

agricultural production. Based on the statistical data of Nanchang [33, 34], the average
growth rates in the last two years are used. During 2017–2020, the annual average
growth rates of population, gross industrial production, and gross agricultural produc-
tion are 0.682%, 9.30%, and 4.00%, respectively.

T
Fig. 4 presents the estimated water demand from 2017 to 2020. For the linear

IP
model, the estimated water demand gradually increases, and all FA variants almost
achieve the same predicted values. For the exponential model, the water demand pre-

CR
dicted by FA, MFA, and ApFA increases during 2017–2020, while VSSFA and NDFA
achieve opposite results. For the hybrid model, the water demand estimated by ApFA
and NDFA maintains a fixed level, which is between 31.52 × 108 m3 and 31.86 × 108 m3 .

6. Conclusions US
AN
In this paper, we propose a new dynamic FA (NDFA) for demand estimation of
water resources in Nanchang city of China. In NDFA, the step factor α is not fixed
and it can be dynamically updated during the search process. Three estimation models
M

based on linear, exponential and hybrid forms are developed according to the historical
water use and local economic structure. In addition, the normalization method is used
to eliminate the effects of different units of test data.
ED

In the experiments, data from 2003 to 2012 are used to optimize the weighting fac-
tors of the estimation models, and the rest of data (2013–2015) are applied to test the
models. Based on the historical data, the annual average growth rates of population,
PT

gross industrial production, and gross agricultural production are used to estimate the
water demand of Nanchang during 2017–2020. The performance of NDFA is com-
CE

pared with FA, MFA, VSSFA, and ApFA. Simulation results show that all FAs can
achieve good accuracy for forecasting the water demand from 2013–2015. The pre-
diction accuracy is between 93.79% and 97.91%. Among five FA variants, NDFA
AC

achieves the best prediction accuracy on three models.

This paper only uses three factors (population, gross industrial production, and
gross agricultural production) to construct the estimation model. However, there are
some uncertain factors, such as water price, weather, and environment, which may

15
 ACCEPTED MANUSCRIPT

affect the water demand. These will be further investigated in the future work.

Acknowledgment

This work was supported by the National Natural Science Foundation of China (No.

T
61663028, 61603163, 51669014, and 61461032), the Natural Science Foundation of

IP
Jiangxi Province (No. 20171BAB202035), and the Distinguished Young Talents Plan
of Jiangxi Province (No. 20171BCB23075).

CR
References

US
[1] M.A. Al-Zahrani and A. Abo-Monasar, Urban residential water demand predic-
tion based on artificial neural networks and time series models, Water Resources
Management 29(10) (2015) 3651–3662.
AN
[2] B. Amiri, L. Hossain, J. W. Crawford, R. T. Wigand, Community detection
in complex networks: Multi-objective enhanced firefly algorithm, Knowledge-
M

Based Systems 46 (2013) 1–11.

[3] F. Arbués, M.A. García-Valiñas, R. Martínez-Espiñeira, Estimation of residential

ED

water demand: a state-of-the-art review, The Journal of Socio-Economics 32(1)

(2003) 81–102.

[4] E. Assareh, M.A. Behrang, M.R. Assari, A. Ghanbarzadeh, Application of PSO

PT

(particle swarm optimization) and GA (genetic algorithm) techniques on demand

estimation of oil in Iran, Energy 35 (2010) 5223–5229.
CE

[5] Y. Bai, P. Wang, C. Li, J.J. Xie, Y. Wang, A multi-scale relevance vector regres-
sion approach for daily urban water demand forecasting, Journal of Hydrology
AC

517 (2014) 236–245.

[6] B.M. Brentan, E. Luvizotto Jr, M. Herrera, J. Izquierdo, R. Pérez-García, Hybrid

regression model for near real-time urban water demand forecasting, Journal of
Computational and Applied Mathematics 309 (2017) 532–541.

16
 ACCEPTED MANUSCRIPT

[7] X.J. Cai, X.Z. Gao, Y. Xue, Improved bat algorithm with optimal forage strategy
and random disturbance strategy, International Journal of Bio-inspired Computa-
tion 8(4) (2016) 205–214.

[8] X.J. Cai, H. Wang, Z.H. Cui, J.H. Cai, Y. Xue and L. Wang, Bat algorithm with

T
triangle-flipping strategy for numerical optimization, International Journal of Ma-

IP
chine Learning and Cybernetics, 2017, doi: 10.1007/s13042-017-0739-8, to be
published.

CR
[9] J. Caiado, Performance of combined double seasonal univariate time series mod-
els for forecasting water demand, Journal of Hydrologic Engineering 15(3) (2009)
215–222.
US
[10] N.J. Cheung, X.M. Ding, H.B. Shen, Adaptive firefly algorithm: parameter anal-
AN
ysis and its application, PloS one 9(11) (2014) e112634.

[11] L.Z Cui, G.H. Li, X.Z. Wang, Q.Z. Lin, J.Y. Chen, N. Lu, J. Lu, A ranking-
based adaptive artificial bee colony algorithm for global numerical optimization,
M

Information Sciences 417 (2017) 169–185.

[12] Z.H. Cui, B. Sun, G.G. Wang, Y. Xue, J.J. Chen, A novel oriented cuckoo search
ED

algorithm to improve DV-Hop performance for cyber-physical systems, Journal

of Parallel and Distributed Computing 103 (2017) 42–52.
PT

[13] M.H. Davijani, M.E. Banihabib, A.N. Anvar, S.R. Hashemi, Optimization model
for the allocation of water resources based on the maximization of employment in
the agriculture and industry sectors, Journal of Hydrology 533 (2016) 430–438.
CE

[14] M. Dorigo, V. Maniezzo, A. Colorni, The ant system: optimization by a colony

of cooperating agents, IEEE Transactions on Systems, Man and Cybernetics-Part
AC

B: Cybernetics 26 (1996) 29–41.

[15] I. Fister, I. Fister Jr., X.S. Yang, J. Brest, A comprehensive review of firefly algo-
rithms, Swarm and Evolutionary Computation 13 (2013) 34–46.

17
 ACCEPTED MANUSCRIPT

[16] I. Fister Jr., X.S. Yang, I. Fister, J. Brest, Memetic firefly algorithm for combi-
natorial optimization, Bioinspired Optimization Methods and their Applications
(BIOMA), 2012, pp. 1–14.

[17] A.P. Florence and V. Shanthi, A load balancing model using firefly algorithm in

T
cloud computing, Journal of Computer Science 10(7) (2014) 1156–1165.

IP
[18] A. Gálvez and A. Iglesias, New memetic self-adaptive firefly algorithm for contin-
uous optimisation, International Journal of Bio-Inspired Computation 8(5) (2016)

CR
300–317.

[19] A.H. Gandomi, X.S. Yang, S. Talatahari, A.R. Alavi, Firefly algorithm with

US
chaos, Communications in Nonlinear Science and Numerical Simulation 18(1)
(2013) 89–98.
AN
[20] D. Karaboga, B. Akay, A comparative study of artificial bee colony algorithm,
Applied Mathematics and Computation 214 (2009) 108–132.

[21] H. Kasdirin, Adaptive bio-inspired firefly and invasive weed algorithms for global
M

optimisation with application to engineering problems, doctoral dissertation, Uni-

versity of Sheffield, UK. 2016.
ED

[22] A. Kazem, E. Sharifi, F. Hussain, M. Saberi, O. K. Hussain, Support vector re-

gression with chaos-based firefly algorithm for stock market price forecasting,
Applied Soft Computing 13(2) (2013) 947–958.
PT

[23] J. Kennedy, R.C. Eberhart, Particle swarm optimization, Proc. IEEE Int. Conf.
Neural Networks, 1995, pp. 1942–1948.
CE

[24] H. Li, C. Zhang, D. Miao, T. Wang, Y. Feng, H. Fu, C. Zhang, M. Zhao, Water
demand prediction of grey Markov model based on GM(1,1), The 3rd Interna-
AC

tional Conference on Mechatronics and Information Technology (ICMIT), 2016,

pp. 524–529.

[25] J. Li, Y. Zhang, X. Chen, Y. Xiang, Secure attribute-based data sharing for
resource-limited users in cloud computing, Computers & Security 72 (2018) 1–
12.

18
 ACCEPTED MANUSCRIPT

[26] P. Li, J. Li, Z. Huang, C. Gao, W. Chen, K. Chen, Privacy-preserving out-

sourced classification in cloud computing, Cluster Computing, 2017, doi:
10.1007/s10586-017-0849-9, to be published.

[27] L. Lv and J. Zhao, The firefly algorithm with Gaussian disturbance and local

T
search, Journal of Signal Processing Systems, 2017, doi: 10.1007/s11265-017-

IP
1278-y, to be published.

CR
[28] M.K. Marichelvam and M. Geetha, A hybrid discrete firefly algorithm to solve
flow shop scheduling problems to minimise total flow time, International Journal
of Bio-Inspired Computation 8(5) (2016) 318–325.

US
[29] B. Nasiri and M.R. Meybodi, History-driven firefly algorithm for optimisation in
dynamic and uncertain environments, International Journal of Bio-Inspired Com-
AN
putation 8(5) (2016) 326–339.

[30] P.J. Oliveira, J.L. Steffen, P. Cheung, Parameter estimation of seasonal Arima
models for water demand forecasting using the harmony search algorithm, Pro-
M

cedia Engineering 186 (2017) 177–185.

[31] I. Pulido-Calvo and J.C. Gutiérrez-Estrada, Improved irrigation water demand

ED

forecasting using a soft-computing hybrid model, Biosystems Engineering 102(2)

(2009) 202–218.
PT

[32] M. Romano and Z. Kapelan, Adaptive water demand forecasting for near real-
time management of smart water distribution systems, Environmental Modelling
& Software 60 (2014) 265–276.
CE

[33] Statistic Bureau of Jiangxi: Jiangxi statistical yearbook, Chinese Statistical Press,
Beijing, 2004–2016.
AC

[34] Statistics Bureau of Nanchang: Statistical bulletin of national economic and so-
cial development of Nanchang, Nanchang, 2003–2015.

[35] O.P. Verma, D. Aggarwal, T. Patodi, Opposition and dimensional based modified
firefly algorithm, Expert Systems with Applications 44 (2016) 168–176.

19
 ACCEPTED MANUSCRIPT

[36] F. Wang, Y. Zhang, Q. Rao, H. Zhang, Exploring mutual information based senti-
mental analysis with kernel based extreme learning machine for stock prediction,
Soft Computing 21(12) (2017) 3193–3205.

[37] H. Wang, Z.H. Cui, H. Sun, S. Rahnamayan, X.S. Yang, Randomly attracted

T
firefly algorithm with neighborhood search and dynamic parameter adjustment

IP
mechanism, Soft Computing 21(18) (2017) 5325–5339.

CR
[38] H. Wang, Z.H. Cui, W.J. Wang, X.Y. Zhou, J. Zhao, L. Lv, H. Sun, Firefly algo-
rithm for demand estimation of water resources, The 24th International Confer-
ence on Neural Information Processing (ICONIP 2017), LNCS 10637, 2017, pp.
11–20.
US
[39] H. Wang, H. Sun, C. Li, S. Rahnamayan, J.S. Pan, Diversity enhanced particle
AN
swarm optimization with neighborhood search, Information Sciences 223 (2013)
119–135.

[40] H. Wang, W.J. Wang, H. Sun, S. Rahnamayan, Firefly algorithm with random
M

attraction, International Journal of Bio-Inspired Computation 8(1) (2016) 33–41.

[41] H. Wang, W.J. Wang, X.Y. Zhou, H. Sun, J. Zhao, X. Yu, Z.H. Cui, Firefly algo-
ED

rithm with neighborhood attraction, Information Sciences 382-383 (2017) 374–

387.
PT

[42] H. Wang, X.Y. Zhou, H. Sun, X. Yu, J. Zhao, H. Zhang, L.Z. Cui, Firefly al-
gorithm with adaptive control parameters, Soft Computing 21(17) (2017) 5091–
5102.
CE

[43] W.J. Wang, H. Wang, X.Y. Zhou, J. Zhao, L. Lv, H. Sun, Dynamic step factor
based firefly algorithm for optimization problems, IEEE International Conference
AC

on Computational Science and Engineering, 2017, pp. 128–134.

[44] H. Wu, L. Kuang, F. Wang, Q. Rao, M. Gong, Y. Li, A Multiobjective box-

covering algorithm for fractal modularity on complex Networks, Applied Soft
Computing 61 (2017) 294–313.

20
 ACCEPTED MANUSCRIPT

[45] X.S. Yang, Nature-inspired metaheuristic algorithms, Luniver Press, UK, 2008.

[46] X.S. Yang, Engineering optimization: an introduction with metaheuristic appli-

cations, John Wiley & Sons, USA, 2010.

T
[47] X.S. Yang, S.S.S. Hosseini, A.H. Gandomi, Firefly algorithm for solving non-
convex economic dispatch problems with valve loading effect, Applied Soft Com-

IP
puting 12(3) (2012) 1180–1186.

CR
[48] S.H. Yu, S.L. Zhu, Y. Ma, D.M. Mao, A variable step size firefly algorithm for
numerical optimization, Applied Mathematics and Computation 263 (2015) 214–
220.

US
[49] M.Q. Zhang, H. Wang, Z.H. Cui and J.J. Chen, Hybrid multi-objective
cuckoo search with dynamical local search, Memetic Computing, 2017, doi:
AN
10.1007/s12293-017-0237-2, to be published.
M
ED
PT
CE
AC

21
 ACCEPTED MANUSCRIPT

T
IP
CR
US
AN
M
ED

Industrial water use

28%
PT

Residential water use

12%
CE

Ecological water use

3%

Agricultural water use

57%
AC

Figure 1: The distribution of average water use in different departments.

22
 ACCEPTED MANUSCRIPT

T
IP
CR
0.5

0.4
US
AN
0.3

0.2
M

0.1

0.0
ED

0 200 400 600 800 1000

Generations (t)
PT

Figure 2: The changes of α with growth of generations.

CE
AC

23
 ACCEPTED MANUSCRIPT

35 35

30 30

T
Water use (10 m )

Water use (10 m )

3

3
25 25
8

IP
20 Actual Data 20 Actual Data

FA-linear MFA-linear

FA-exponential MFA-exponential

FA-hybrid MFA-hybrid
15 15
FA-linear (2013-2015) MFA-linear (2013-2015)

CR
FA-exponential (2013-2015) MFA-exponential (2013-2015)

FA-hybrid (2013-2015) MFA-hybrid (2013-2015)

10 10
2002 2004 2006 2008 2010 2012 2014 2016 2002 2004 2006 2008 2010 2012 2014 2016
Year Year

(a) FA (b) MFA

35

30
US 35

30
AN
Water use (10 m )

Water use (10 m )

3

25 25
8

20 Actual Data 20 Actual Data

VSSFA-linear ApFA-linear

VSSFA-exponential ApFA-exponential

VSSFA-hybrid ApFA-hybrid
15 15
M

VSSFA-linear (2013-2015) ApFA-linear (2013-2015)

VSSFA-exponential (2013-2015) ApFA-exponential (2013-2015)

VSSFA-hybrid (2013-2015) ApFA-hybrid (2013-2015)

10 10
2002 2004 2006 2008 2010 2012 2014 2016 2002 2004 2006 2008 2010 2012 2014 2016
Year Year
ED

(c) VSSFA (d) ApFA

35
PT

30
Water use (10 m )
3

25
8
CE

20 Actual Data

NDFA-linear

NDFA-exponential

NDFA-hybrid
15
NDFA-linear (2013-2015)

NDFA-exponential (2013-2015)

NDFA-hybrid (2013-2015)
10
2002 2004 2006 2008 2010 2012 2014 2016
AC

Year

(e) NDFA

Figure 3: Comparison between actual and estimated values for water demand.

24
 ACCEPTED MANUSCRIPT

40

38

36

T
Water use (10 m )
3

34
8

32

IP
30 FA

MFA

VSSFA
28
ApFA

CR
NDFA
26
2017 2018 2019 2020
Year

(a) Linear estimation model

34

32
US
AN
Water use (10 m )
3
8

30

FA
28
MFA
M

VSSFA

ApFA

NDFA
26
2017 2018 2019 2020
Year
ED

(b) Exponential estimation model

34
PT

32
Water use (10 m )
3
8

30
CE

FA
28
MFA

VSSFA

ApFA

NDFA
26
AC

2017 2018 2019 2020

Year

(c) Hybrid estimation model

Figure 4: Comparison for predicting water demand from 2017 to 2020.

25