Expert Systems With Applications 245 (2024) 123088

Contents lists available at ScienceDirect

Expert Systems With Applications

journal homepage: www.elsevier.com/locate/eswa

Football team training algorithm: A novel sport-inspired meta-heuristic

optimization algorithm for global optimization
Zhirui Tian a, Mei Gai b, c, *
a
School of Science and Engineering, The Chinese University of Hong Kong (Shenzhen), Shenzhen 518172, Guangdong, China
b
Key Research Base of Humanities and Social Sciences of the Ministry of Education, Center for Studies of Marine Economy and Sustainable Development, Liaoning
Normal University, Dalian 116029, Liaoning, China
c
University Collaborative Institute Center of Marine Economy High-Quality Development of Liaoning Province, Dalian 116029, Liaoning, China

A R T I C L E I N F O A B S T R A C T

Keywords: A more efficient optimization algorithm has always been the pursuit of researchers, but the performance of the
Football team training algorithm current optimization algorithm in some complex test functions is not always satisfactory. In order to solve this
Wind speed prediction problem, a new meta-heuristic optimization algorithm—Football Team Training Algorithm (FTTA) is proposed
Unconstrained weighting method
according to the training method of the football team, which simulates the three stages of the training session:
Neural network
Data preprocessing strategy
Collective Training, Group Training and Individual Extra Training. By the test on two groups of test functions,
CEC2005 and CEC2020, the proposed optimization algorithm (FTTA) achieves the best results, which far exceeds
the traditional Grey Wolf Optimization (GWO), Whale Optimization Algorithm (WOA) algorithms and so on. In
the engineering application, a new hybrid wind speed prediction system is proposed based on FTTA. The FTTA is
used to optimize variational mode decomposition (VMD) to improve the effect of data denoising. At the same
time, based on unconstrained weighting algorithm, FTTA and combination prediction model build a new hybrid
prediction strategy. Through the experiments on four groups of wind speed data in Dalian, the accuracy, sta­
bility, advancement, and CPU running speed of the system are verified. It is obvious that the practical application
ability of the system is much better than previous methods, which can effectively improve the utilization effi­
ciency of renewable energy.

1. Introduction simulating the phenomena of reproduction, hybridization and mutation

in the process of natural selection and natural genetics and finally ob­
Meta-heuristic algorithm has always been one of the optimal ways to tained the optimal solution. The Particle Swarm Optimization is
solve complex functions. Traditional optimization methods are easy to designed by simulating the predation behavior of the flock of birds. All
fall into the local optimal solution, while the meta-heuristic algorithm the birds want to find the food, and they constantly transmit the infor­
can better escape from the local optimal solution through a series of mation of their respective locations to each other in the whole process of
behavior simulation and find the global optimal solution. At the same searching, to make other birds know the location of the food source.
time, meta-heuristic algorithm, as a black box algorithm, is very Finally, the whole flock of birds can gather around the food source to the
convenient to use, so it has been widely used as soon as it is released. global optimal solution is obtained.
Since the last century, scholars have made a lot of efforts on opti­ From 2010 to 2020, the optimization algorithm has made a new
mization algorithms. The earliest and most popular optimization algo­ breakthrough, and the optimization algorithm based on animal behavior
rithms are Genetic Algorithm (GA) (Chen et al., 1995) and Particle has gradually come into everyone’s field of vision, such as Grey Wolf
Swarm Optimization (PSO) (Parsopoulos & Vrahatis, 2007), and they Optimization (GWO) (Mirjalili et al., 2014), Whale Optimization Algo­
are still very popular today. Genetic algorithm is an optimization algo­ rithm (WOA) (Mirjalili & Lewis, 2016) and Grasshopper Optimization
rithm which simulates natural selection, it continuously eliminates Algorithm (GOA) (Saremi et al., 2017). WOA simulates the predation
backward individuals to maintain the good state of the population by mode of humpback whales. GOA is inspired by gray wolf hunting, which

* Corresponding author at: Key Research Base of Humanities and Social Sciences of the Ministry of Education, Center for Studies of Marine Economy and Sus­
tainable Development, Liaoning Normal University, Dalian 116029, Liaoning, China.
E-mail addresses: zhiruitian@link.cuhk.edu.cn (Z. Tian), gaimei71@lnnu.edu.cn (M. Gai).

https://doi.org/10.1016/j.eswa.2023.123088
Received 29 October 2023; Received in revised form 17 December 2023; Accepted 26 December 2023
Available online 2 January 2024
0957-4174/© 2023 Elsevier Ltd. All rights reserved.
Z. Tian and M. Gai Expert Systems With Applications 245 (2024) 123088

mainly simulates four types of grey wolves in nature: including leader­ agricultural and sideline products processing, wind power generation, wind
ship mechanism and hunting mechanism, to obtain the global optimal heating and so on (Wang et al., 2024). But wind speed prediction has al­
solution. The performance of these algorithms is much better than that ways been a difficult problem in wind power generation (Wang et al.,
of GA, DE etc. on most test functions, which further improved the per­ 2024). According to a research, accurate wind speed prediction can in­
formance of meta-heuristic algorithm. crease the utilization rate of wind power by at least 30 % (Tian & Gai,
In addition to some bionic algorithms, some algorithms that simulate 2023).
social behavior or physical theories are also popular, such as the Harmony Nowadays, various optimization algorithms are widely used in wind
Search (HS) (Lee & Geem, 2004) and the Multi-Verse Optimization (MVO) speed prediction models. For example, Xin et al. (2022) used SSA opti­
(Ali et al., 2016). The Multi-verse theory is the inspiration for the MVO mization algorithm to optimize BP neural network and achieved accu­
algorithm, which proposed that there are other universes besides the rate prediction results. Wang et al. (2023) effectively combined MOSCA
universes in which we live. The emergence of individuals in the universe is with the combined neural networks, which improved the wind speed
the result of a single huge explosion, and multiple huge explosions con­ prediction accuracy. Sun et al. (2024) proposed a multi-objective arti­
tributes to the birth of the entire multiverse population, and white hole, ficial hummingbird algorithm and applied it to the interval prediction of
black hole and wormhole are the three core concepts in multiverse theory. wind speed. The experimental results show that the algorithm signifi­
The Harmony Search is designed to simulate a orchestra, in which cantly improves the quality of the interval. Shi et al. (2024) adopted
everyone plays a musical instrument, and their performances add up to a sparrow search optimizer to optimize the wind speed prediction model
group of harmonies, and they will continue to cooperate and rehearse to and significantly improved the interpretability. Although the optimiza­
get the best harmony effect. Each group of harmonies they play is weighed tion algorithm optimizes the ability of the prediction model to some
by the commander-in-chief, and if they fail to meet the requirements, extent, the prediction effect of most models is still unsatisfactory, which
continue to play until they get a satisfactory set of harmonies. Thus, the is due to several reasons. On the one hand, the convergence ability of
global optimal solution is obtained. some optimization algorithms is not enough. In engineering applica­
In addition, there is a class of optimization algorithms which draw tions, it’s almost impossible to have enough time to carry out a large
inspiration from sports, because co-operation is often crucial for optimal number of iterations, so the optimization effect is very poor when the
performance in a team. For example, the Soccer Team Optimization (SGO) number of iterations is small. On the other hand, the optimization effect
(Purnomo & Wee, 2013) simulates the behavior of players and realizes of some meta-heuristic optimization algorithms vary greatly in different
powerful information sharing ability through two basic operators. The objective functions. The convergence effect is very good in some
Football Game Algorithm (FGA) (Fadakar & Ebrahimi, 2016) simulates objective functions, but almost no convergence in some other objective
the winning process of players in a game under the command of a coach. functions, which will cause very serious trouble. Due to the above
Volleyball Premier League Algorithm (VPL) (Moghdani & Salimifard, problems, the engineering application of optimization algorithms are
2018) is inspired from the competition and interaction of a volleyball team limited in wind speed prediction. Therefore, an optimization method
in a season, which can effectively solve the problem of complex search with good convergence effect and fast convergence speed is urgently
space; Team Game Algorithm (TGA) (Mahmoodabadi et al., 2018) simu­ needed in the wind speed prediction system.
lates all sports that require teamwork, such as basketball, and obtains Due to the excellent performance of FTTA, a new hybrid wind speed
excellent optimization results by the idea of game theory. prediction system is proposed based on this algorithm. First of all, a new
Although numerous meta-heuristic algorithms have been developed, FTTA-VMD data preprocessing method is proposed, which uses FTTA to
there are still some problems in the current algorithms. For example, the automatically optimize the super-parameters of variational mode
PSO algorithm is sensitive to the setting of parameters, if the parameters decomposition (VMD) (Tian & Gai, 2022) to reduce the impact of the
are not set properly, it is easy to lead to the deterioration of the opti­ influence of operator’s lack of experience on prediction accuracy as
mization results. However, in practical application, few people know much as possible. We also apply the combination prediction model (Tian
how to set the parameters correctly, resulting in the algorithm often & Wang, 2023), which effectively combines prediction models with
cannot make a difference. In addition, some algorithms resulting in poor different principles, and gives it dynamic weight through FTTA to
results in the process of solving some complex functions. For example, further improve its prediction accuracy and stability. Through four
GA works well at the initial stage of the algorithm, but it is weak in the groups of experiments and two groups of discussions, the prediction
later stage, so it is difficult to get the real global optimal solution. effect of the wind speed prediction system is verified. It is obvious that
In order to improve the above shortcomings, a new meta-heuristic the results are satisfactory.
optimization algorithm—Football Team Training Algorithm (FTTA) is The innovation of this paper is as follows:
proposed, with the simulation of football lessons, training players in A new meta-heuristic optimization algorithm is proposed.
stages to improve the ability of players to achieve the best optimization In this paper, a new meta-heuristic optimization algorithm is pro­
effect. The algorithm has a good convergence ability, its performance on posed. By simulating the training mode of the football team, the algo­
several groups of test functions (CEC2005, CEC2020) is much better rithm continuously improves the ability of the convergence, and finally
than that of classical algorithms such as GWO and WOA. In addition, the obtains the global optimal solution. FTTA algorithm is proved to have
algorithm has excellent CPU running speed, which further improves the strong performance in CEC test functions.
practical application ability of the algorithm. Excellent engineering application ability is showed.
In this paper, we verify the application ability of FTTA in practical The football team training algorithm has a good convergence effect
engineering problems by wind speed prediction. As a kind of clean energy, on almost all test functions. At the same time, it also has excellent
wind energy has been more and more widely used. Through the wind operation speed and can complete enough iterations in a short time. The
turbine, wind energy is converted into electric energy, thermal energy and engineering application showed that the method has strong practical
mechanical energy, which can be used in sailing, wind water lifting, application ability.

2
Z. Tian and M. Gai Expert Systems With Applications 245 (2024) 123088

Fig. 1. The diagram of collective training.

Fig. 2. The diagram of group training.

3
Z. Tian and M. Gai Expert Systems With Applications 245 (2024) 123088

Fig. 3. Full flow chart of the FTTA algorithm.

Advanced high-accuracy wind speed prediction system is the ability of FTTA by two groups of test functions and CPU running
obtained. speed. The third part is the construction strategy of the hybrid high-
Based on FTTA, a new data preprocessing method FTTA-VMD is accuracy wind speed prediction system. The fourth part mainly in­
proposed, which can directly provide users with the optimal hyper­ troduces the selection of data, the super parameters of the model and the
parameters of VMD through the optimization algorithm. At the same mathematical formula of the evaluation indexes; The fifth part is four
time, a new unconstrained weighting system (FTTA-CP) is developed, groups of experiments, focusing on verifying the performance of the
which weighs different prediction models according to statistical eval­ wind speed prediction system. The last part is the conclusion, in which
uation indexes, which not only improves the advanced nature of the we summarize the full text and discuss the future development direction.
model, but also ensures that the model will not suddenly deteriorate. The list of abbreviations in this paper are as follows:
The paper is divided into six parts, the second part mainly introduces
the proposed FTTA meta-heuristic optimization algorithm and verified

4
Z. Tian and M. Gai Expert Systems With Applications 245 (2024) 123088

Table 1
The function expression of CEC2005.
Objective Function Dim Range Fmin
∑
CEC2005_F1 f1 (x) = ni=1 x2i 30 [-100,100] 0
∑ ∏
CEC2005_F2 f2 (x) = ni=1 |xi | + ni=1 |xi | 30 [-10,10] 0
CEC2005_F3 ∑n (∑i 2
)2
30 [-100,100] 0
f3 (x) = i=1 j− 1 xj

CEC2005_F4 f4 (x) = maxi {|xi |, 1⩽i⩽n } 30 [-100,100] 0

CEC2005_F5 f5 (x) =
∑n− 1 2 2 2 30 [-30,30] 0
i=1 100(xi+1 − xi ) + (xi − 1)
∑
CEC2005_F6 f6 (x) = ni=1 [xi + 0.5]2 30 [-100,100] 0
∑
CEC2005_F7 f7 (x) = ni=1 ix4i + random(0, 1) 30 [-1.28.1.28] 0
∑ ( √̅̅̅̅̅̅̅ )
CEC2005_F8 f8 (x) = ni=1 − xi sin |xi | 30 [-500,500] − 12569.5
∑n [ 2 ]
CEC2005_F9 f9 (x) = i=1 xi − 10cos(2πxi ) + 10 30 [-5.12,5.12] 0
√ ̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
̅ )
CEC2005_F10 ( 1∑n 2 ) ( 1∑n 30 [–32,32] 0
f10 (x) = − 20exp − 0.2 xi − exp cos(2πxi )) + 20 + e
n i=1 n i=1

CEC2005_F11 1 ∑n 2 ∏n xi 30 [-600,600] 0
f11 (x) = x − cos( √ ̅ ) + 1
4000 i=1 i i=1
i
⎧ ⎫
CEC2005_F12 ⎪
⎪ ⎪
⎪ 30 [-50,50] 0
⎪
⎪
⎪ 10sin(πyi ) ⎪
⎪
⎪
⎪ ⎪
π ⎨ ∑n− 1 ⎬ ∑ x +1
f12 (x) = + i=1 (yi − 1) [1 + 10sin2 (πyi+1 )] + ni=1 u(xi , 10, 100, 4)yi = 1 + i
2
n⎪⎪ ⎪
⎪ 4
⎪ ⎪
⎪
⎪
⎪
⎩ +(yn − 1)2 ⎪
⎪
⎪
⎭
⎧ ∑n ⎫
CEC2005_F13 ⎪
⎪ sin2 (3πxi ) + (x − 1)2 ⎪ ⎪ 30 [-50,50] 0
⎪
⎨ i=1 i ⎪
⎬ ∑
f13 (x) = 0.1 [1 + sin2 (3πxi + 1)]*(xn − 1)2 + ni=1 (xi , 5, 100, 4)
⎪
⎪ ⎪
⎪
⎪ ⎪
⎩ [1 + sin2 (3πxn ) ⎭
)
CEC2005_F14 ( − 1 2 [-65,65] 0.998
1 ∑25 1
f14 (x) = + ∑
500 j=1
j + 2i=1 (xi − aij )6
CEC2005_F15 [
⎤2 4 [-5,5] 0.0003
∑ xi (b2i + bi x2 ) 2
f15 (x) = 11 i=1 ai − bi + bi x3 + x4 ⎦
bi + bi x3 + x4
2

CEC2005_F16 1 2 [-5,5] − 1.0316

f16 (x) = 4x21 − 2.1x41 + x61 + x1 x2 − 4x22 + 4x42
3 )
CEC2005_F17 ( 5.1 2 5 1 2 [-5,5] 0.398
f17 (x) = x2 − x1 + x1 − 62 + 10(1 − )cosx1 + 10
4π 2 π 8π
CEC2005_F18 f18 (x) = [1 + (x1 + x2 + 1)2 (19 − 14x1 + 3x21 2 [-2,2] 3
− 14x2 + 6x1 x2 + 3x22 )]*[30 + (2x1 − 3x2 )2 *
(18 − 32x1 + 12x21 + 48x2 − 36x1 x2 + 27x22 )]
∑4 ∑3
CEC2005_F19 f19 (x) = − i=1 ci exp(− j=1 aij (xj − pij ) )
2 3 [1,3] − 3.86
( )
CEC2005_F20 ∑4 ∑6 2 6 [0,1] − 3.32
f20 (x) = − i=1 ci exp(− j=1 aij (xj − pij )
∑5
CEC2005_F21 f21 (x) = − T
i=1 (X − ai )(X − ai ) + ci
− 1 4 [0,10] − 10.1532
∑7
CEC2005_F22 f22 (x) = − T
i=1 (X − ai )(X − ai ) + ci
− 1 4 [0,10] − 10.4028
∑10
CEC2005_F23 f22 (x) = − i=1 (X − ai )(X − ai )T
+ c − 1
i
4 [0,10] − 10.5363

List of abbreviations. football training class is divided into three parts: Collective training,
Group training and Individual extra training. The specific behavior and
VMD Variational mode ARIMA Autoregressive Integrated formula are as follows:
decomposition Moving Average
BP BP Neural Network GRU Gate Recurrent Unit
2.1.1. Collective training
CNN Convolutional Neural GWO Grey Wolf Optimization
Network At the beginning of the training, the players will conduct collective
DE Differential Evolution PSO Particle Swarm Optimization training under the guidance of the coaches, who will first let the players
SGO Soccer Team Optimization FGA Football Game Algorithm know their own level through a series of tests (fitness function). Then the
VPL Volleyball Premier League TGA Team Game Algorithm players will make their own collective training plan according to their
Algorithm
WOA Whale Optimization RMSE Root Mean Square Error
own level. We divide players into four different types: Followers, Dis­
Algorithm coverers, Thinkers and Volatilities. In each iteration, players will
PICP PI coverage probability NMSE Normalized Mean Square Error randomly change their own types.
MAPE Mean Absolute Percentage MAE Mean Absolute Error Followers
Error
Followers are enthusiastic followers of the current best player, and
MSE Mean Square Error SSE Sum of squares due to error
he works hard towards the best player in every dimension, hoping to
reach the level of the current best player, but due to the limitation of
strength, often each dimension can only move randomly to the best
player. The equation is as follows:
2. Football team training algorithm k
Fi,j k
new = Fi,j old + rand × ( FkBest,j − Fi,j
k
old) (1)
2.1. Methodology
k
where the current best player is defined as FBest , where k is the number of
k
Football Team Training Algorithm (FTTA) is realized by simulating iterations, FBest,j is its value on dimension j; the current player is defined
the behavior of players in a high-level football training class. Usually, a

5
Z. Tian and M. Gai Expert Systems With Applications 245 (2024) 123088

Table 2A
The test results of CEC2005 (F1 - F12).
Mean Best Worst Std Mean Best Worst Std

CEC2005_F1 CEC2005_F2
PSO 179.2816 42.65244 602.2784 109.4029 PSO 34.06093 7.564448 86.9716 18.17329
DE 43.49036 4.25E-07 1518.494 224.7077 DE 1.203645 1.83E-06 20.00088 3.730586
GWO 1.65E-33 1.83E-34 8.42E-33 2.09E-33 GWO 7.59E-20 1.14E-20 2.41E-19 6.25E-20
WOA 5.54E-85 2.03E-96 5.62E-84 1.64E-84 WOA 6.70E-54 3.75E-60 1.28E-52 2.87E-53
FTTA 2.0335E-184 1.0119E-276 2.0318E-182 0 FTTA 7.3433E-106 2.4081E-142 5.5175E-104 5.6902E-105
CEC2005_F3 CEC2005_F4
PSO 8547.311 1473.866 26360.35 6299.691 PSO 6.779684 3.166777 12.95799 2.095797
DE 7664.078 370.0237 24821.57 7040.677 DE 57.31421 34.33122 82.25686 13.08247
GWO 5.96E-08 2.41E-11 5.87E-07 1.53E-07 GWO 1.32E-08 3.39E-10 3.68E-08 9.63E-09
WOA 29350.53 7282.497 51903.75 11705.66 WOA 38.78278 4.68E-05 86.48047 28.93503
FTTA 4.45E-166 3.05E-251 4.45E-164 0 FTTA 1.47E-94 1.43E-131 6.02E-93 8.68E-94
CEC2005_F5 CEC2005_F6
PSO 15821.5 675.8881 96560.34 23958.14 PSO 0.163703 0.002624 1.163627 0.198393
DE 46508.75 72.77292 787,732 177084.6 DE 127.7037 7.40E-07 844.7566 269.2271
GWO 26.55697 25.27304 27.93929 0.729583 GWO 0.582519 3.50E-05 1.248794 0.368
WOA 27.57139 26.98362 28.72361 0.451081 WOA 0.071486 0.012628 0.260498 0.076391
FTTA 21.75252 20.52372 22.69543 0.642508 FTTA 0 0 0 0
CEC2005_F7 CEC2005_F8
PSO 0.595702 0.034712 10.86252 1.822274 PSO − 7743.89 − 9439.24 − 5399.83 978.7279
DE 1.115112 0.210989 6.966475 1.527486 DE − 8209.51 − 9204.48 − 7314.32 536.6523
GWO 0.001019 0.000309 0.002436 0.000461 GWO − 6412.75 − 7569.59 − 5009.23 627.4872
WOA 0.001459 0.000132 0.005572 0.001505 WOA − 11239.8 − 12569 − 8475.21 1636.519
FTTA 0.00178 0.000536 0.003812 0.00083 FTTA ¡12415 ¡12569.4 ¡11616.5 236.8746
CEC2005_F9 CEC2005_F10
PSO 201.0894 152.2727 282.7807 32.93316 PSO 6.906082 3.136277 19.96677 4.58151
DE 88.27844 51.04062 157.0982 27.32999 DE 19.96329 19.95721 19.96677 0.002801
GWO 2.089344 0 11.35414 3.327346 GWO 4.55E-14 4.00E-14 5.77E-14 4.53E-15
WOA 0 0 0 0 WOA 4.09E-15 8.88E-16 7.99E-15 2.55E-15
FTTA 0 0 0 0 FTTA 8.88E-16 8.88E-16 8.88E-16 0
CEC2005_F11 CEC2005_F12
PSO 2.35 1.71 4.58 0.691258 PSO 4.488788 1.47415 9.742755 2.187985
DE 6.944247 3.56E-06 129.4541 28.87814 DE 501827.5 0.656741 3,880,011 975565.4
GWO 0.002007 0 0.028455 0.00675 GWO 0.030707 0.006851 0.065384 0.017626
WOA 0.007711 0 0.099239 0.024795 WOA 0.007714 0.001415 0.02106 0.006249
FTTA 0 0 0 0 FTTA 2.24E-10 4.22E-12 1.92E-09 4.34E-10

as Fik , i is its player number, Fi,j

k k
is its value on dimension j, Fi,j new is the Volatilities refuse to learn from anyone, and they do their own
state of the player on dimension j after training. training, so the state will fluctuate to a certain extent. Of course, with the
Discoverer increase of the number of training (Iterations). The fluctuation of the
Discoverers are more rational than the followers. They see not only player’s state well be getting smaller and smaller, we define the fluc­
the best players, but also the worst players, so they not only work to­ tuation of the player’s state as follows:
wards the best players, but also try their best to avoid becoming the k
Fi,j k
new = Fi,j old × (1 + t(k)) (4)
worst players. The equation is as follows:
k
Fi,j k
new = Fi,j old + rand1 × ( FkBest,j − Fi,j
k
old) − rand2 × ( FkWorst,j Among them, t(k) is a random number with t-distribution which it’s
degree of freedom is the current number of iterations, with the increase
− k
Fi,j old) (2)
of degree of freedom, the probability of t distribution close to the in­
termediate value (0) becomes higher and higher, and the distribution at
k
where the current worst player is defined as FWorst , k is the number of both ends gradually decreases, which will be closer and closer to the
k k
iterations, where FWorst,j is its value in dimension j. Fi,j new is the status normal distribution. Therefore, with the increase of the number of it­
of the player on dimension j after training. erations, the degree of fluctuation will become smaller and smaller, and
Thinkers gradually change from global search to local search.
Thinkers are more alert than those in front of them, they directly see The diagram of collective training is as follows (Fig. 1):
the gap between the best players and the worst players, and strive to
reach the gap in each dimension, and the equation is as follows: 2.1.2. Group training
After the completion of the collective training, the football training
k
Fi,j new = k
Fi,j old + rand ∗ ( FkBest,j − FkWorst,j ) (3)
process has reached the stage of group training, and the coaches divide
the players into four categories according to the characteristics of the
In dimension j, FkBest,j − FkWorst,j is the difference vector between the cur­ players (Each dimension is a characteristic value): Strikers, Midfielders,
rent best player and the worst player, k is the number of iterations. Defenders and Goalkeepers. In the group training, we use the MGEM
Volatilities adaptive clustering method (MixGaussEM), through the clustering

6
Z. Tian and M. Gai Expert Systems With Applications 245 (2024) 123088

Table 2B
The test results of CEC2005 (F13 - F23).
Mean Best Worst Std Mean Best Worst Std

CEC2005_F13 CEC2005_F14
PSO 16.362 6.495678 50.29283 9.898945 PSO 0.998004 0.998004 0.998004 9.76E-10
DE 8,431,744 0.081198 1.33E + 08 29,710,555 DE 2.304479 0.998004 9.745292 2.252207
GWO 0.408351 4.39E-05 0.761053 0.21891 GWO 3.004668 0.998004 12.67051 3.721667
WOA 0.171483 0.035033 0.53528 0.12567 WOA 2.272137 0.998004 10.76318 2.991735
FTTA 1.41E-05 3.21E-08 0.000218 4.84E-05 FTTA 0.998004 0.998004 0.998004 1.02E-16
CEC2005_F15 CEC2005_F16
PSO 0.006788 0.000837 0.022553 0.008718 PSO − 1.03163 − 1.03163 − 1.03122 8.98E-05
DE 0.00525 0.000307 0.020363 0.008117 DE − 1.03163 − 1.03163 − 1.03163 1.61E-16
GWO 0.003331 0.000307 0.020363 0.007341 GWO − 1.03163 − 1.03163 − 1.03163 9.79E-09
WOA 0.000636 0.000312 0.001531 0.000349 WOA − 1.03163 − 1.03163 − 1.03163 2.43E-10
FTTA 0.000392 0.000307 0.001223 0.00023 FTTA ¡1.03163 ¡1.03163 ¡1.03163 2.32E-19
CEC2005_F17 CEC2005_F18
PSO 0.397893 0.397887 0.397979 2.04E-05 PSO 3.000072 3 3.000293 8.77E-05
DE 0.397977 0.397887 0.399671 0.000399 DE 3 3 3 2.71E-15
GWO 0.397888 0.397887 0.397891 9.55E-07 GWO 3.000013 3 3.000066 1.63E-05
WOA 0.397889 0.397887 0.397896 2.59E-06 WOA 3.000001 3 3.000009 2.31E-06
FTTA 0.397888 0.397887 0.397974 2.73E-05 FTTA 3 3 3 9.22E-19
CEC2005_F19 CEC2005_F20
PSO − 3.86081 − 3.86278 − 3.8549 0.003499 PSO − 3.11638 − 3.32134 − 2.83997 0.163759
DE − 3.86278 − 3.86278 − 3.86278 2.22E-15 DE − 3.22115 − 3.322 − 2.86854 0.105872
GWO − 3.86211 − 3.86278 − 3.85925 0.001211 GWO − 3.23817 − 3.32199 − 3.11921 0.079516
WOA − 3.85911 − 3.86278 − 3.84942 0.003664 WOA − 3.19498 − 3.32176 − 3.07643 0.09337
FTTA ¡3.86215 ¡3.86278 ¡3.85702 0.001752 FTTA ¡3.2427 ¡3.32199 ¡3.11446 0.067355
CEC2005_F21 CEC2005_F22
PSO − 9.11121 − 10.152 − 2.63025 2.226692 PSO − 9.41285 − 10.4023 − 1.83753 2.410237
DE − 7.74278 − 10.1532 − 2.68286 3.077703 DE − 9.08676 − 10.4029 − 2.75193 2.727559
GWO − 8.63386 − 10.1526 − 5.05518 2.378653 GWO − 10.4016 − 10.4027 − 10.4001 0.000664
WOA − 8.1243 − 10.1528 − 2.62965 2.909317 WOA − 7.88055 − 10.4028 − 1.83751 3.560239
FTTA ¡9.88247 ¡10.1532 ¡5.0552 1.137951 FTTA ¡10.4029 ¡10.4029 ¡10.4027 5.29E-05
CEC2005_F23
PSO − 10.5145 − 10.5345 − 10.4495 0.022717
DE − 8.95519 − 10.5364 − 2.42173 3.246023
GWO − 10.1293 − 10.5359 − 2.42171 1.814171
WOA − 8.32419 − 10.5357 − 1.67648 3.185321
FTTA ¡10.5352 ¡10.5364 ¡10.5126 0.005333

method to simulate the behavior of coaches, and divide the population ⎧

into four categories according to their own characteristics. The specific ⎨ k,teaml
FBest,j if rand⩽pstudy
k,teaml
Fi,j new = (6)
forms of classification are as follows: k,teaml
⎩ Fi,j old if rand > pstudy
MGEM
All Players ̅̅̅̅→[Team 1, Team 2, Team 3, Team 4]
k,teaml
if Any team⩽Team number where FBest is the best player in group l, k is the number of iterations,
k,teaml
teaml representing group l, FBest,j is the j dimension of the best player in
Randomuniformgrouping
All Players ̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅→[Team 1, Team 2, Team 3, Team 4]
(5) k,teaml
group l, Fi,j new is the state of the player in j dimension after
Our strategy is to set a threshold: Team number (Teamnumber⩾2), optimal learning.
which is the minimum number of people in each group. We assume that Random learning
if the number of groups is less than this value, the group training can’t be In each dimension, the player has a certain probability to learn the
implemented. Once it happens, the coaches will carry out the second ability value of any random player in the group directly. The formula is
grouping, which adopts the strategy of uniform random grouping, which defined as follows:
divides the players into four groups randomly and evenly. ⎧
⎨ F k,teaml if rand⩽p
After the coaches complete the grouping, the players will learn or
(7)
k,teaml Random,j study
Fi,j new =
communicate with other players in the group. We define the group k,teaml
⎩ Fi,j old if rand > pstudy
training as three states: Optimal learning, Random learning and
Random communication. We define the learning probability as pstudy , k,teaml
where FRandom is a random player in group l, k is the number of iterations,
the communication probability is pcomm , the players will randomly select k,teaml
states in each iteration. teaml representing group l, FRandom,j is the j dimension of the random
k,teaml
Optimal learning player in group l, Fi,j new is the state of the player in j dimension
In each dimension, the player has a certain probability to learn the after random learning.
ability value of the best player in the group directly. The formula is Random communication
defined as follows: In training, learning is only a part of it, and the communication

7
Z. Tian and M. Gai Expert Systems With Applications 245 (2024) 123088

Table 2C
The test results of CEC2005 (F1 - F12).
Mean Best Worst Std Mean Best Worst Std

CEC2005_F1 CEC2005_F2
SGO 42.0836 25.96319 70.70697 14.15893 SGO 4.312717 3.028322 6.352676 1.179946
FGA 144.1049384 64.16416 268.982 6.42341 FGA 28.51981329 7.754778 66.9318 19.02627
VPL 1.56E-164 4.11E-182 1.56E-153 0 VPL 3.96E-103 8.72E-122 3.90E-102 1.23E-102
TGA 7.25E-37 1.79E-39 2.01E-36 7.61E-37 TGA 2.44E-19 5.17E-20 7.57E-19 2.25E-19
FTTA 2.0335E-184 1.0119E-276 2.0318E-182 0 FTTA 7.3433E-106 2.4081E-142 5.5175E-104 5.6902E-105
CEC2005_F3 CEC2005_F4
SGO 4.25E + 04 3.50E + 04 4.74E + 04 4.36E + 03 SGO 29.88088 21.27727 46.59181 7.961258
FGA 1.18E + 04 2.81E + 03 3.07E + 04 9.44E + 03 FGA 6.942059 4.376714 10.12716 1.860504
VPL 4.93E-139 2.61E-168 4.93E-118 0 VPL 3.72E-57 3.20E-102 3.71E-46 1.17E-26
TGA 9.77E-27 1.18E-29 2.80E-26 1.19E-26 TGA 1.98E-17 2.07E-18 5.00E-17 1.62E-17
FTTA 4.45E-166 3.05E-251 4.45E-164 0 FTTA 1.47E-94 1.43E-131 6.02E-93 8.68E-94
CEC2005_F5 CEC2005_F6
SGO 3.17E + 03 1.70E + 03 8.37E + 03 1.91E + 03 SGO 0.011433 0.004372 0.019229 0.005467
FGA 1.18E + 04 1.47E + 03 6.37E + 04 1.97E + 04 FGA 0.23889964 0.049438 0.771912 0.225492
VPL 25.70593 22.37538 27.73778 0.747014 VPL 0 0 0 0
TGA 25.66095 24.89185 26.19329 0.426503 TGA 1.55E-14 2.66E-15 3.34E-14 1.04E-14
FTTA 21.75252 20.52372 22.69543 0.642508 FTTA 0 0 0 0
CEC2005_F7 CEC2005_F8
SGO 0.410367 0.306808 0.514343 0.073241 SGO − 5257 − 5925.89 − 4359.79 472.8163
FGA 0.217419809 0.013778 0.608655 0.166222 FGA − 7914.122923 − 8862.53 − 6354.34 769.9387
VPL 0.002126 0.00095 0.002808 0.00668 VPL − 11815.36 − 12211.59 − 10908.1 374.2894
TGA 0.294203 0.17408 0.457292 0.089857 TGA − 4781.63 − 5909.32 − 4128.26 558.457
FTTA 0.00178 0.000536 0.003812 0.00083 FTTA ¡12415 ¡12569.4 ¡11616.5 236.8746
CEC2005_F9 CEC2005_F10
SGO 256.0329 212.1057 288.9829 24.23104 SGO 5.848077 3.764015 19.95769 4.982364
FGA 195.1280665 137.1355 250.4264 34.94543 FGA 13.4117826 4.104801 19.96677 7.654732
VPL 0 0 0 0 VPL 8.88E-16 8.88E-16 8.88E-16 0
TGA 0 0 0 0 TGA 4.44E-16 4.44E-16 4.44E-16 0
FTTA 0 0 0 0 FTTA 8.88E-16 8.88E-16 8.88E-16 0
CEC2005_F11 CEC2005_F12
SGO 1.369219 1.219506 1.467387 0.07677 SGO 16.67907 7.268321 32.32636 8.295491
FGA 2.404389711 1.861704 3.300536 0.455241 FGA 5.729093689 2.71816 8.803219 2.180432
VPL 0 0 0 0 VPL 1.25E-08 7.57E-11 1.25E-07 3.94E-08
TGA 0 0 0 0 TGA 0.001662 0.001186 0.002288 0.000329
FTTA 0 0 0 0 FTTA 2.24E-10 4.22E-12 1.92E-09 4.34E-10

between the two players is more important for ability improvement. In The diagram of group training is as follows (Fig. 2):
each dimension, players have a certain probability of communicating
with any player in the group. 2.1.3. Individual extra training
When rand⩽pcomm , the formula is defined as follows: After the end of the group training, it is necessary to recalculate the
⎡ ⎤ new fitness value and replace the poor fitness value with a better one to
F k,teaml new = FRandom,j update the status of the players. After the update, the coach will select
k,teaml
old ∗ (1 + randn)
⎣ i,j ⎦ if rand⩽pcomm (8)
k,teaml
FRandom,j k,teaml
new = Fi,j old ∗ (1 + randn) the best player and let him practice to make him better, so that he can
better drive the training state of others, the formula for training is as
k,teaml follows:
where FRandom is a random player in group l, k is the number of iterations,
k,teaml
teaml representing group l, FRandom,j is the j dimension of the random
k
FBest k
new = FBest old × (1 + (1 − 1/k) × Gauss + 1/k × Cauchy) (11)
player in group l, the two players exchanged their abilities in dimension Cauchy and Gauss joint variation is used to describe individual extra
j. Randn is a random number with normal distribution, and multiplying training and k is the number of iterations. The reason for the Gaussian-
with (1 + randn) represents the two players’ understanding of the Cauchy distribution is chosen is that the level of everyone is generally
abilities of others. not high in the early stage of training, so the best player has a greater
When rand > pcomm , the formula is defined as follows: probability to get a greater promotion, at this time, the Cauchy distri­
⎡
k,teaml k,teaml
⎤ bution function occupies a large proportion, which can effectively pro­
Fi,j new = Fi,j old vide a large range of improvement for players, which is conducive to the
⎣ ⎦ if rand > pcomm (9)
global search. With the increase of the number of iterations, the
k,teaml k,teaml
FRandom,j new = FRandom,j old
improvement of players’ ability becomes more and more difficult, at this
Random error time the Gaussian distribution accounts for a relatively large proportion,
We assume that during the group training, there is a certain proba­ and the promotion range of the players gradually decreases, which is
bility that errors will occur, that is, they accidentally learn the contents more conducive to local search.
of other dimensions of others. The probability of this situation is very The flow chart and the pseudo code of the FTTA algorithm are as
low, but it is real and objective. We define the error probability as perror . follows (Fig. 3):
⎧
⎨ Fk,teaml new = F k,teaml The pseudo code of the FTTA algorithm are as follows:
Random1 ,Random2 old if rand⩽perror
(10)
i,j
⎩ Fi,j k,teaml k,teaml
new = Fi,j old if rand > perror

8
Z. Tian and M. Gai Expert Systems With Applications 245 (2024) 123088

Pseudo code of the FTTA

CLC
CLEAR ALL
Set Objective Function
Parameters:
Set Maxgen = 500（The number of iterations is 500）
Set nPop = 50 （The population is 50）
pstudy = 0.2; pcomm = 0.2;perror = 0.001
/*Set the paraments of FTTA*/
/*Initialize population */
FOR I ¼ 1: MaxGen DO
FOR EACH i : 1⩽i⩽nPop DO
Calculate the fitness value of each football player Fi
End For
Find the Best & Find the Worst
Do Collective Training:
⎧
⎪
⎪ k
old + rand ∗ ( FkBest,j − Fki,j old)
⎪
⎪
⎪
⎪ Fi,j
⎪
⎪ k
⎪
⎨ F old + rand1 *( Fk k k k
k i,j Best,j − Fi,j old) − rand2 *( FWorst,j − Fi,j old)
Fi,j new = k k k
Divide the Population into Groups and Do Group Training:
⎪
⎪
⎪ Fi,j old + rand ∗ ( FBest,j − FWorst,j )
⎪
⎪
⎪ k
⎪
⎪
⎪ Fi,j old ∗ (1 + t(k))
⎩
⎧
⎪ k,teaml
⎪
⎪
⎪
⎪ FBest,j if rand⩽pstudy
⎨
k,teaml
Fi,j new = Fk,team
Random,j if rand⩽pstudy
l

⎪
⎪
⎪
⎪ Fk,teaml ∗(1 + randn) if rand⩽p
⎪
⎩ Random,j comm

Fk,team
i,j
l
new = Fk,team
Random1 ,Random2 if rand⩽perror
l

FOR EACH i : 1⩽i⩽nPop DO

Calculate the fitness value of each football player Fi

END FOR
Update the Players
Find the Best
Do Individual Extra Training
⎧
⎨ Fk old*(1 + (1 − 1)*Gauss + 1*Cauchy) if F(Fk new) < F(Fk old)
Best Best Best
k
FBest new = k k
⎩ k k k
FBest old if F(FBest new) > F(FBest old)
END FOR
/*Store the Optimal solution in Archive*/
t=t+1
END WHILE
END FOR

2.2. Proof of global convergence of FTTA 2. Proof of the ability of transfer from one local optimal state to
another better local optimal state: In order to proof the issue, we
For meta-heuristic optimization algorithm, to achieve global opti­ assume a nonlinear optimization problem: maxf(X), X = (x1 , x2 , ⋯,
mization function, it needs to meet two conditions at the same time xn ) ∈ K⊂Rn , where f is the fitness function. Suppose that there are m
(Pandey et al., 2014): First, it should have the ability to achieve local local maximum points with different values in this formula, and we
optimization. Second, it should be able to transfer from one local arrange them in the order from small to large, and get:
optimal state to another better local optimal state. In this part, I will f(X1 ) < f(X2 ) < ⋯f(Xi ) < ⋯f(Xj ) < ⋯f(Xm ), we assume that at the
prove why FTTA algorithm can effectively converge and achieve the current number of iterations, the local optimal state the algorithm
function of global optimization from these two aspects (Li et al., 2011). achieved is at position Xi , we define Xi+ as follows:
1. Proof of the ability to achieve local optimization: In the FTTA
optimization algorithm, we simulate the existence of a coach, whose role Xi+ = {X : f (X) > f (Xi )} (13)
is to observe the state of the players. At the end of an iteration, the fitness We set the number of states contained in Xi+ as Ω, so Ω is a monotone
function will be used to evaluate the status of the players to ensure that subtractive function of i, because the ergodic search of the solution space
better individuals enter the next round of competition. This practice can can be realized by a finite number of disturbance operations on any
be expressed by formula as follows: individual, so:
{
Fki+1 = Fki if Fit(Fki )⩽Fit(Fki− 1 ) Prob(Xi →Xi+ )⩾Prob(Xi →Xj ) > 0(j > i) (14)
(12)
Fi+1 = Fki− 1 if Fit(Fki ) > Fit(Fki− 1 )
k
In other words, when the current situation is a non-global maximum,
the FTTA algorithm can realize the directional transfer from the current
where i represents the number of iterations and k represents the number situation to the region with larger extreme points in a finite step.
of players, this formula can ensure that the whole population will only It can be seen from the formula that the one-way reachability of the
become stronger or at least remain the same in the iterative process, state (Xi →Xj ) is a sufficient condition for the implementation of the al­
which not only ensures the iterative stability of the algorithm, and gorithm, and it can be further inferred that the one-way reachability of
makes it have the function of realizing local optimization. the maximum state is also a sufficient condition for the implementation

9
Z. Tian and M. Gai Expert Systems With Applications 245 (2024) 123088

Table 2D
The test results of CEC2005 (F13 - F23).
Mean Best Worst Std Mean Best Worst Std

CEC2005_F13 CEC2005_F14
SGO 54.72347 13.59653 326.1205 95.95181 SGO 0.99833 0.998004 0.998874 0.000373
FGA 17.7442127 6.648108 47.09149 12.14839 FGA 0.998004 0.998004 0.998004 7.17E-10
VPL 8.20E-04 1.03E-9 4.45E-03 1.59E-04 VPL 0.998004 0.998004 0.998004 2.14E-14
TGA 1.865453 0.063419 2.966305 1.186859 TGA 0.998004 0.998004 0.998006 6.79E-07
FTTA 1.41E-05 3.21E-08 0.000218 4.84E-05 FTTA 0.998004 0.998004 0.998004 1.02E-16
CEC2005_F15 CEC2005_F16
SGO 0.000439 0.000323 0.000895 0.000165 SGO − 1.03159 − 1.03163 − 1.0315 4.30E-05
FGA 0.008115 0.001636 0.020363 0.008682 FGA − 1.03158 − 1.03163 − 1.03145 6.18E-05
VPL 0.000491 0.000307 0.001223 0.000386 VPL − 1.03163 − 1.03163 − 1.03162 6.82E-15
TGA 0.001082 0.00045 0.001737 0.000405 TGA − 1.03161 − 1.03162 − 1.03159 9.62E-06
FTTA 0.000392 0.000307 0.001223 0.00023 FTTA ¡1.03163 ¡1.03163 ¡1.03163 2.32E-19
CEC2005_F17 CEC2005_F18
SGO 0.398082 0.397887 0.398475 1.82E-04 SGO 3.000762 3.000095 3.00166 5.97E-04
FGA 0.397889 0.397887 0.397891 1.38E-04 FGA 3.000068 3 3.000337 0.000109
VPL 0.397882 0.397887 0.397867 3.71E-03 VPL 3 3 3 2.32E-15
TGA 0.398595 0.397897 0.401116 0.001039 TGA 3 3 3 1.12E-15
FTTA 0.397888 0.397887 0.397974 2.73E-05 FTTA 3 3 3 9.22E-19
CEC2005_F19 CEC2005_F20
SGO − 3.86278 − 3.86278 − 3.86278 9.36E-16 SGO − 3.18633 − 3.32199 − 3.0031 0.0874
FGA − 3.86041 − 3.86278 − 3.8549 0.003802 FGA − 3.165 − 3.32191 − 2.8617 0.156399
VPL − 3.86164 − 3.86258 − 3.862122 0.039824 VPL − 3.226881 − 3.321995 − 3.103102 0.08234
TGA ¡3.86278 ¡3.86278 ¡3.86278 7.83E-16 TGA − 3.24116 − 3.32199 − 3.2031 0.077084
FTTA − 3.86215 − 3.86278 − 3.85702 0.001752 FTTA ¡3.2427 ¡3.32199 ¡3.11446 0.067355
CEC2005_F21 CEC2005_F22
SGO − 6.82334 − 10.1532 − 3.22916 2.63 SGO − 7.17133 − 10.4029 − 2.7573 3.45
FGA − 9.09713 − 10.1508 − 5.0547 2.13069 FGA − 10.1604 − 10.3998 − 8.21562 0.683966
VPL − 7.629054 − 10.1532 − 2.630472 3.3602 VPL − 7.48747 − 10.4029 − 1.837593 3.85
TGA − 5.22898 − 6.79306 − 5.0552 0.549561 TGA − 5.35116 − 7.73855 − 5.07168 0.838858
FTTA ¡9.88247 ¡10.1532 ¡5.0552 1.137951 FTTA ¡10.4029 ¡10.4029 ¡10.4027 5.29E-05
CEC2005_F23
SGO − 8.43141 − 10.5364 − 3.80988 2.51
FGA − 10.5125 − 10.5349 − 10.4748 0.023291
VPL − 9.18471 − 10.5364 − 2.4273 2.9221
TGA − 5.26839 − 7.02301 − 4.83578 0.623624
FTTA ¡10.5352 ¡10.5364 ¡10.5126 0.005333

Fig. 4. Theoretical image of CEC function and performance of FTTA.

10
Z. Tian and M. Gai Expert Systems With Applications 245 (2024) 123088

Fig. 5. Theoretical image of CEC function and performance of FTTA.

CEC2005 test function is shown in Table 1, and the test results are shown
Table 3 in Table 2ATable 2BTable 2CTable 2DA and Fig. 4 and Fig. 5.
The function expression of CEC2005.
From the results, it is obvious that our optimization algorithm has
Dim Range Fmin achieved the optimal solution in almost all the test functions.
CEC2020_F1 5 [-100,100] 100
CEC2020_F2 5 [-100,100] 1100
2.4. FTTA performance testing based on CEC2020
CEC2020_F3 5 [-100,100] 700
CEC2020_F4 5 [-100,100] 1900
CEC2020_F5 5 [-100,100] 1700 The function expression of CEC2020 is shown in Table 3. We tested
CEC2020_F6 5 [-100,100] 1600 the performance of PSO, DE, GWO, WOA and FTTA on the test function,
CEC2020_F7 10 [-100,100] 2100 and the test results are shown in Table 4. From the results, we can clearly
CEC2020_F8 5 [-100,100] 2200
see that our optimization algorithm has achieved the optimal solution in
CEC2020_F9 5 [-100,100] 2300
CEC2020_F10 5 [-100,100] 2500 almost all the test functions.

2.5. FTTA performance testing based on CPU operation time

of the algorithm. Therefore, the conditions for FTTA to achieve global
optimization are very loose.
The CPU running speed of FTTA is tested in this part. In practical
applications, we often expect to get the best results in the shortest
2.3. FTTA performance testing based on CEC2005 possible time. Therefore, the running speed of the algorithm is very
important. In this discussion, we unify the running environment for all
The performance of the FTTA algorithm is verified through a set of algorithms: Matlab2020a, as well as the configuration of the computer:
traditional test functions CEC2005. We repeat the algorithm 50 times, and Intel 10-generation CPU to verify the running rate of the algorithm. All
then make a performance evaluation table of the meta-heuristic optimi­ the settings of the algorithm are consistent with those in 2.1, and the
zation algorithm according to the Max, Min, Std and Mean of the result. results are shown in Table 5.
The optimization algorithms involved in the comparison include GWO and From the above table, it is obvious that the performance of other
WOA, which is widely used at present, and the early algorithm PSO algorithms is not very different except GOA algorithm is very slow. The
(Anggraeni et al., 2024), DE (Nandhini & Balasundaram, 2014). We also running speed of FTTA algorithm is basically the same as that of most
compare the proposed FTTA algorithm with some current algorithms algorithms with fast running speed, but its accuracy is significantly
inspired by sports, including SGO, FGA, VPL and TGA. Iterations is 500 higher, and it can search out global optimal solutions more easily, which
and population is 50. The other super parameters follow the initial settings further proves the progressiveness and practical application ability of
when the algorithm is released. The mathematical expression of the FTTA algorithm.

11
Z. Tian and M. Gai Expert Systems With Applications 245 (2024) 123088

Table 4 Table 5
The test results of CEC2020. CPU operation time of each model.
Mean Best Worst Median Std Model CPU operation time（
（second）
）
CEC2020_F1 PSO 2.501
PSO 33548.66 3339.246 119924.8 27604.19 30135.71 DE 2.602
DE 174.6295 100.0094 507.5319 110.2058 124.625 GWO 2.401
GWO 5934.89 589.3267 13600.32 4110.405 4132.677 GOA 170.634
WOA 8606.279 239.6059 36265.04 7434.096 8088.376 WOA 2.417
FTTA 148.2527 100.0023 540.6072 107.6469 100.0294 FTTA 2.087
CEC2020_F2
PSO 1269.405 1111.019 1643.674 1235.321 125.4442
DE 1426.986 1135.484 1911.994 1392.103 214.901
composed of variational mode decomposition (VMD) and FTTA. FTTA
GWO 1283.49 1100.3153 1515.859 1245.21 111.6266
WOA 1418.036 1119.258 1870.849 1388.646 186.2054 is used to optimize the punishment factor (α) and decomposition layers
FTTA 1231.927 1100.25 1399.046 1219.468 97.55719 (k) of VMD to achieve the best data preprocessing result. We regard the
CEC2020_F3 most fluctuating component of the decomposed IMF component as
PSO 713.0736 702.7388 720.0333 713.0864 4.027884
white noise and delete it, and recombine other IMF components to get
DE 714.9471 706.608 723.7202 715.6559 5.654764
GWO 709.8157 705.3118 713.9562 709.4445 2.375721
the time series data after denoising.
WOA 717.8281 705.2546 728.9874 716.4592 6.312567 Objective Function Design 1: FTTA-VMD model is an advanced
FTTA 708.4315 701.9748 714.4212 708.2137 2.433894 data preprocessing model. In the model, the fitness function of FTTA is
CEC2020_F4 the minimum of envelope entropy, and envelope entropy represents the
PSO 1900.934 1900.291 1903.647 1900.739 0.705121
sparse characteristic of the original signal. When there is more noise and
DE 1900.647 1900.124 1901.565 1900.585 0.341603
GWO 1900.726 1900.114 1901.726 1900.428 0.171537 less feature information in IMF, the envelope entropy is larger, on the
WOA 1902.678 1900.446 1915.782 1901.705 3.273529 contrary, the envelope entropy is smaller.
FTTA 1900.474 1900.05 1901.27 1900.421 0.293041 In signal x(i), the envelope entropy Ep is calculated using the
CEC2020_F5
following formula:
PSO 2806.949 1747.231 10625.41 1915.547 2216.394
DE 1737.387 1700.995 1841.531 1727.878 45.47495 ∑
N ∑
N
GWO 2409.921 1720.382 8519.266 1954.612 1473.432 Ep = − pj lgpj pj = a(j)/ a(j) (12)
WOA 5043.637 1785.08 13183.29 2871.227 3884.722 j=1 j=1
FTTA 1706.419 1700 1719.899 1701.99 7.012298
CEC2020_F6 where a(j) is the envelope signal of k modal components decomposed by
PSO 1602.574 1600.559 1617.737 1601.329 4.170798
DE 1617.994 1600 1761.884 1600.396 43.09678
VMD and demodulated by Hilbert. pj is the probability distribution
GWO 1611.422 1601.004 1718.945 1602.964 26.15279 sequence obtained by calculating the normalization of a(j). N is the
WOA 1615.677 1600.908 1671.877 1616.567 17.5112 number of sampling points, and the entropy of the probability distri­
FTTA 1600.144 1600 1600.624 1600.005 0.190484 bution sequence is the envelope entropy Ep . After the optimization of
CEC2020_F7
FTTA, the optimal hyperparameters [α, k] are obtained, and the best data
PSO 363,554 2801.04 3,541,247 7608.164 1,086,759
DE 3861.694 2107.461 31752.88 2301.184 6577.971 preprocessing result is achieved.
GWO 8411.091 2489.311 20719.03 6664.766 5800.069 Part 2: The second part is the combination prediction model, which
WOA 193743.7 7550.098 1,499,453 37249.64 370917.9 is mainly composed of four prediction methods of ARIMA (Ray et al.,
FTTA 3368.508 2102.144 16286.87 2286.778 3261.626
2023), GRU (Huang & Qian, 2023), CNN (Sharifuzzaman Sagar et al.,
CEC2020_F8
PSO 2212.563 2200.062 2227.259 2212.509 6.078788
2024), BP (Du et al., 2024) and unconstrained weighting algorithm
DE 2220.807 2200 2302.907 2206.612 35.42774 based on FTTA. The principle of each prediction method is different
GWO 2256.201 2200.043 2302.98 2258.735 47.05896 (three neural networks with different principles and one statistical
WOA 2248.443 2200.1 2316.829 2219.888 46.32988 method). In different data sets, some methods have better prediction
FTTA 2211.749 2200 2284.693 2218.432 46.83901
results and some have poor results. Therefore, in order to ensure the
CEC2020_F9
PSO 2616.072 2403.163 2748.943 2604.954 113.5551 accuracy and stability of prediction, FTTA-CP (Combined Prediction)
DE 2580.058 2500 2727.06 2500 108.5278 model is proposed. After obtain the prediction results, the prediction
GWO 2538.06 2500.035 2729.755 2500.5 80.57918 performance of each model is judged by statistical indicators, and un­
WOA 2601.88 2500.38 2745.521 2528.027 109.8927
constrained weighting is carried out through FTTA, which gives less
FTTA 2533.509 2400 2726.159 2500 87.9262
CEC2020_F10
weight to the models with poor accuracy and larger weights to the
PSO 2835.005 2800.512 2891.944 2847.425 23.3725 models with better accuracy to achieve the optimization of prediction
DE 2847.381 2847.367 2847.401 2847.382 0.011228 accuracy.
GWO 2847.407 2847.381 2847.433 2847.395 0.016992 Objective Function Design 2: As a combination prediction model,
WOA 2851.779 2847.391 2891.807 2847.489 13.17999
the objective function of FTTA-CP model is to maximize the prediction
FTTA 2847.305 2847.367 2847.535 2847.391 0.045012
accuracy (MAPE value), and the final result is the optimal weight of
each prediction model [x(1), x(2), x(3), x(4)].
From the experimental results, no matter the basic CEC2005 or the
MAPE = mean(abs(Yt - actual)./actual)*100 (13)
latest test function CEC2020, FTTA both shows the best performance.
Our algorithm converges fast, the computing time is short, the optimi­
Yt = x(1)∗ARIMA + x(2)∗CNN + x(3)∗GRU + x(4)∗BP
zation effect is good, and the results are concentrated.
In the traditional constrained weighting model, the weights of each
3. Design of a hybrid wind speed prediction system model are between 0 and 1, and the sum of all weights is required to be
1, but in FTTA-CP, unconstrained weighting method is used, and we set
The new proposed wind speed prediction system is mainly composed the upper and lower limits of-5 and 5 respectively. There are no re­
of two parts: data-preprocessing and combination prediction (CP) based quirements for the sum of weights or the positive and negative of
on FTTA. weights, so better results can be obtained (Tian & Wang, 2022).
Part 1: The first part is data preprocessing, which is mainly Proof

12
Z. Tian and M. Gai Expert Systems With Applications 245 (2024) 123088

Fig. 6. Full flow chart of the wind speed prediction system.

Table 6 Table 7b
Basic characteristics of the data. 2 Parameters of Part 2: FTTA-CP.
Max Min Mean Var Std Model Parameters Value

Spring 18.2 0.9 7.446313 9.295596 3.048868 GRU Max Epochs 500
Summer 11.6 0.5 4.235813 3.62674 1.9044 Threshold 1
Autumn 10.7 0.7 4.613875 3.218056 1.793894 Learning Rate 0.005
Winter 17.2 1.7 8.206063 7.777218 2.788766 Learn Rate Drop When 250
Learn Rate is Dropped To 0.001
CNN Max Epochs 50
Mini Batch Size 16
Table 7a Learn Rate 0.001
1 Parameters of Part 1: FTTA-VMD. Threshold 1
BP Max Epochs 1000
Model Parameters Value
Goal 0.001
VMD Signal Data Learning Rate 0.01
Alpha Based on FTTA [1000,2500] ARIMA p [1,5]
Tau 0 d [0,3]
K Based on FTTA [4,15] q [1,5]
DC 0 FTTA Population 50
Init 1 Max-Iteration 100
Tol 1e-6 Communication Operator 0.2
FTTA Population 20 Study Operator 0.2
Max-Iteration 15 Error Operator 0.001
Communication Operator 0.2 Team Number 5
Study Operator 0.2
Error Operator 0.001
Team Number 2 CNN, GRU and BP method.
The flow chart of the wind speed prediction system is as follows
(Fig. 6):
Meta-heuristic algorithm is a method based on the mechanism of
computational intelligence to solve the optimal or satisfactory solution
of complex optimization problems. In the iterative process of FTTA, we 4. Selection of data and model parameters
assume that the current number of iterations is N, when there is a M* =
N* + 1, B(N* ) is the best individual of the population in iteration N, Dalian, located in the south of Liaoning Province in China, in the
Bohai economic belt, is a beautiful coastal city. Dalian is not only an
satisfies FITness(B(M* ))⩾FITness(B(N* )), then B(M* ) will be sorted in
outstanding tourist city in China, but also an important industrial city.
front of B(N* ). After the completion of the iteration, the final gradient
Dalian has well-developed wind energy resources. There are large-scale
descent curve is obtained. We believe that in Part 2, after enough iter­
wind power stations in Zhuang-He and other places. In order to test the
ations, FTTA will obtain the theoretical optimal weight for ARIMA,
effect of our model in practical application throughout the year, we

13
Z. Tian and M. Gai Expert Systems With Applications 245 (2024) 123088

Table 8 selected a wind speed measuring machine and obtained 1600 sets of
Mathematical expression of the Evaluation Index. data in Spring, Summer, Autumn and Winter, in which the first 1500
Index Mathematical expression sets of data were used as the training set. the last 100 sets of data are
used as the testing set. The frequency of wind speed measurement is
MAPE 1∑n ⃒⃒ ∧ ⃒
every 10 min. The basic characteristics of the data are shown in Table 6:
⃒
MAPE = ⃒(wm i − wmi )/wm i⃒*100
n i=1
MAE ∑
MAE = ni=1 |wmi − wmi |*
∧ 1 In order to effectively evaluate the point prediction results of the
n model, we use multiple indicators to comprehensively evaluate, the
SSE ∑n ( ∧ )2
SSE = i=1 wm i − wmi selected indexes are: MAPE, MAE, SSE, MSE, RMSE, NMSE, R2, IA,
MSE MSE = SSE/n
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
U1, and the formula of the index is shown in Table 8. In the model, we
RMSE )2 ̅
RMSE =
∑n ( ∧ use the combination of six algorithms including: FTTA, ARIMA (Luo &
i=1 wm i − wmi /n
Gong, 2023), GRU (Wu et al., 2023), CNN (Zhang & Yin, 2024), BP
NMSE 1∑n
NMSE = (wm i − wmi )2 /var(wmi ) (Chen et al., 2022), VMD (Parri et al., 2024), and the super parameters
n i=1
R2
R2 = 1 −
∑n ( ∧ )2 ∑ (
/ ni=1 wm i − wmi
)2
of each model are shown in Table 7a.Table 7b.
i=1 wm i − wmi

IA ∑n ( ∧ )2 1
i=1 wm i − wmi * 5. Experiments
IA = 1 − n
∑n ( ⃒ ∧ ⃒ )2
⃒ ⃒
i=1 |w mi − w mi | − ⃒wm i − wm i⃒
U1
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
( )2 5.1. Experiment 1
∑n ⌢
i=1 wmi − wmi /n
U1 = √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅ √ ̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
( )2 ̅ In experiment 1, we mainly explore the improvement of the pre­
∑n ( )/
diction accuracy of our combined wind speed prediction system
2 ∑n ⌢
i=1 (w mi ) n + i=1 w mi /n
compared with the single model in the system, to verify the necessity of
the combined prediction system. The results are shown in Table 9. It is

Table 9
The results of experiment 1.
MAPE MAE SSE MSE RMSE NMSE R2 IA U1
Spring FTTA-VMD-ARIMA 6.2013 0.7661 95.7281 0.9573 0.9784 0.1334 0.8652 0.9989 0.0379

FTTA-VMD-CNN 6.6771 0.8341 113.8568 1.1386 1.0671 0.1587 0.8397 0.9988 0.0419
FTTA-VMD-GRU 4.4287 0.5717 61.8466 0.6185 0.7864 0.0862 0.9129 0.9992 0.0308
FTTA-VMD-BP 4.5507 0.6263 65.3194 0.6532 0.8082 0.091 0.908 0.9992 0.0322
FTTA-VMD-CP 3.5479 0.4869 43.5819 0.4358 0.6602 0.0607 0.9386 0.9994 0.0261
Summer FTTA-VMD-ARIMA 12.8884 0.4854 40.6579 0.4066 0.6377 0.0635 0.9358 0.9995 0.0566
FTTA-VMD-CNN 14.8786 0.533 39.1447 0.3914 0.6256 0.0612 0.9382 0.9995 0.0558
FTTA-VMD-GRU 10.708 0.4137 26.4017 0.264 0.5138 0.0413 0.9583 0.9997 0.0464
FTTA-VMD-BP 4.5534 0.1539 3.8483 0.0385 0.1962 0.006 0.9939 0.9999 0.0176
FTTA-VMD-CP 3.5781 0.1438 3.1896 0.0319 0.1786 0.005 0.995 0.9999 0.0161
Autumn FTTA-VMD-ARIMA 14.3204 0.4061 25.0602 0.2506 0.5006 0.4824 0.5127 0.9975 0.0831
FTTA-VMD-CNN 11.0122 0.308 14.935 0.1493 0.3864 0.2875 0.7096 0.9982 0.064
FTTA-VMD-GRU 9.8076 0.2735 11.8343 0.1183 0.3439 0.2278 0.7699 0.9985 0.0571
FTTA-VMD-BP 4.2502 0.1089 1.8079 0.0181 0.1345 0.0348 0.9648 0.9996 0.0224
FTTA-VMD-CP 3.8605 0.0956 1.4687 0.0147 0.1212 0.0283 0.9714 0.9997 0.0202
Winter FTTA-VMD-ARIMA 9.3175 0.6483 69.2335 0.6923 0.832 0.1472 0.8513 0.9988 0.0532
FTTA-VMD-CNN 10.7424 0.7114 73.9209 0.7392 0.8598 0.1572 0.8412 0.9988 0.0541
FTTA-VMD-GRU 9.7236 0.6767 66.479 0.6648 0.8154 0.1414 0.8572 0.9989 0.052
FTTA-VMD-BP 4.4486 0.3177 16.0668 0.1607 0.4009 0.0342 0.9655 0.9996 0.0258
FTTA-VMD-CP 3.6542 0.2655 12.1011 0.121 0.3479 0.0257 0.974 0.9997 0.0224

Table 10
The results of experiment 2.
MAPE MAE SSE MSE RMSE NMSE R2 IA U1
Spring PSO-VMD-CP 3.6103 0.4892 43.8875 0.4389 0.6625 0.0612 0.9382 0.9994 0.0262

DE-VMD-CP 3.5622 0.4889 43.8837 0.4388 0.6624 0.0612 0.9382 0.9994 0.0262
GWO-VMD-CP 3.6044 0.4952 45.0835 0.4508 0.6714 0.0628 0.9365 0.9994 0.0266
WOA-VMD-CP 3.5734 0.4908 44.9939 0.4499 0.6707 0.0627 0.9367 0.9994 0.0265
FTTA-VMD-CP 3.5479 0.4869 43.5819 0.4358 0.6602 0.0607 0.9386 0.9994 0.0261
Summer PSO-VMD-CP 3.6827 0.1475 3.3867 0.0339 0.1841 0.0053 0.9947 0.9999 0.0166
DE-VMD-CP 3.5806 0.1442 3.2074 0.0321 0.1792 0.005 0.9949 0.9999 0.0162
GWO-VMD-CP 3.9062 0.1564 3.6738 0.0367 0.1916 0.0057 0.9942 0.9999 0.0173
WOA-VMD-CP 6.8979 0.3014 14.7829 0.1478 0.3844 0.0231 0.9767 0.9998 0.0351
FTTA-VMD-CP 3.5781 0.1438 3.1896 0.0319 0.1786 0.005 0.995 0.9999 0.0161
Autumn PSO-VMD-CP 4.5821 0.1198 2.1216 0.0212 0.1456 0.0408 0.9587 0.9996 0.0242
DE-VMD-CP 4.1898 0.1051 1.6516 0.0165 0.1285 0.0318 0.9679 0.9997 0.0214
GWO-VMD-CP 4.212 0.1061 1.6552 0.0166 0.1288 0.0319 0.9678 0.9997 0.0214
WOA-VMD-CP 4.9811 0.1379 3.4831 0.0348 0.1865 0.067 0.9323 0.9994 0.031
FTTA-VMD-CP 3.8605 0.0956 1.4687 0.0147 0.1212 0.0283 0.9714 0.9997 0.0202
Winter PSO-VMD-CP 3.9729 0.2921 14.3292 0.1433 0.3785 0.0305 0.9692 0.9997 0.0245
DE-VMD-CP 3.9536 0.2903 14.1074 0.1411 0.3756 0.03 0.9697 0.9997 0.0243
GWO-VMD-CP 3.9557 0.2914 14.0372 0.1404 0.3747 0.0298 0.9699 0.9997 0.0242
WOA-VMD-CP 4.3732 0.3208 17.0484 0.1705 0.4129 0.0362 0.9634 0.9996 0.0267
FTTA-VMD-CP 3.6542 0.2655 12.1011 0.121 0.3479 0.0257 0.974 0.9997 0.0224

14
Z. Tian and M. Gai Expert Systems With Applications 245 (2024) 123088

Fig. 7. Point prediction results of Experiment 1 and Experiment 2.

50 % higher; The prediction accuracy of the single CNN model is MAPE

Table 12 (6.7), MAE (0.8). the prediction accuracy of our proposed combination
CPU operation time of each model. prediction system is also nearly 50 % higher than that of the CNN.
Model CPU operation time（
（second）
） From the Summer data, the MAPE value of the prediction system is
VMD 4.1 about 3.6 and the MAE value is about 0.14, which is the best of all the
FTTA-VMD 150.8 models. The MAPE value and MAE value of the GRU are 10.7 and 0.4.
BP 4.7 Compared with the GRU, the prediction accuracy of the proposed
CNN 71.3
combination prediction system is nearly 70 % higher than that of the
GRU 46.2
ARIMA 25.8 GRU model; The prediction accuracy of the BP Neural Network is MAPE
FTTA-CP 75.4 (4.55), MAE (0.15). the prediction accuracy of our proposed prediction
FTTA-VMD-CP 226.2 system is about 20 % higher than that of BP neural network.
Whether in Autumn data or Winter data, the proposed prediction
system shows the best performance in all evaluation indicators. For data
obvious that our proposed combined wind speed prediction system has
with large fluctuations in spring or winter, the MAE will be relatively
achieved the best performance in both spring, summer, autumn and
large, about 0.4. For data with small fluctuations in spring or autumn,
winter, which further verifies that our prediction system is available for
the MAE will be relatively small, about 0.1. For any data set, the MAPE
the whole year of wind power plants, which is a system of practical
value is below 4, indicating that our proposed wind speed prediction
significance.
system has achieved excellent performance and stability.
From the Spring data, the MAPE value of the prediction system is
about 3.55 and the MAE value is about 0.5, both of which have the best Remark 1. In experiment 1, the advanced nature of the combined pre­
performance. The MAPE value and MAE value of the single ARIMA diction system is verified compared with the single prediction model. The
model are 6.2 and 0.8. Compared with the ARIMA model, the predic­ experimental results show that compared with the single neural network or
tion accuracy of the proposed combination prediction system is nearly statistical prediction method, the wind speed prediction system shows the

15
Z. Tian and M. Gai Expert Systems With Applications 245 (2024) 123088

Table 11
The results of experiment 3.
MAPE MAE SSE MSE RMSE NMSE R2 IA U1
Spring EMD-CP 3.9271 0.5046 46.4746 0.4647 0.6817 0.0648 0.9346 0.9994 0.0267

EEMD-CP 3.7952 0.5107 49.0707 0.4907 0.7005 0.0684 0.9309 0.9993 0.0276
CEEMDAN-CP 3.8663 0.5174 46.1289 0.4613 0.6792 0.0643 0.9351 0.9994 0.0269
ICEEMDAN-CP 3.7023 0.4989 45.0361 0.4504 0.6711 0.0628 0.9366 0.9994 0.0265
FTTA-VMD-CP 3.5479 0.4869 43.5819 0.4358 0.6602 0.0607 0.9386 0.9994 0.0261
Summer EMD-CP 3.7073 0.1494 3.3683 0.0337 0.1836 0.0053 0.9947 0.9999 0.0166
EEMD-CP 4.0728 0.1661 4.0697 0.0407 0.2017 0.0064 0.9936 0.9999 0.0182
CEEMDAN-CP 3.8549 0.1504 3.4857 0.0349 0.1868 0.0054 0.9945 0.9999 0.0168
ICEEMDAN-CP 3.6221 0.1502 3.5412 0.0354 0.1881 0.0055 0.9944 0.9999 0.017
FTTA-VMD-CP 3.5781 0.1438 3.1896 0.0319 0.1786 0.005 0.995 0.9999 0.0161
Autumn EMD-CP 4.6199 0.1167 2.0609 0.0206 0.1435 0.0397 0.9599 0.9996 0.0238
EEMD-CP 4.3358 0.1089 1.7794 0.0178 0.1334 0.0343 0.9654 0.9996 0.0222
CEEMDAN-CP 4.6646 0.1239 2.3079 0.0231 0.152 0.0444 0.9551 0.9996 0.0252
ICEEMDAN-CP 4.13 0.1094 1.8046 0.018 0.1342 0.0347 0.9649 0.9996 0.0223
FTTA-VMD-CP 3.8605 0.0956 1.4687 0.0147 0.1212 0.0283 0.9714 0.9997 0.0202
Winter EMD-CP 4.1124 0.3016 15.158 0.1516 0.3894 0.0322 0.9674 0.9996 0.0251
EEMD-CP 4.4033 0.3237 16.7152 0.1672 0.4089 0.0355 0.9641 0.9996 0.0264
CEEMDAN-CP 4.1119 0.3004 15.4781 0.1548 0.3934 0.0329 0.9668 0.9996 0.0254
ICEEMDAN-CP 3.7193 0.2687 12.8817 0.1288 0.3589 0.0274 0.9723 0.9997 0.0232
FTTA-VMD-CP 3.6542 0.2655 12.1011 0.121 0.3479 0.0257 0.974 0.9997 0.0224

highest accuracy and stability, which verified that the combined wind speed more suitable for the wind speed prediction system. It not only shows the
prediction system we developed is necessary. extremely strong accuracy and stability of the model, but also verifies the
practical application ability of the system, because it performs best in the data
of four seasons.
5.2. Experiment 2
5.3. Experiment 3
In Experiment 2, we mainly verified the applicability of FTTA
method in the Part 2 of wind speed prediction system. As we all know, In Experiment 3, the advanced nature of our data preprocessing
there are many optimization algorithms being developed and applied at module is verified. In the data preprocessing module, we use FTTA to
present. Therefore, in this experiment, we selected some classic meta optimize the two super-parameters of VMD. Therefore, in this experi­
heuristic optimization algorithms to compare with FTTA to verify the ment, we choose other data preprocessing methods to compare with
progressiveness of the prediction system we proposed. GWO and WOA FTTA-VMD to verify the advanced nature of our proposed prediction
meta-heuristic optimization algorithms, which are the method of simu­ system. The method we chose includes EMD (Xiong et al., 2024), EEMD
lating animal behavior widely and used at present, are chosen to (Li et al., 2023), CEEMDAN (Karijadi et al., 2023), ICEEMDAN (Bom­
compare with FTTA. We also bring the classic PSO and DE algorithm midi et al., 2023), and the experimental results are shown in Table 11.
into our system and compare them with FTTA to verify the advanced Obviously, the FTTA-VMD method is most suitable for our wind
role of FTTA in the wind speed prediction system. The experimental speed prediction system in the year-round data set. In the autumn and
results are shown in Table 10. In this experiment, the population number winter data, changing the data preprocessing method will lead to a
and maximum number of iterations of all optimization algorithms are significant decline in prediction accuracy. In the spring and summer
the same, and the remaining parameters follow the settings when the data, other data preprocessing methods have little difference with the
algorithm is released. FTTA-VMD method, but there are still some gaps. In general, the data
It is clear from Table 10 that FTTA has the best performance of all preprocessing method FTTA-VMD shows a strong fitness, and also re­
meta-heuristic algorithms in the year-round dataset. For Spring data, flects the progressiveness and applicability of the proposed prediction
FTTA is about 0.1 higher than PSO, DE, GWO, WOA according to system.
MAPE. For Summer data, FTTA is about 0.1 higher than PSO, DE and
0.3 higher than GWO, while WOA’s results on summer datasets are not Remark 3. In experiment 3, the applicability of FTTA-VMD algorithm is
satisfactory. For Autumn data, FTTA is about 0. 7 higher compared to verified. Compared with the traditional data preprocessing method, it is
PSO, 0. 3 to DE, 0. 4 to GWO, and 1. 1 to WOA according to MAPE. For obvious that the use of FTTA-VMD method makes the wind speed prediction
Winter data, FTTA is about 0. 3 higher than PSO, DE, GWO and 0. 7 system obtain higher accuracy and stability, and enhances the practical
higher than WOA in MAPE value. application ability of the system.
Further analysis of the experimental results shows that PSO opti­
mization algorithm performs better in spring and summer data, but 5.4. Experiment 4
worse in autumn and winter data; DE optimization algorithm performs
better in spring, summer and autumn data, but worse in winter data; In this experiment, we mainly analyze the running speed of the
GWO only performed well in spring data, while WOA did not perform combined prediction system to verify the applicability of the prediction
satisfactorily in all data sets. However, FTTA has shown strong appli­ system in the wind speed prediction. The results are shown in Table 12.
cability. In the four data sets we tested, it has shown the best perfor­ From Table 12, it is obvious that the running time of the whole
mance and the best stability. It is very suitable for our proposed wind prediction system is about 230 s, that is, about 4 min. In actual wind
speed prediction system, further verifying the progressiveness and ac­ power plant applications, the interval between predictions is 10 min.
curacy of the prediction system. According to Experiment 1 and Therefore, sufficient time can be reserved to allow the staff of the wind
Experiment 2, we draw the results of the experiment as shown in Fig. 7. power plant to adjust the facilities accordingly, so as to achieve the
optimal working effect, which further verifies the practicality and pro­
Remark 2. In experiment 2, The applicability of FTTA-CP algorithm in the
gressiveness of the proposed system.
combined wind speed prediction system is verified. Compared with the
traditional meta-heuristic optimization algorithm, it is obvious that FTTA is Remark 4. In experiment 4, the CPU operation time of hybrid wind speed

16
Z. Tian and M. Gai Expert Systems With Applications 245 (2024) 123088

prediction system is verified. Obviously, the operation speed of the wind speed Appendix A. Supplementary data
prediction system based on FTTA is satisfied, which further confirms its
practical application ability. Supplementary data to this article can be found online at https://doi.
org/10.1016/j.eswa.2023.123088.
6. Conclusion
References
In this paper, a new meta-heuristic optimization algorithm-FTTA
algorithm is proposed, which simulates the training process of football Ali, E. E., El-Hameed, M. A., El-Fergany, A. A., & El-Arini, M. M. (2016). Parameter
extraction of photovoltaic generating units using multi-verse optimizer. Sustainable
team. Through three processes: collective training, group training and Energy Technologies and Assessments, 17, 68–76. https://doi.org/10.1016/j.
individual extra training, the level of the whole team is improved finally. seta.2016.08.004
We further verify the advanced nature of the FTTA algorithm through Anggraeni, W., Yuniarno, E. M., Rachmadi, R. F., Sumpeno, S., Pujiadi, P., Sugiyanto, S.,
… Purnomo, M. H. (2024). A hybrid EMD-GRNN-PSO in intermittent time-series
the standard CEC test function and the model CPU operation time. data for dengue fever forecasting. Expert Systems with Applications, 237, Article
Obviously, the optimization effect of the proposed FTTA optimization 121438. https://doi.org/10.1016/j.eswa.2023.121438
algorithm is very good, which far exceeds the traditional meta-heuristic Bommidi, B., Teeparthi, K., & Kosana, V. (2023). Hybrid wind speed forecasting using
ICEEMDAN and transformer model with novel loss function. Energy, 265, Article
optimization algorithm. 126383. https://doi.org/10.1016/j.energy.2022.126383
In the engineering application, a new hybrid wind speed prediction Chen, C.-L., Vempati, V. S., & Aljaber, N. (1995). An application of genetic algorithms for
system is proposed based on FTTA. Firstly, we use the FTTA algorithm flow shop problems. European Journal of Operational Research, 80(2), 389–396.
https://doi.org/10.1016/0377-2217(93)e0228-p
combined with the VMD method to form a new data preprocessing
Chen, G., Tang, B., Zeng, X., Zhou, P., Kang, P., & Long, H. (2022). Short-term wind
model, combined with the combination prediction method to form a speed forecasting based on long short-term memory and improved BP Neural
new prediction model. Through four groups of experiments, the accu­ Network. International Journal of Electrical Power & Energy Systems, 134, Article
racy, advancement and generalization ability of the model are verified. 107365. https://doi.org/10.1016/j.ijepes.2021.107365
Du, X., Han, X., Shen, T., Meng, Z., Chen, K., Yao, X., … Castro-García, S. (2024). Natural
In Experiment 1, we mainly focus on verifying the difference be­ frequency identification model based on BP neural network for Camellia oleifera
tween the combination prediction model and the single neural network fruit harvesting. Biosystems Engineering, 237, 38–49. https://doi.org/10.1016/j.
or statistical prediction. It is found that the combination prediction biosystemseng.2023.11.012
Fadakar, E., & Ebrahimi, M. (2016). A new metaheuristic football game inspired
model has the highest accuracy and stability. In Experiment 2, we algorithm. In 2016 1st Conference on Swarm Intelligence and Evolutionary Computation
mainly verify the applicability of FTTA in the combination prediction (CSIEC). https://doi.org/10.1109/csiec.2016.7482120
model, and verifies that FTTA is the most suitable for the wind speed Huang, H., & Qian, C. (2023). Modeling PM2.5 forecast using a self-weighted ensemble
GRU network: Method Optimization and Evaluation. Ecological Indicators, 156,
prediction system by comparing with other optimization algorithms. In Article 111138. https://doi.org/10.1016/j.ecolind.2023.111138
Experiment 3, we mainly verify the advanced nature of FTTA-VMD Karijadi, I., Chou, S.-Y., & Dewabharata, A. (2023). Wind power forecasting based on
method compared with other data preprocessing methods. In Experi­ hybrid CEEMDAN-EWT deep learning method. Renewable Energy, 218, Article
119357. https://doi.org/10.1016/j.renene.2023.119357
ment 4, we focus on model CPU operation time. Overall, the new wind Lee, K. S., & Geem, Z. W. (2004). A new structural optimization method based on the
speed prediction system we proposed has high applicability and harmony search algorithm. Computers & Structures, 82(9–10), 781–798. https://doi.
advanced nature, and is very suitable for the practical application of org/10.1016/j.compstruc.2004.01.002
Li, D., Jiang, M.-R., Li, M.-W., Hong, W.-C., & Xu, R.-Z. (2023). A floating offshore
wind farms. You can get the code of FTTA from the Supplementary data.
platform Motion Forecasting Approach based on EEMD hybrid ConvLSTM and
Nevertheless, I think there are still some shortcomings in the pro­ Chaotic Quantum Alo. Applied Soft Computing, 144, Article 110487. https://doi.org/
posed model, which is what we need to research in depth in the future: 10.1016/j.asoc.2023.110487
Li, F., Xu, L. D., Jin, C., & Wang, H. (2011). Intelligent Bionic Genetic Algorithm (IB-GA)
and its convergence. Expert Systems with Applications, 38(7), 8804–8811. https://doi.
1) The consideration of terrain factors can be increased. org/10.1016/j.eswa.2011.01.091
2) Other air factors such as temperature and humidity can be increased. Luo, J., & Gong, Y. (2023). Air pollutant prediction based on Arima-WOA-LSTM model.
3) The running speed of the model should be further improved that the Atmospheric Pollution Research, 14(6), Article 101761. https://doi.org/10.1016/j.
apr.2023.101761
wind power plant can get information in time. Mahmoodabadi, M. J., Rasekh, M., & Zohari, T. (2018). TGA: Team Game Algorithm.
Future Computing and Informatics Journal, 3(2), 191–199. https://doi.org/10.1016/j.
CRediT authorship contribution statement fcij.2018.03.002
Mirjalili, S., & Lewis, A. (2016). The whale optimization algorithm. Advances in
Engineering Software, 95, 51–67. https://doi.org/10.1016/j.advengsoft.2016.01.008
Zhirui Tian: Formal analysis, Methodology, Writing – original draft, Mirjalili, S., Mirjalili, S. M., & Lewis, A. (2014). Grey Wolf optimizer. Advances in
Writing – review & editing, Software. Mei Gai: Supervision, Engineering Software, 69, 46–61. https://doi.org/10.1016/j.advengsoft.2013.12.007
Moghdani, R., & Salimifard, K. (2018). Volleyball Premier League algorithm. Applied Soft
Conceptualization. Computing, 64, 161–185. https://doi.org/10.1016/j.asoc.2017.11.043
Nandhini, K., & Balasundaram, S. R. (2014). Extracting easy to understand summary
Declaration of competing interest using differential evolution algorithm. Swarm and Evolutionary Computation, 16,
19–27. https://doi.org/10.1016/j.swevo.2013.12.004
Pandey, H. M., Chaudhary, A., & Mehrotra, D. (2014). A comparative review of
The authors declare that they have no known competing financial approaches to prevent premature convergence in Ga. Applied Soft Computing, 24,
interests or personal relationships that could have appeared to influence 1047–1077. https://doi.org/10.1016/j.asoc.2014.08.025
Parri, S., Teeparthi, K., & Kosana, V. (2024). A hybrid methodology using VMD and
the work reported in this paper.
disentangled features for wind speed forecasting. Energy, 288, Article 129824.
https://doi.org/10.1016/j.energy.2023.129824
Data availability Parsopoulos, K. E., & Vrahatis, M. N. (2007). Parameter selection and adaptation in
unified particle swarm optimization. Mathematical and Computer Modelling, 46(1–2),
198–213. https://doi.org/10.1016/j.mcm.2006.12.019
Data will be made available on request. Purnomo, H. D., & Wee, H.-M. (2013). Soccer game optimization. Meta-Heuristics
Optimization Algorithms in Engineering, Business, Economics, and Finance, 386–420.
Acknowledgements https://doi.org/10.4018/978-1-4666-2086-5.ch013
Ray, S., Lama, A., Mishra, P., Biswas, T., Sankar Das, S., & Gurung, B. (2023). An Arima-
LSTM model for predicting volatile agricultural price series with Random Forest
This work is supported by the National Natural Science Foundation of Technique. Applied Soft Computing, 149, Article 110939. https://doi.org/10.1016/j.
China (42276231), the Humanities and Social Science Fund of Ministry of asoc.2023.110939
Saremi, S., Mirjalili, S., & Lewis, A. (2017). Grasshopper optimisation algorithm: Theory
Education of the People’s Republic of China (22JJD790028), the Project of and application. Advances in Engineering Software, 105, 30–47. https://doi.org/
Liaoning Provincial Department of Education (LJKMZ20221408) and 10.1016/j.advengsoft.2017.01.004
Project of the cultivation of high-end scientific research achievements of Sharifuzzaman Sagar, A. S. M., Chen, Y., Xie, Y., & Kim, H. S. (2024). MSA R-CNN: A
comprehensive approach to remote sensing object detection and scene
Liaoning Normal University in 2023 (23GDL005).

17
Z. Tian and M. Gai Expert Systems With Applications 245 (2024) 123088

understanding. Expert Systems with Applications, 241, Article 122788. https://doi. Wang, J., Lv, M., Li, Z., & Zeng, B. (2023). Multivariate selection-combination short-term
org/10.1016/j.eswa.2023.122788 wind speed forecasting system based on convolution-recurrent network and multi-
Shi, X., Wang, J., & Zhang, B. (2024). A fuzzy time series forecasting model with both objective Chameleon Swarm algorithm. Expert Systems with Applications, 214, Article
accuracy and interpretability is used to forecast wind power. Applied Energy, 353, 119129. https://doi.org/10.1016/j.eswa.2022.119129
Article 122015. https://doi.org/10.1016/j.apenergy.2023.122015 Wang, J., Niu, X., Zhang, L., Liu, Z., & Huang, X. (2024). A wind speed forecasting system
Sun, P., Liu, Z., Wang, J., & Zhao, W. (2024). Interval forecasting for wind speed using a for the construction of a smart grid with two-stage data processing based on
combination model based on multiobjective artificial hummingbird algorithm. improved elm and Deep Learning Strategies. Expert Systems with Applications, 241,
Applied Soft Computing, 150, Article 111090. https://doi.org/10.1016/j. Article 122487. https://doi.org/10.1016/j.eswa.2023.122487
asoc.2023.111090 Wu, J., Huang, J., Wu, X., & Dai, H. (2023). A novel graph-based hybrid deep learning of
Tian, Z., & Gai, M. (2022). New PM2.5 forecasting system based on combined neural cumulative GRU and deeper GCN for recognition of abnormal gait patterns using
network and an improved multi-objective optimization algorithm: Taking the wearable sensors. Expert Systems with Applications, 233, Article 120968. https://doi.
economic belt surrounding the Bohai Sea as an example. Journal of Cleaner org/10.1016/j.eswa.2023.120968
Production, 375, Article 134048. https://doi.org/10.1016/j.jclepro.2022.134048 Xin, J., Chen, J., Li, C., Lu, R., Li, X., Wang, C., … He, R. (2022). Deformation
Tian, Z., & Gai, M. (2023). A novel hybrid wind speed prediction framework based on characterization of oil and gas pipeline by ACM technique based on SSA-BP Neural
multi-strategy Improved Optimizer and new data pre-processing system with Network model. Measurement, 189, Article 110654. https://doi.org/10.1016/j.
feedback mechanism. Energy, 281, Article 128225. https://doi.org/10.1016/j. measurement.2021.110654
energy.2023.128225 Xiong, Z., Yao, J., Huang, Y., Yu, Z., & Liu, Y. (2024). A wind speed forecasting method
Tian, Z., & Wang, J. (2022). Variable frequency wind speed trend prediction system based on EMD-MGM with switching QR loss function and novel subsequence
based on combined neural network and improved multi-objective optimization superposition. Applied Energy, 353, Article 122248. https://doi.org/10.1016/j.
algorithm. Energy, 254, Article 124249. https://doi.org/10.1016/j. apenergy.2023.122248
energy.2022.124249 Zhang, Z., & Yin, J. (2024). Spatial-temporal offshore wind speed characteristics
Tian, Z., & Wang, J. (2023). A wind speed prediction system based on new data prediction based on an improved purely 2d CNN approach in a large-scale
preprocessing strategy and improved multi-objective optimizer. Renewable Energy, perspective using reanalysis dataset. Energy Conversion and Management, 299, Article
215, Article 118932. https://doi.org/10.1016/j.renene.2023.118932 117880. https://doi.org/10.1016/j.enconman.2023.117880

18