The chaotic time series can be expanded to the multidimensional space by phase space reconstruction, in order to reconstruct the dynamic characteristics of the original system. It is difficult to obtain complete phase space for chaotic time series, as a result of the inconsistency of phase space reconstruction. This paper presents an idea of subspace approximation. The chaotic time series prediction based on the phase space reconstruction can be considered as the subspace approximation problem in different neighborhood at different time. The common static neural network approximation is suitable for a trained neighborhood, but it cannot ensure its generalization performance in other untrained neighborhood. The subspace approximation of neural network based on the nonlinear extended Kalman filtering (EKF) is a dynamic evolution approximation from one neighborhood to another. Therefore, in view of incomplete phase space, due to the chaos phase space reconstruction, we put forward subspace adaptive evolution approximation method based on nonlinear Kalman filtering. This method is verified by multiple sets of wind speed prediction experiments in Wulong city, and the results demonstrate that it possesses higher chaotic prediction accuracy.
1. Introduction
In recent years, industrial disasters and accidents occurred frequently, the meteorological and hydrological conditions were complicated and changeable, and financial markets fluctuated drastically. These phenomena often contain chaotic characteristics [1, 2], and prediction [3] for these phenomena is imminent. For a long time, there was no scientific tool for handling this issue, because the changing mechanisms of characteristics in these phenomena were not understood very well. Hence, aiming at the chaotic characteristics, some scholars worked with structures and made a lot of new researches on the prediction of chaotic time series [4–8].
To study and deal with the measurement data of chaotic system, Kennel et al. presented the reconstruction method of phase space system. Two parameters, the embedding dimension m and delay time τ, needed to be determined before the phase space reconstruction [9, 10]. At present, time delay selection methods that are commonly used in the chaotic short-term prediction mainly include autocorrelation method [11], mutual information method [12], and singular value fraction method [13]. Calculating methods of embedding dimension mainly include saturated correlation dimension [14], false nearest neighbors method [15], and Cao's method [16]. Hu and Chen put forward the C-C method [17], which can simultaneously estimate the delay time τ and embedding dimension m with the correlation integral. Autocorrelation method extracts only linear correlation degree between time series, which is hard to be applied to high-dimensional chaos system and nonlinear dynamical system. Mutual information method, which can determine the optimal delay time by calculating the first minimal value of mutual information function, is a nonlinear analysis method, but it cannot avoid massive calculation and cannot satisfy the requirement of complicated space division. It is difficult to determine the threshold of the singular value, because the singular value fraction method is largely affected by noise. When the embedded dimension with the saturated correlation dimension is calculated, the main question is to choose the different neighborhood radius. The radius selection has certain randomness, and the result will be in large deviation with improper choice, because of the influence of noise in the data and excessive concentration of the data. The determination of threshold has very strong subjectivity when we use false nearest neighbors method to determine the embedding dimension. There is no objective standard to determine the threshold value, especially for the experimental data, which may get a wrong result. Cao's method, an improved false nearest neighbors method, can effectively distinguish random signals and deterministic signals, and embedding dimension can be obtained through a less amount of data. C-C method is based on the statistical theory, so m cannot be precisely determined.
Researches have showed that different phase space reconstruction methods get different m and τ. Moreover, the same chaotic time series with the same kind of method in different times may get different m and τ. There is no phase space reconstruction that can obtain complete and independent phase space. After phase space reconstruction, prediction model is often established through the functional approximation method.
The prediction model based on phase space reconstruction has been used to adopt the functional approximation method based on the neural network [18–21], which has strong nonlinear fitting capability and can approximate any complex nonlinear relationships. However, since neural network is only suitable for approximation of a deterministic system, it is difficult to guarantee the time-varying system performance and ensure its generalization performance in other untrained neighborhood. Meanwhile, the prediction effect of neural network is not good, because the chaotic time series is a complex nonlinear uncertain system.
In this study, we introduce Kalman filtering to neural network model [22], inspired by Kalman iteration and Bucy and Sunahara’s nonlinear extended Kalman filtering theory [23]. The subspace approximation of neural network based on the nonlinear extended Kalman filtering (EKF) has a function which is dynamic evolution approximation from one neighborhood to another. Therefore, we can constitute a phase space by choosing a kind of phase space reconstruction method, and the space may be incomplete, not separate, and can be seen as a subspace of the ideal phase space. On this basis, we put forward adaptive neural network model based on nonlinear Kalman filtering and finally realize the subspace approximation of dynamic evolution system. In addition, we simulate wind speed series in Wulong city using the proposed method. By comparing with BP neural network prediction model, the results show that our method possesses higher prediction accuracy.
The paper is organized as follows. Section 2 discusses about the subspace approximation of phase space reconstruction. In Section 3, we describe the neural network model based on nonlinear Kalman filtering. Section 4 uses practical examples and series tests to verify the proposed method, while Section 5 contains the conclusions of the present work.
2. Subspace Approximation of Phase Space Reconstruction
Reconstructing phase space by chaos theory needs to identify the chaos of time series. Single variable time series can be reconstructed into a phase space by Takens’ embedding theorem in phase space reconstruction [24, 25]; that is, the original dynamical system can be restored in the sense of topological equivalence as long as the embedding dimension is sufficiently high. For the observed time series x(1),x(2),…,x(t), after time delay reconstruction by Takens embedding theorem, it will receive a set of space vector
(1)X(t)={x(t),x(t+τ),…,x(t+(m-1)τ)},000000000t=1,2,…M,M=N-(m-1)τ.
After phase space reconstruction, the data space is
(2)[x(1)x(2)⋯x(t)x(1+τ)x(2+τ)⋯x(t+τ)⋮⋮⋱⋮x(1+(m-1)τ)x(2+(m-1)τ)⋯x(t+(m-1)τ)].
Accordingly, we acquire
(3)f:Rm⟶R,
where f is a single-valued function. Then, we have
(4)x(t+mτ)=f(x(t),x(t+τ),…,x(t+(m-1)τ)).
However, it cannot be really obtained as the data are often limited. Hence, f^:Rm→R can only be constituted by limited measurement data, making f^ sufficiently approximate to f, consequently we can get a nonlinear prediction model.
This paper employs the neural network to predict chaotic series. However, the neural network cannot readily handle the inconsistency of the phase space reconstruction because of uncertain nonlinear chaotic time series. Therefore, it is crucial to adaptively construct subspace to approximate chaotic series through the incomplete phase space. The feature of adaptive subspace approximation is that it can add new data in real time and forget old data in the process of training. Consequently, weights and thresholds of the neural network are continuously modified to realize the dynamic evolution modeling.
3. Neural Network Model Based on Nonlinear Kalman Filtering
Kalman filtering has good adaptability. It can dynamically update and forecast the system information in real time with limited data. However, it cannot be readily used for complicated nonlinear model. Meanwhile, the extended Kalman filtering (EKF) is a kind of effective method to handle nonlinear filtering.
The mathematical model of EKF is as follows:
(5)Xk+1=f(Xk,k)+Γ(Xk,k)WkZk=h(Xk,k)+Vk,
where Wk and Vk are independent, zero mean, and Gaussian random processes with covariance matrices Q and R, respectively. The statistical properties are as follows:
(6)p(w)~N(0,Q),p(v)~N(0,R).
EKF spreads nonlinear functions f(·) and h(·) to Taylor series around filtering value X^k and predicted value X^k-, respectively, only retaining the first-order information. Hence, the linearization model of the nonlinear system is obtained, and then we can obtain the EKF formula in nonlinear system by basic equations of Kalman filtering.
Given a forward network with N layers, the numbers of neurons in each layer are Sk(k=1,2,…,N). Suppose that input layer is the first layer and output layer is the Nth layer. The weights of the kth layer neurons are Wijk(i=1,2,…,Sk-1;j=1,2,…,Sk). In order to convert the calculation of connection weights Wijk in the above problem into filter recursive estimation form, we let all of the network weights constitute the state vector
(7)W=[W111⋯WS1S21W112⋯WS2S32⋯W11N-1WSN-1SNN-1]T,
where state vector W consists of all of the weights according to the linear array, and its dimension is as follows:
(8)NW=∑i=1N-1SiSi+1.
Then the state equation and measurement equation of the system can be expressed as
(9)Wk=Wk-1,(10)Yek=h(Wk,Xk)+Vk=Yrk+Vk,
where Yek is the expected output, Xk is the input vector, and Yrk is the actual output.
The measurement noise Vk is assumed to be additive, white, and Gaussian, with zero mean and with covariance matrix defined by
(11)E(Vk)=0,E(VkVkT)=Rk.
Suppose that the output of the jth node for the lth layer in the kth iteration is
(12)Ojkl=Fjl(Wjkl,Okl-1).
From (10) and (12), we have
(13)Yek=h(Wk,Xk)+Vk=FN(WkN,FN-1(WkN-1⋯F2(Wk2,Xk)))+Vk,Yek=h(W^k-,Xk)+∂h∂W|Wk=W^k-(Wk-W^k-)+Vk.
Assume that
(14)∂h∂W|Wk=W^k-=Hk,h(W^k-,Xk)-∂h∂W|Wk=W^k-W^k-=Ck.
Accordingly, the measurement equation may also be expressed as
(15)Yek=HkWk+Ck+Vk.
The Jacobian matrix of the function h(·) is described by
(16)Hk=[∂h1∂w1∂h1∂w2⋯∂h1∂wn∂h2∂w1∂h2∂w2⋯∂h2∂wn⋮⋮⋱⋮∂hn∂w1∂hn∂w2⋯∂hn∂wn].
Similarly, all thresholds of the network constitute the state vector
(17)b=[b11⋯bS21b12⋯bS32⋯b1N-1⋯bSNN-1]T,
where the dimension is
(18)Nb=∑i=1N-1Si+1.
Suppose that W and b are both state variable; that is, the state vector composed of weights and thresholds is described by
(19)θ=[W,b]T=[W111⋯WS1S21b11⋯bS21W112⋯WS2S32b12000⋯bS32⋯W11N-1WSN-1SNN-1b1N-1⋯bSNN-1]T.
Kalman filtering algorithm on training weights and thresholds of the neural network is as in Table 1.
Qk-1 and Rk are process noise covariance and measurement noise covariance, respectively, Hk is the Jacobian matrix of observable model, θ^k- is the optimal predictive value for step k according to step k-1, and θ^k is the optimal filter estimate for step k.
4. Simulation Examples4.1. Determining of Embedding Dimension and Delay Time
One of the most popular chaos logistic mapper is selected as the study object. Logistic equation is
(20)xn+1=αxn(1-xn),α∈[0,4].
The related time series are produced according to (20). It is a chaotic system when α=4. Assume that initial value of series is 0.1, and 4000 points are calculated. The first 1000 points are eliminated as transition phenomenon, leaving the remaining 3000 points to reconstruct phase space. Before the phase space reconstruction, we determine the embedding dimension m and delay time τ. A comparison among several methods is present in Table 2.
Comparison among phase space reconstruction methods.
Parameter
Method
Autocorrelation
Mutual information
False nearest neighbors
Cao
C-C
τ
1
12
—
—
5
m
—
—
4
3
5
“—” means nothing.
Obviously, the optimal embedding dimension and delay time are generally different by different methods of phase space reconstruction.
In order to verify the fact that data at different time will obtain different embedding dimension m and delay time τ with the same phase space reconstruction method, we have the following experiment.
The remaining 3000 points (k=1,2,…,3000) are divided into five parts, with time intervals T1, T2, T3, T4, and T5, respectively. Embedding dimension and delay time are present in Table 3 by C-C method.
Parameters of the same phase space reconstruction during different time periods.
Parameter
Interval
T1(k=1,2,…,600)
T2(k=601,602,…,1200)
T3(k=1201,…,1800)
T4(k=1801,…,2400)
T5(k=2401,…,3000)
τ
5
3
4
4
5
m
5
5
4
9
10
Apparently, the data during different time periods will acquire different embedding dimension and delay time by using the same phase space reconstruction method.
4.2. Wind Speed Chaotic Series Forecasting Simulation
Analysis about the chaotic characteristics of wind speed in the process of wind power generation has been presented in a related article [26]. We record one of the wind speed data every 10 minutes, and 150 groups of wind speed data in Wulong city are used to simulate experiments in our study. We obtain the corresponding m and τ by different phase space reconstruction methods, as shown in Table 4.
Comparison among phase space reconstruction methods.
Parameter
Method
Autocorrelation
Mutual information
False nearest neighbors
Cao
C-C
τ
1
12
—
—
3
m
—
—
3
7
4
Various combinations are present in Table 5.
Various combinations on two forecasting methods.
Model
Combination
Parameter
Forecasting
a1
Autocorrelation + false nearest neighbors
τ=1, m=3
BPNN
b1
Autocorrelation + false nearest neighbors
τ=1, m=3
EKFNN
a2
Mutual information + false nearest neighbors
τ=12, m=3
BPNN
b2
Mutual information + false nearest neighbors
τ=12, m=3
EKFNN
a3
Autocorrelation + Cao
τ=1, m=7
BPNN
b3
Autocorrelation + Cao
τ=1, m=7
EKFNN
a4
Mutual information + Cao
τ=12, m=7
BPNN
b4
Mutual information + Cao
τ=12, m=7
EKFNN
a5
C-C
τ=3, m=4
BPNN
b5
C-C
τ=3, m=4
EKFNN
Wind speed prediction [27, 28] of chaotic time series about neural network model usually extracts phase space reference points as the BP neural network training samples on the basis of phase space reconstruction. We establish the neural network model based on nonlinear Kalman filtering, including two parts: predict wind speed and constantly modify weights and thresholds of the neural network by Kalman recursion. In this paper, BPNN model in the same structure is employed to forecast wind speed time series, in order to illustrate the validity of EKFNN on predicting the chaotic time series. The same 150 groups of wind speed data are used to simulate experiments. The predicted curves and error curves are shown in Figures 1, 2, 3, 4, and 5.
The effect comparison between a1 and b1.
The predicted wind speed data
Relative wind speed error
The effect comparison between a2 and b2.
The predicted wind speed data
Relative wind speed error
The effect comparison between a3 and b3.
The predicted wind speed data
Relative wind speed error
The effect comparison between a4 and b4.
The predicted wind speed data
Relative wind speed error
The effect comparison between a5 and b5.
The predicted wind speed data
Relative wind speed error
Comparisons among several models in four indices are present in Table 6.
Different wind speed model index.
Model
Error
MAE
MRE
MSE
SSE
a1
0.8482
0.1076
1.3786
206.7967
b1
0.3365
0.0416
0.1919
28.7834
a2
1.2279
0.1674
2.9614
444.2152
b2
0.6190
0.0759
0.6209
93.1357
a3
0.5106
0.0656
0.4089
61.3361
b3
0.3783
0.0458
0.2654
39.8130
a4
2.4005
0.3079
8.4766
1.2715e+003
b4
0.0852
0.0110
0.0177
2.6584
a5
0.8952
0.1173
1.7970
269.5495
b5
0.2867
0.0357
0.1359
20.3850
MAE, MRE, MSE, and SSE are Mean Absolute Error, Mean Relative Error, Mean Square Error, and Sum of Squared Error, respectively.
We list 12 groups, a total of 2 hours of wind speed forecasting results in two methods, under the same phase space reconstruction. Compare the prediction performance in the next 10 min, 20 min, 30 min, and up to, 120 min.
Comparisons among several prediction results in two methods are present in Table 7.
Observed wind speed data and predicted data.
Time (min)
Observed value (m/s)
Predicted value (m/s)
Relative error (%)
Predicted value (m/s)
Relative error (%)
a1
b1
10
10.4900
10.0801
3.9561
10.5101
0.1823
20
10.3700
9.9783
3.7762
10.4011
0.2630
30
10.6900
10.1024
5.4902
10.4621
2.1476
40
10.2100
10.8802
6.5374
10.6202
4.0235
50
10.3000
9.1424
11.2402
10.1531
1.4947
60
9.8000
10.3508
5.6401
10.2210
4.2446
70
9.3000
8.6881
6.5810
9.6032
3.2585
80
9.4900
8.1883
13.7201
9.3163
1.8295
90
10.1100
9.6013
5.0321
9.7631
3.4282
100
9.2100
10.4726
13.7200
10.0300
8.9132
110
8.5500
7.2082
15.6903
8.8621
3.6533
120
9.0000
7.0824
21.3101
8.5470
5.0347
a2
b2
10
10.4900
9.6930
7.5977
10.4785
0.1095
20
10.3700
10.4006
0.2955
10.4249
0.5292
30
10.6900
10.0489
5.9968
10.4992
1.7845
40
10.2100
10.5699
3.5250
10.3357
1.2311
50
10.3000
10.2559
0.4285
10.2518
0.4675
60
9.8000
10.1445
3.5151
10.1132
3.1960
70
9.3000
9.2583
0.4488
10.0703
8.2826
80
9.4900
9.8944
4.2609
9.9547
4.8963
90
10.1100
10.6937
5.7739
9.7503
3.5580
100
9.2100
11.4420
24.2351
9.5212
3.3787
110
8.5500
11.8315
38.3804
9.1823
7.3954
120
9.0000
11.2543
25.0475
9.3012
3.3463
a3
b3
10
10.4900
10.9775
4.6468
10.7194
2.1870
20
10.3700
10.4931
1.1872
10.4277
0.5567
30
10.6900
10.3878
2.8269
10.4168
2.5559
40
10.2100
13.7393
34.5672
10.1893
0.2032
50
10.3000
9.8085
4.7716
10.2918
0.0793
60
9.8000
12.1826
24.3126
10.0993
3.0540
70
9.3000
9.7924
5.2943
9.8370
5.7739
80
9.4900
9.3521
1.4529
9.6689
1.8855
90
10.1100
14.5803
44.2168
9.7772
3.2920
100
9.2100
14.6095
58.6264
9.5805
4.0233
110
8.5500
8.6425
0.0821
9.3442
9.2884
120
9.0000
9.7379
8.1988
9.3314
3.6817
a4
b4
10
10.4900
8.9295
14.8759
10.1137
3.5868
20
10.3700
9.2037
11.2466
9.7344
6.1296
30
10.6900
8.6494
19.0887
10.3308
3.3597
40
10.2100
8.5516
16.2426
9.9803
2.2501
50
10.3000
9.3038
9.6717
10.5685
2.6064
60
9.8000
8.7467
10.7479
10.1183
3.2477
70
9.3000
8.7126
6.3161
9.7779
5.1391
80
9.4900
8.6188
9.1802
9.8413
3.7021
90
10.1100
9.4713
6.3176
10.9886
8.6901
100
9.2100
9.3204
1.1992
8.8816
3.5657
110
8.5500
9.5029
11.1452
8.2497
3.5118
120
9.0000
9.6909
7.6768
8.8558
1.6017
a5
b5
10
10.4900
10.1227
3.5017
10.4227
0.6412
20
10.3700
10.5150
1.3985
10.4890
1.1473
30
10.6900
10.3714
2.9805
10.6268
0.5909
40
10.2100
10.1625
0.4652
10.5638
3.4649
50
10.3000
10.5236
2.1705
10.4778
1.7260
60
9.8000
10.5366
7.5159
10.1741
3.8174
70
9.3000
9.8748
6.1807
9.8224
5.6174
80
9.4900
9.4500
0.4217
9.6744
1.9432
90
10.1100
9.7279
3.7794
9.8950
2.1267
100
9.2100
8.9728
2.5755
9.6195
4.4458
110
8.5500
9.3838
9.7519
9.2464
8.1455
120
9.0000
9.3328
3.6973
9.0710
0.7889
Figures 1–5 show that relative error of wind speed prediction by EKF neural network is much smaller than that by BP neural network, through observing the future wind speed prediction of 150 groups. As can be seen in Table 6, the prediction effects are largely different by different kinds of phase space reconstruction methods. Four performance indices, which are Mean Absolute Error (MAE), Mean Relative Error (MRE), Mean Square Error (MSE), and Sum of Squared Error (SSE), of EKF neural network, are also far less than those of corresponding general neural network.
Apparently, EKF neural network can solve the inconsistency problem of phase space reconstruction and approximate chaotic time series well through subspace. The neural network model based on EKF has outstanding adaptability, so it can predict the wind speed chaotic time series with higher precision, compared with BP neural network.
Furthermore, we can conclude that in Table 7, prediction accuracy of EKF neural network is higher than that of BP neural network, by comparing the prediction performance of wind speed in the next 10 min, 20 min, 30 min, and up to, 120 min. It demonstrates that EKF neural network model, which has better dynamic adaptability, can better the prediction of wind speed time series with nonlinear chaotic characteristics. Therefore, the proposed phase space reconstruction method of the adaptive evolution approximation in this paper is an effective approach.
5. Conclusion and Further Work
The phase space reconstruction cannot meet characteristics of the completeness and independence, and the results with different reconstruction methods are obviously inconsistent. The reconstructed phase space is a subspace of the ideal space. If a subspace approximation can make the real-time dynamic evolution, then the initial constructed phase space, for which the evolution is adaptive subspace approximation, can finally approximate to the ideal phase space much better.
In this paper, neural network model based on nonlinear Kalman filter is established, by dynamic adaptivity of nonlinear Kalman filter. The model will add new samples in real time and gradually eliminate previous data, as a moving samples window, and the evolution of the training sample continually updates weights and thresholds of the neural network. As a result, adaptive subspace approximation is implemented by reconstructed incomplete phase space.
The optimized plan, which combines the nonlinear Kalman filter with neural network, sufficiently utilizes the nonlinear approximation capability of neural network and dynamic adaptive ability of real-time update correction of nonlinear Kalman filter. Consequently, it can realize subspace adaptive evolution approximation and solve the inconsistency problem of phase space reconstruction. Therefore, it is a nice direction in research into chaotic prediction. Future research can be performed in a number of areas. It provides a good technical support in studying problems of meteorology, hydrology, and finance fields.
Acknowledgments
This work was supported by the National Science Foundation of China (no. 51075418), the National Science Foundation of China (no. 61174015), Chongqing CMEC Foundations of China (no. CSTC 2013jjB40007), and Chongqing Scientific Personnel Training Plan of China (no. CSTC 2013kjrc-qnrc40008).
KaddoumG.GagnonF.ChargéP.RovirasD.A generalized BER prediction method for differential chaos shift keying system through different communication channels20126424254372-s2.0-8485973608710.1007/s11277-010-0207-1SivakumarB.Chaos theory in geophysics: past, present and future200419244146210.1016/S0960-0779(03)00055-9MR2007457MastorocostasP. A.TheocharisJ. B.BakirtzisA. G.Fuzzy modeling for short term load forecasting using the orthogonal least squares method199914129362-s2.0-003308024010.1109/59.744480YangH. Y.YeH.WangG.KhanJ.HuT.Fuzzy neural very-short-term load forecasting based on chaotic dynamics reconstruction20062924624692-s2.0-3144444303210.1016/j.chaos.2005.08.095SivakumarB.A phase-space reconstruction approach to prediction of suspended sediment concentration in rivers20022581–41491622-s2.0-003718638110.1016/S0022-1694(01)00573-XPorporatoA.RidolfiL.Nonlinear analysis of river flow time sequences1997336135313672-s2.0-0030619033ZhaoP.XingL.YuJ.Chaotic time series prediction: from one to another200937325217421772-s2.0-6734922741110.1016/j.physleta.2009.04.033AbarbanelH. D. I.1996New York, NY, USASpringer10.1007/978-1-4612-0763-4MR1363486KennelM. B.BrownR.AbarbanelH. D. I.Determining embedding dimension for phase-space reconstruction using a geometrical construction1992456340334112-s2.0-3594900679110.1103/PhysRevA.45.3403MaH.HanC.Selection of embedding dimension and delay time in phase space reconstruction2006111111142-s2.0-3374608170110.1088/0957-0233/17/1/018ChenB.LiuG.TangJ.ZhangY.CaiP.HuangJ.WuY.-S.Research on chaotic sequence autocorrelation by phase space method20103968598632-s2.0-7865002515810.3969/j.issn.1001-0548.2010.06.012JiangA.HuangX.ZhangZ.LiJ.ZhangZ.-Y.HuaH.-X.Mutual information algorithms2010248294729602-s2.0-7804942272010.1016/j.ymssp.2010.05.015Abu-ShikhahN.ElkarmiF.Medium-term electric load forecasting using singular value decomposition2011367425942712-s2.0-7995937552510.1016/j.energy.2011.04.017GaoC.LiuX.Chaotic identification of BF ironmaking process I: the calculation of saturated correlative dimension20044043473502-s2.0-3142695593CarriónI. M.AntúnezE. A.Thread-based implementations of the false nearest neighbors method20093510-1152353410.1016/j.parco.2009.09.003HiteJ.Jr.1999Amsterdam, The NetherlandsGulf Professional PublishingHuY.ChenT.Phase-space reconstruction technology of chaotic attractor based on C-C method20123510-11425430FadareD. A.The application of artificial neural networks to mapping of wind speed profile for energy application in Nigeria20108739349422-s2.0-7184908512710.1016/j.apenergy.2009.09.005Salcedo-SanzS.Ángel M. Pérez-BellidoA. M.Ortiz-GarcíaE. G.Portilla-FiguerasA.PrietoL.ParedesD.Hybridizing the fifth generation mesoscale model with artificial neural networks for short-term wind speed prediction2009346145114572-s2.0-5904909294510.1016/j.renene.2008.10.017DracopoulosD. C.1997London, UKSpringerXueJ.ShiZ.Short-time traffic flow prediction based on chaos time series theory20088568722-s2.0-55649085084YangH.LiJ.DingF.A neural network learning algorithm of chemical process modeling based on the extended Kalman filter2007704–66256322-s2.0-3384600728010.1016/j.neucom.2006.10.033BucyR. S.SenneK. D.Digital synthesis of non-linear filters1971732872982-s2.0-0001231547WangJ.XieY.Solar radiation prediction based on phase space reconstruction of wavelet neural network20111514603460710.1016/j.proeng.2011.08.864KantzH.SchreiberT.1997Cambridge, UKCambridge University PressMR1472976KarakasidisT. E.CharakopoulosA.Detection of low-dimensional chaos in wind time series2009414172317322-s2.0-6734916548810.1016/j.chaos.2008.07.020LoukaP.GalanisG.SiebertN.KariniotakisG.KatsafadosP.PytharoulisI.KallosG.Improvements in wind speed forecasts for wind power prediction purposes using Kalman filtering20089612234823622-s2.0-4974913892310.1016/j.jweia.2008.03.013PoncelaM.PoncelaP.PeránJ. R.Automatic tuning of Kalman filters by maximum likelihood methods for wind energy forecasting20131081234936210.1016/j.apenergy.2013.03.041