Wind power prediction method and system based on dominant component analysis and IOOA-KELM

Abstract

The wind power prediction method based on dominant component analysis and IOOA-KELM comprises the following steps: acquiring original data of wind power, and decomposing a time sequence of the wind power through ICEEMDAN to obtain a plurality of subsequences; performing dominant component extraction on a plurality of subsequences obtained after decomposition by adopting SSA, and superposing and reconstructing residual errors of the subsequences; dividing the data after the dominant component extraction into a test set, a training set and a verification set; the method comprises the following steps: changing random initialization into Logistic chaos initialization, introducing a Levy flight strategy when updating a local development position, and improving an OOA (Octave Optimization Algorithm) to obtain an improved IOOA; constructing a KELM model, respectively predicting all subsequences, and calibrating hyper-parameters of the KELM model by using an improved eagle optimization algorithm IOOA to obtain optimal parameters of an IOOA-KELM prediction model; and superposing the prediction results of all the subsequences to obtain a final wind power prediction result. According to the method, the problem of difficulty in selecting hyper-parameters of a kernel extreme learning machine prediction model is solved, and the precision of wind power prediction is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of wind power prediction, and in particular to a wind power prediction method and system based on dominant component analysis and IOOA-KELM. Background Art

[0002] As a clean and renewable energy, wind power has become an important part of the global energy transformation. However, due to the randomness and intermittency of wind resources, the prediction of wind power has great uncertainty, which poses a severe challenge to the stable operation of the power grid and the effective operation of the power market. Therefore, it is particularly important to make accurate short-term wind power forecasts.

[0003] Wind power forecasting methods are mainly divided into three categories: physical method, statistical method and artificial intelligence method. The physical method has a complex calculation process, high cost and large short-term wind power forecasting error. Statistical method and artificial intelligence method only need to use historical wind power data, which is easy to implement and more suitable for short-term wind power forecasting. Since the original wind power data has the characteristics of nonlinearity and non-stationarity, the prediction performance of statistical methods is restricted. Artificial intelligence methods have outstanding advantages in dealing with nonlinear problems, so they have been applied to the field of wind power forecasting by many researchers. For example, the literature [1]: Yang Mao, Zhang Shutian, Wang Bo. Short-term wind power forecasting method based on causal regularized extreme learning machine [J]. Power System Protection and Control, 2024, 52(11): 127-136.

[0004] Artificial intelligence methods mainly include support vector machines, long short-term memory neural networks, extreme learning machines, etc. Although extreme learning machines have the advantages of fast training speed and simple implementation, their reliance on random initialization and difficulty in selecting the number of hidden layer neurons limit their application. In contrast, the kernel extreme learning machine introduces kernel functions, which enables the model to effectively handle complex nonlinear relationships in high-dimensional feature space. However, the kernel extreme learning machine has the problem of difficulty in selecting hyperparameters, so an accurate and effective optimization algorithm is needed to optimize its parameters to improve the performance of the prediction model. For example, the literature [2]: Wang Hui, Zou Zhichao, Li Xin, et al. Short-term wind power forecasting based on VMD-ISSA-GRU combined model [J]. Thermal Power Generation, 2024, 53(05): 122-131.

[0005] At the same time, in order to reduce the non-stationarity of the original data of wind power and improve the prediction accuracy of the model. Signal decomposition is used to preprocess the wind power sequence, such as empirical mode decomposition EMD, fully integrated empirical mode decomposition CEEMDAN, and improved fully integrated empirical mode decomposition ICEEMDAN. However, EMD has problems such as modal aliasing and endpoint effect, and the results of CEEMDAN depend on the type of noise added. The improved fully integrated empirical mode decomposition ICEEMDAN further reduces modal aliasing and improves the stability and robustness of the decomposition results by introducing an adaptive noise control strategy. And because there is still noise in the subsequence, the denoising method is used to perform dominant component analysis to achieve denoising effect, thereby improving the prediction accuracy of the model. Summary of the invention

[0006] The present invention provides a wind power prediction method and system based on dominant component analysis and IOOA-KELM. The method processes the original time series through an improved adaptive noise fully integrated empirical mode decomposition (ICEEMDAN) to obtain multiple subsequences, and then uses singular spectrum analysis (SSA) to extract the dominant component of each subsequence and reconstruct the residuals by superposition, so as to reduce the non-stationarity of the original wind power time series. In addition, an improved Osprey optimization algorithm (IOA) is proposed to solve the problem of difficult hyperparameter selection of a kernel extreme learning machine prediction model, obtain the optimal hyperparameters of the IOOA-KELM prediction model, and improve the accuracy of wind power prediction.

[0007] The technical solution adopted by the present invention is:

[0008] The wind power prediction method based on dominant component analysis and IOOA-KELM includes the following steps:

[0009] Step 1: Obtain the original data of wind power, decompose the time series of wind power through the improved adaptive noise fully integrated empirical mode decomposition ICEEMDAN, reduce its non-stationarity, and obtain multiple subsequences;

[0010] Step 2: Use singular spectrum analysis (SSA) to extract the dominant components of the multiple subsequences obtained after decomposition in step 1 and reconstruct the residuals of each subsequence by superposition to further reduce its non-stationarity;

[0011] Step 3: Divide the data after extracting the dominant components into a test set, a training set, and a validation set according to a ratio of 8:1:1;

[0012] Step 4: By changing random initialization to Logistic chaos initialization and introducing Levy flight strategy when updating the local development position, the traditional Osprey optimization algorithm OOA is improved to obtain the improved Osprey optimization algorithm IOOA;

[0013] Step 5: Construct a KELM model, predict all subsequences separately, and use the improved Osprey optimization algorithm IOOA obtained in step 4 to calibrate the KELM model hyperparameters to obtain the optimal parameters of the IOOA-KELM prediction model;

[0014] Step 6: Superimpose the prediction results of all subsequences to obtain the final wind power prediction result;

[0015] Step 7: Use the root mean square error RMSE, mean absolute error MAE and determination coefficient R 2 Three key evaluation indicators are used to analyze the results of the wind power prediction model.

[0016] The step 1 comprises the following steps:

[0017] S1.1: Add adaptive white noise E1(w) to the original signal x(t) (i) );

[0018] x (i) =x(t)+β0E1(w (i) )

[0019] Where: x (i) represents the signal with white noise added for the i-th time; x(t) represents the original signal; w (i) represents the i-th added white noise i∈{1,2,…,M}, M represents the number of added white noises; β0 represents the noise standard deviation of the first decomposition signal, E i (·) represents the operator for calculating the i-th EMD decomposition;

[0020] S1.2: Calculate each The local average of the signal and the residual r1 of the first decomposition are obtained:

[0021]

[0022] Where: <·> represents the operator for calculating the average value of the signal, and M(·) represents the operator for calculating the local average value of the signal.

[0023] S1.3: Subtract the residual r1 of the first decomposition from the original signal x(t) to obtain the first component IMF1 of the original signal x(t);

[0024] IMF1=x-r1

[0025] S1.4: Determine the second component IMF2:

[0026] r2= <M(r1+β1E2(w (i) ))>

[0027] IMF2=r1-r2

[0028] Where: r2 represents the residual of the second decomposition; β1 represents the noise standard deviation of the second decomposition signal; E2(w (i) ) represents the operator for calculating the second EMD decomposition; M(r1+β1E2(w (i) )) represents an operator for calculating the local average value of a signal;

[0029] S1.5: For k = 3, 4, ..., N1, k represents the current number of decompositions, N1 represents the maximum number of decompositions; calculate the kth component IMF k :

[0030] r k = <M(r k-1 +β k-1 E k (w (i) ))>

[0031] IMF k =r k-1 -r k

[0032] Where: r k represents the residual of the kth decomposition, r k-1 represents the residual of the k-th decomposition, β k-1 represents the noise standard deviation of the second decomposition signal, E k (w (i) ) represents the operator for calculating the kth EMD decomposition, M(r k-1 +β k-1 E k (w (i) )) represents an operator for calculating the local average value of a signal;

[0033] S1.6: k = k + 1, repeat S1.5 until function r k = <M(r k-1 +β k-1 E k (w (i) ))>Cannot be decomposed, the improved adaptive noise fully integrated empirical mode decomposition algorithm ends.

[0034] The step 2 comprises the following steps:

[0035] S2.1: First, suppose the time series data {h i ,i=1,2,…,N2},h i Represents the i-th data, N2 represents the total number of data, and constructs the Hankel trajectory matrix;

[0036]

[0037] Where: h1 represents the first data, h2 represents the second data, h3 represents the third data, h k represents the k-th data, h k+1 represents the (k + 1)-th data, h L represents the L-th data, h L+1 represents the (L + 1)-th data, represents the N2-th data, L represents the sliding window length parameter, and 1 < L < N2, k is N2 - L + 1; the matrix H is the trajectory matrix.

[0038] S2.2: Perform singular value decomposition (SVD) on the trajectory matrix. At this time, the matrix H can be rewritten as:

[0039]

[0040] Where: d = rank(H) represents the intrinsic component with non-zero eigenvalues, H i represents the elementary matrix, λ i represents the singular values arranged in descending order, U i and V i represent the relevant left and right eigenvectors respectively, V i T represents the right eigenvector V i 's transpose matrix.

[0041] S2.3: According to the magnitudes of the singular values, arrange the corresponding singular vectors in descending order, and extract the dominant components of the trajectory matrix H, that is, reconstruct the matrix H using the combination of the first r elementary matrices:

[0042]

[0043] Where: represents the reconstructed signal, r represents the number of features used for signal reconstruction, the residual R represents noise, and i represents the current i-th elementary matrix.

[0044] The said step 4 includes the following steps:

[0045] Step S4.1: Using a chaotic sequence to replace the pseudo-random number for population initialization can increase the diversity of the initial population and improve the quality of the initial solution. Introduce Logistic chaotic initialization to reduce the randomness fluctuation of population initialization, thereby improving the robustness of the algorithm. Let the number of osprey populations be N3, the upper and lower bounds of the search space be ub and lb respectively, and the initialization formula of the Logistic chaotic mapping can be shown as:

[0046] x i = α·x i-1 ·(1 - x i-1),0≤x0≤1,i=1,2,...,N3

[0047] Where: x i represents the i-th iteration value, x i-1 represents the i-1th iteration value, α∈[3.57,4], x0 represents the initial value, N3 represents the number of osprey populations, and the chaotic value x i The formula for mapping to the search space is:

[0048]

[0049] Where: Represents the initial value of population optimization;

[0050] Step S4.2: In order to reduce the probability of the OOA algorithm falling into the local optimum during the iteration process and improve the local search accuracy, the Levy flight strategy is introduced to perturb the population when updating the local development position to balance the global search and local development. The Levy flight formula is expressed as follows:

[0051]

[0052] In the formula: α is a constant, usually 1.5; x and y are two variables that obey the normal distribution. The calculation formula is as follows:

[0053] x=Normal(0,σx2)

[0054] y=Normal(0,σy2)

[0055]

[0056] The first stage search and fishing position update formula of the improved OOA algorithm is as follows:

[0057]

[0058] Where: represents the updated new position of the ith osprey in the jth dimension in the first stage, l represents Levy flight, x i.j represents the position of the i-th osprey in the j-th dimension before the update, SF ij represents the i-th fish selected in the j-th dimension, r i.j and I i.j Respectively represent random numbers between the interval 0 and 1 and random numbers between the interval 1 and 2;

[0059] The position update formula for the second stage of hunting and eating is as follows:

[0060]

[0061] Where: represents the updated new position of the i-th osprey in the j-th dimension in the second stage; ub and lb represent the upper and lower bounds of the search space respectively; r represents a random number between 0 and 1.

[0062] In step 5, the KELM model is a single hidden layer feedforward neural network learning algorithm based on ELM. Its basic idea is to map data to a high-dimensional space through a kernel function and perform linear operations in the space, thereby avoiding explicit calculation mapping. Its model structure is as follows: Figure 5 shown.

[0063] The KELM model predicts all subsequences separately, and for each subsequence x i , the goal of KELM is to find a function mapping relationship to predict y i The calculation formula is as follows:

[0064] h new =[K(x new1 ,x1),…,K(x newm ,x m )]

[0065]

[0066] Where: h new represents the kernel function value, x1 represents the first subsequence; x new1 and x newm Respectively represent the new input subsequences corresponding to the first subsequence and the mth subsequence; K(x new1 ,x1) represents the kernel function; represents the predicted value; β represents the output weight between the hidden layer and the output layer.

[0067] The step 5 comprises the following steps:

[0068] S5.1: The population is initialized using the Logistic chaotic map in step S4.1, and the regularization coefficient C and the kernel parameter γ of the KELM model are used as the dimension of the osprey position in the improved osprey optimization algorithm 10A, that is, the dimension size is 2. In wind power prediction, the regularization coefficient C is used to control the complexity of the KELM model, and the kernel parameter γ determines the width of the radial basis function kernel in the KELM.

[0069] S5.2: Calculate the fitness value F of each osprey. The fitness value F represents the error between the wind power value predicted by the KELM model and the actual value. The root mean square error (RMSE) is used as a fitness evaluation index to determine the initial fish school position FP corresponding to the osprey.

[0070] S5.3: According to the position update strategy of the osprey at different stages, the position of each osprey is updated, and the updated position is used as the new regularization coefficient C and kernel parameter γ of the KELM model for verification, and the corresponding fitness value is calculated;

[0071] S5.4: Compare the fitness values ​​of the current osprey position and the local optimal position of the osprey to obtain the optimal position X of the osprey best ;

[0072] S5.5: Determine whether the current number of iterations has reached the maximum number of iterations. If so, output the optimal osprey position X best , and use it as the regularization coefficient C and kernel parameter γ of the KELM model, otherwise the number of iterations is increased by 1, and S5.2 to S5.4 are repeated. In S5.3, the position update strategy of the Osprey algorithm includes a positioning phase and a feeding phase:

[0073] In the positioning phase, for each osprey, the positions of other ospreys with better fitness values ​​in the search space are considered to be underwater fish. At this time, the position formula of the fish corresponding to each osprey is expressed as follows:

[0074] FP i ={X k |k∈{1,2,…,N3}∧F k <F i}∪{X best}

[0075] Where: FP i represents the location set of the fish of the i-th osprey, X k represents the position of the fish corresponding to the kth osprey, k represents the kth osprey, N3 represents the number of osprey populations, and F k and F i represent the objective function values ​​corresponding to the kth and ith ospreys, respectively, X best Indicates optimal Osprey location.

[0076] The osprey will randomly detect the position of one of the fish and attack it. Based on the simulation of the osprey's movement towards the fish, the osprey's position update formula is expressed as follows:

[0077]

[0078] Where: represents the new position of the i-th osprey in the j-th dimension in the first stage, x i,j represents the initial position of the i-th osprey in the j-th dimension in the first stage, r i,j Represents a random number in the interval [0,1], SF i,j represents the preferred fish of the i-th osprey in the j-th dimension, I i,jrepresents a random number in the set {1,2};

[0079]

[0080] Where: lb j represents the lower bound of the j-th dimension, ub j represents the upper bound of the j-th dimension;

[0081]

[0082] Where: X i represents the position of the i-th osprey after the update, represents the position of the i-th osprey in the first stage, F i P1 Indicates its corresponding fitness value, F i Represents the objective function value corresponding to the i-th osprey.

[0083] In the feeding phase, after the osprey catches the fish, it will take it to a suitable location for feeding. The osprey's position update formula is as follows:

[0084]

[0085] Where: represents the position of the i-th osprey in the j-th dimension after the second stage update, r represents a random number in the interval [0,1], t represents the number of iterations, N3 represents the number of osprey populations, m represents the maximum dimension, and T represents the maximum number of iterations;

[0086]

[0087] Where: represents the position of the i-th osprey in the second stage, F i P2 Represents the corresponding objective function value.

[0088] In S5.3, the optimal parameters of the IOOA-KELM prediction model are the updated regularization coefficient C and kernel parameter γ.

[0089] In step 6, the predicted values ​​of each subsequence are superimposed using the following formula:

[0090]

[0091] Where: i represents the number of subsequences, I represents the total number of subsequences, represents the predicted value corresponding to the i-th subsequence, represents the prediction result of the residual sequence, Represents the final prediction result obtained by superposition.

[0092] In step 7, the three evaluation indicators specifically include:

[0093]

[0094] Where: N4 represents the number of samples, y i , and They represent the actual value, predicted value and average value of the sample respectively.

[0095] In a second aspect, the present invention provides a wind power prediction system based on dominant component analysis and IOOA-KELM, the test system comprising:

[0096] A data acquisition module collects an original power data set of a wind turbine generator set, wherein the original data set includes instantaneous power data of the wind turbine generator set with a sampling period of 15 minutes;

[0097] The data preprocessing module decomposes the original power data set into multiple subsequences based on ICEEMDAN, and uses singular spectrum analysis SSA to extract the dominant components of each subsequence.

[0098] The model training module uses the subsequence after data processing as the input of the IOOA-KELM prediction model for training to obtain a wind power prediction model;

[0099] The prediction module inputs the subsequence of the wind turbine to be predicted that has undergone data preprocessing into the trained wind power prediction model for prediction.

[0100] The present invention provides a wind power prediction method and system based on dominant component analysis and IOOA-KELM, and the technical effects are as follows:

[0101] 1) The present invention adopts the dominant component analysis method, decomposes the original time series into multiple subsequences through ICEEMDAN, and uses SSA to extract the dominant component of each subsequence to reduce the non-stationarity of the sequence, thereby improving the prediction accuracy of the model, so it is suitable for wind power data preprocessing.

[0102] 2) The present invention introduces Logistic chaos initialization and Levy flight strategy into OOA. Compared with OOA, the improved Osprey optimization algorithm IOOA proposed in the present invention can not only increase the diversity of the initial population and improve the quality of the initial solution, but also reduce the probability of the OOA algorithm falling into the local optimum during the iteration process, thereby improving the local search accuracy.

[0103] 3) The present invention combines the improved Osprey optimization algorithm IOOA with the KELM model, uses IOOA to calibrate the hyperparameters of the KELM model, and obtains the optimal hyperparameter combination of the IOOA-KELM prediction model, so that the prediction performance of the KELM model reaches the optimal state, thereby improving the prediction accuracy of the KELM model. BRIEF DESCRIPTION OF THE DRAWINGS

[0104] Figure 1 is a flow chart of the present invention;

[0105] Figure 2 It is a data preprocessing result diagram in the present invention;

[0106] Figure 3 It is a graph of the subsequence prediction results of the present invention;

[0107] Figure 4 It is a comparison chart of the predicted value and the actual value using the present invention.

[0108] Figure 5 This is the KELM model structure diagram. DETAILED DESCRIPTION

[0109] The wind power prediction method based on dominant component analysis and IOOA-KELM, first, obtain the original time series data of wind power, decompose the time series of the original wind power by ICEEMDAN, reduce its non-stationarity, and obtain multiple subsequences; then use SSA to extract the dominant components of the obtained multiple subsequences and reconstruct the residuals of each subsequence by superposition to further reduce its non-stationarity; secondly, divide the data after the dominant component extraction into a test set, a training set and a validation set; further, on the basis of the traditional OOA algorithm, by introducing Logistic chaos initialization and Levy flight strategy, an improved Osprey optimization algorithm IOOA is obtained; a KELM model is constructed to train all subsequences respectively, and the hyperparameters of the KELM model are calibrated by IOOA; finally, all subsequences are predicted using the prediction model, and the prediction results of all subsequences are superimposed to obtain the final prediction result. The present invention adopts the proposed dominant component analysis method based on ICEEMDAN and SSA, and proposes an improved Osprey optimization algorithm IOOA on the basis of traditional OOA, and calibrates the hyperparameters of the KELM model. The method has strong data fitting ability and prediction performance.

[0110] The prediction process of the present invention is as follows Figure 1 As shown, the specific steps include:

[0111] Step a: Get the original wind power data and decompose it through ICEEMDAN. The specific process is as follows:

[0112] Step a1: Add adaptive white noise E1(w (i) ):

[0113] x (i) =x+β0E1(w (i) )

[0114] Where: x (i) represents the signal with white noise added for the i-th time, x(t) represents the original signal, and w (i) represents the i-th added white noise i∈{1,2,…,M}, M represents the number of added white noises; β0 represents the noise standard deviation of the first decomposition signal, E i (·) represents the operator for calculating the i-th EMD decomposition;

[0115] Step a2: Calculate each The local average of the signal and the residual r1 of the first decomposition are obtained:

[0116]

[0117] Where: <·> represents the operator for calculating the average value of the signal, M(·) represents the operator for calculating the local average value of the signal;

[0118] Step a3: Subtract the first residual r1 from the original signal x(t) to obtain the first component IMF1 of the original signal;

[0119] IMF1=x-r1

[0120] Step a4: Determine the second component IMF2:

[0121] r2= <M(r1+β1E2(w (i) ))>

[0122] IMF2=r1-r2

[0123] Where: r2 represents the residual of the second decomposition, β1 represents the noise standard deviation of the second decomposition signal, E2(w (i) ) represents the operator for calculating the second EMD decomposition, M(r1+β1E2(w (i) )) represents an operator for calculating the local average value of a signal;

[0124] Step a5: For k = 3, 4, ..., N1, where k is the current number of decompositions and N1 is the total number of decompositions, calculate the kth component IMF k :

[0125] r k = <M(r k-1 +β k-1 E k (w(i) ))>

[0126] IMF k = r k-1 -r k

[0127] where: r k represents the residual of the k-th decomposition, r k-1 represents the residual of the k-th decomposition, β k-1 represents the noise standard deviation of the second decomposition signal, E k (w (i) ) represents the operator for calculating the k-th EMD decomposition, M(r k-1 + β k-1 E k (w (i) )) represents the operator for calculating the local average value of the signal;

[0128] Step a6: k = k + 1, repeat a5 until the function r k = <M(r k-1 + β k-1 E k (w (i) ))> cannot be decomposed, and the adaptive noise complete ensemble empirical mode decomposition algorithm ends.

[0129] Step b: Use SSA to extract the dominant components of the multiple subsequences obtained by ICEEMDAN decomposition and stack and reconstruct the residuals of each subsequence. The preprocessed subsequences are as Figure 2 shown. The specific process is as follows:

[0130] Step b1: First, set the time series data {h i , i = 1, 2,..., N2}, h i represents the i-th data, N2 represents the total number of data, and construct a Hankel trajectory matrix;

[0131]

[0132] where: h1 represents the first data, h2 represents the second data, h3 represents the third data, h k represents the k-th data, h k+1 represents the (k + 1)-th data, h L represents the L-th data, h L+1 represents the (L + 1)-th data, represents the N2-th data, L represents the sliding window length parameter, and 1 < L < N2, k is N2 - L + 1; the matrix H is the trajectory matrix.

[0133] Step b2: Perform singular value decomposition (SVD) on the trajectory matrix. At this time, H can be rewritten as:

[0134]

[0135] Where: d = rank (H) represents the intrinsic component with non-zero eigenvalues, H i represents the elementary matrix, λ i represents the singular values ​​in descending order, U i and V i denote the relevant left and right eigenvectors, V i T represents the right eigenvector V i The transposed matrix of .

[0136] Step b3: Arrange the corresponding singular vectors in descending order according to the size of the singular values, extract the dominant components of the trajectory matrix H, that is, reconstruct the matrix H using the combination of the first r basic matrices:

[0137]

[0138] Where: represents the reconstructed signal, r represents the number of features used for signal reconstruction, the residual R represents the noise, and i represents the current i-th basic matrix.

[0139] Step c: Divide the data after the dominant component extraction into a test set, a training set, and a validation set in a ratio of 8:1:1.

[0140] Step d: By changing random initialization to Logistic chaos initialization and introducing Levy flight strategy when updating the local development position, the traditional Osprey optimization algorithm OOA is improved to obtain the improved Osprey optimization algorithm IOOA. The specific steps are as follows:

[0141] Step d1: Using chaotic sequences instead of pseudo-random numbers to initialize the population can increase the diversity of the initial population and improve the quality of the initial solution. Introducing Logistic chaos initialization can reduce the random fluctuations of population initialization, thereby improving the robustness of the algorithm. Assuming that the number of osprey populations in the algorithm is N3, the upper and lower bounds of the search space are ub and lb respectively, the initialization formula of the Logistic chaos mapping can be expressed as:

[0142] x i =α·x i-1 ·(1-x i-1 ),0≤x0≤1,i=1,2,...,N3

[0143] Where: x i represents the value of the i-th iteration, x i-1represents the value of the previous iteration, α∈[3.57,4], x0 represents the initial value, N3 represents the number of osprey populations, and the chaotic value x i The formula for mapping to the search space is:

[0144]

[0145] Where: Represents the initial value of population optimization.

[0146] Step d2: In order to reduce the probability of the OOA algorithm falling into the local optimum during the iteration process and improve the local search accuracy, the Levy flight strategy is introduced to perturb the population when updating the local development position to balance the global search and local development. The Levy flight formula is expressed as follows:

[0147]

[0148] In the formula: α represents a constant, which is generally 1.5, and x and y are two variables that obey the normal distribution.

[0149] The first stage search and fishing position update formula of the improved OOA algorithm is as follows:

[0150]

[0151] Where: represents the updated new position of the ith osprey in the jth dimension in the first stage, l represents Levy flight, x i.j represents the position of the i-th osprey in the j-th dimension before the update, SF ij represents the i-th fish selected in the j-th dimension, r i.j and I i.j They represent random numbers between 0 and 1 and between 1 and 2 respectively. The position update formulas for hunting and eating in the second stage are as follows:

[0152]

[0153] Where: represents the updated new position of the i-th osprey in the j-th dimension in the second stage, ub and lb represent the upper and lower bounds of the search space respectively, and r represents a random number between 0 and 1.

[0154] Step e: Construct the KELM model to predict all subsequences respectively, and use the proposed improved Osprey optimization algorithm IOOA to calibrate the KELM model hyperparameters to obtain the optimal parameters of the prediction model IOOA-KELM. The prediction results of each subsequence are as follows Figure 3 The specific steps are as follows:

[0155] Step e1: using the Logistic chaotic map in step d1 to initialize the population and set the algorithm parameters, and at the same time using the regularization coefficient C and the kernel parameter γ of the KELM model as the dimension of the osprey position in the improved osprey optimization algorithm 10A, that is, the dimension size is 2;

[0156] Step e2: Calculate the fitness value F of each osprey and determine the initial fish school position FP corresponding to the osprey;

[0157] Step e3: According to the position update strategy of the osprey at different stages, the position of each osprey is updated, and the updated position is used as the new regularization coefficient C and kernel parameter γ of KELM for verification, and the corresponding fitness value is calculated;

[0158] Step e3.1: The Osprey algorithm includes the positioning phase and the feeding phase:

[0159] In the positioning phase, for each osprey, the positions of other ospreys with better fitness values ​​in the search space are considered to be underwater fish. At this time, the fish position formula is expressed as follows:

[0160] FP i ={X k |k∈{1,2,…,N3}∧F k <F i}∪{X best}

[0161] Where: FP i represents the location set of the fish of the i-th osprey, X k represents the position of the fish corresponding to the kth osprey, k represents the kth osprey, N3 represents the number of osprey populations, and F k and F i represent the objective function values ​​corresponding to the kth and ith ospreys, respectively, X best indicates the best Osprey position;

[0162] The osprey will randomly detect the position of one of the fish and attack it. Based on the simulation of the osprey's movement towards the fish, the osprey's position update formula is expressed as follows:

[0163]

[0164] Where: represents the new position of the i-th osprey in the j-th dimension in the first stage, x i,j represents the initial position of the i-th osprey in the j-th dimension in the first stage, r i,j Represents a random number in the interval [0,1], SF i,j represents the preferred fish of the i-th osprey in the j-th dimension, I i,j represents a random number in the set {1,2},

[0165]

[0166] Where: lb j represents the lower bound of the j-th dimension, ub j represents the upper bound of the j-th dimension;

[0167]

[0168] Where: X i represents the position of the i-th osprey after update, X i P1 represents the position of the i-th osprey in the first stage, F i P1 Indicates its corresponding fitness value, F i Represents the objective function value corresponding to the i-th osprey.

[0169] In the feeding phase, after the osprey catches the fish, it will take it to a suitable location for feeding. The osprey's position update formula is as follows:

[0170]

[0171] Where: represents the position of the i-th osprey in the j-th dimension after the second stage update, r represents a random number in the interval [0,1], t represents the number of iterations, N3 represents the number of osprey populations, m represents the maximum dimension, and T represents the maximum number of iterations.

[0172]

[0173] Where: represents the position of the i-th osprey in the second stage, F i P2 Represents the corresponding objective function value.

[0174] Step e4: Compare the fitness of the current osprey position with the local optimal position of the osprey to obtain the optimal position X of the osprey best ;

[0175] Step e5: Determine whether the current number of iterations has reached the maximum number of iterations. If so, output the optimal osprey position X. best , and use it as the regularization coefficient C and kernel parameter γ of the KELM model, otherwise the number of iterations increases by 1, and steps e2 to e4 are repeated;

[0176] Step e6: Prediction is performed using the IHHO-KELM model after hyperparameter calibration to obtain the prediction results of each subsequence in the test set;

[0177] Step f: Superimpose the prediction results of all subsequences to obtain the final wind power prediction result.

[0178] Step g: The wind power data collected from a certain wind farm from January 8, 2022 to January 18, 2022 were selected as the research object for analysis. A total of 960 data were obtained with a sampling interval of 15 minutes. The root mean square error (RMSE), mean absolute error (MAE) and determination coefficient (R) were used to calculate the wind power data. 2 Three key indicators are used to analyze the results of the wind power prediction model. The three evaluation indicator formulas are as follows:

[0179]

[0180]

[0181] Where: N4 represents the number of samples, y i , and They represent the actual value, predicted value and average value of the sample respectively.

[0182] Example:

[0183] In order to verify the effectiveness of the model described in the present invention, the ICEEMDAN-SSA-KELM model, ICEEMDAN-KELM model and KELM model are used to model and predict the original data of wind power, and then the final prediction results are analyzed.

[0184] Figure 2 is a graph of data preprocessing results in the present invention; Figure 2 It can be seen that the original wind power data is processed by ICEEMDAN to obtain multiple subsequences. In addition, in order to further weaken the nonlinearity of the subsequence, the dominant components of the subsequence are extracted using the SSA algorithm, and the residual components are accumulated as the residual sequence for prediction to compensate for the prediction error, which can effectively weaken the non-stationarity of the data.

[0185] Figure 3 is a graph of the neutron sequence prediction results of the present invention; Figure 3 It can be seen that the optimized KELM model has a good prediction effect and fitting degree for each preprocessed subsequence. This shows that the proposed dominant component analysis method can effectively reduce the non-stationarity and nonlinearity of wind power series, thereby improving the prediction accuracy of the model.

[0186] The grid search method is used to optimize and determine the hyperparameters in each comparison model. The prediction results of each model are as follows: Figure 4 The corresponding evaluation indicators are shown in Table 1.

[0187] from Figure 4It can be seen that the wind power prediction curve of the ICEEMDAN-SSA-IOOA-KELM model proposed in the present invention is mostly fitted with the true value curve, which has strong data fitting ability and prediction performance, proving the effectiveness and superiority of the present invention.

[0188] Table 1 Error analysis evaluation indicators of different models

[0189]

[0190] It can be observed from Table 1 that the RMSE, MAE, and R 2 The RMSE, MAE, and R of the ICEEMDAN-KELM model are 1.8079, 1.5858, and 0.8903, respectively. The prediction performance is not very ideal. 2 are 1.0289, 0.9257, and 0.9870, respectively. Compared with the KELM model, the RMSE and MAE of the ICEEMDAN-KELM model decrease by 43.09% and 41.63%, respectively. 2 It increased by 10.86%, indicating that the original wind power data with non-stationary and nonlinear characteristics can reduce its volatility impact after being decomposed by ICEEMDAN, thereby improving the prediction accuracy of the model.

[0191] Compared with the ICEEMDAN-KELM model, the RMSE and MAE of the ICEEMDAN-SSA-KELM model decreased by 11.11% and 10.40%, respectively, and R 2 It increased by 0.36%, indicating that after ICEEMDAN decomposition, the dominant component analysis based on SSA can further reduce the non-stationarity of the data, thereby reducing the model prediction error. It also verifies the effectiveness of dominant component analysis in wind power prediction.

[0192] Compared with the ICEEMDAN-SSA-KELM model, the RMSE and MAE of the ICEEMDAN-SSA-IOOA-KELM model decreased by 11.43% and 11.13%, respectively, and R 2 It increased by 0.38%. This shows that the proposed IOOA algorithm has a good performance in optimizing the hyperparameters of the KELM model. In general, compared with other comparative models, the model proposed in the present invention achieves the best prediction performance.

Claims

1. A wind power prediction method based on dominant component analysis and IOOA-KELM is characterized by The following steps are involved: Step 1: Obtain the original wind power data, decompose the time series of wind power through ICEEMDAN, and obtain multiple subsequences; Step 2: Use singular spectrum analysis (SSA) to extract the dominant components of the multiple subsequences obtained after decomposition in step 1 and reconstruct the residuals of each subsequence by superimposing them; Step 3: Divide the data after extracting the dominant components into a test set, a training set, and a validation set; Step 4: By changing random initialization to Logistic chaos initialization and introducing Levy flight strategy when updating the local development position, the Osprey optimization algorithm OOA is improved to obtain the improved Osprey optimization algorithm IOOA; Step 5: Construct a KELM model, predict all subsequences separately, and use the improved Osprey optimization algorithm IOOA obtained in step 4 to calibrate the KELM model hyperparameters to obtain the optimal parameters of the IOOA-KELM prediction model; Step 6: Superimpose the prediction results of all subsequences to obtain the final wind power prediction result.

2. The wind power prediction method based on dominant component analysis and IOOA-KELM according to claim 1 is characterized by: Also includes step 7: using the root mean square error RMSE, mean absolute error MAE and determination coefficient R 2 Three key evaluation indicators are used to analyze the results of the wind power prediction model.

3. The wind power prediction method based on dominant component analysis and IOOA-KELM according to claim 1 is characterized by: The step 1 comprises the following steps: S1.1: Add adaptive white noise E1(w) to the original signal x(t) (i) ); x (i) =x(t)+β0E1(w (i) ) Where: x (i) represents the signal with white noise added for the i-th time; x(t) represents the original signal; w (i) represents the i-th added white noise i∈{1,2,…,M}, M represents the number of added white noises; β0 represents the noise standard deviation of the first decomposition signal, E i (·) represents the operator for calculating the i-th EMD decomposition; S1.2: Calculate each The local average of the signal and the residual r1 of the first decomposition are obtained: Where: <·> represents the operator for calculating the average value of the signal, M(·) represents the operator for calculating the local average value of the signal; S1.3: Subtract the residual r1 of the first decomposition from the original signal x(t) to obtain the first component IMF1 of the original signal x(t); IMF1=x-r1 S1.4: Determine the second component IMF2: r2= <M(r1+β1E2(w (i) )) IMF2=r1-r2 Where: r2 represents the residual of the second decomposition; β1 represents the noise standard deviation of the second decomposition signal; E2(w (i) ) represents the operator for calculating the second EMD decomposition; M(r1+β1E2(w (i) )) represents an operator for calculating the local average value of a signal; S1.5: For k = 3, 4, ..., N1, k represents the current number of decompositions, N1 represents the maximum number of decompositions; calculate the kth component IMF k : r k h <M(r k-1 +β k-1 E k (w (i) )) IMF k =r k-1 -r k Where: r k represents the residual of the kth decomposition, r k-1 represents the residual of the k-th decomposition, β k-1 represents the noise standard deviation of the second decomposition signal, E k (w (i) ) represents the operator for calculating the kth EMD decomposition, M(r k-1 +β k-1 E k (w (i) )) represents an operator for calculating the local average value of a signal; S1.6: k = k + 1, repeat S1.5 until function r k = <M(r k-1 +β k-1 E k (w (i) ))>Cannot be decomposed, the improved adaptive noise fully integrated empirical mode decomposition algorithm ends.

4. The wind power prediction method based on dominant component analysis and IOOA-KELM according to claim 1 is characterized by: The step 2 comprises the following steps: S2.1: First, suppose the time series data {h i ,i=1,2,…,N2},h i Represents the i-th data, N2 represents the total number of data, and constructs the Hankel trajectory matrix; Where: h1 represents the first data, h2 represents the second data, h3 represents the third data, h k represents the k-th data, h k+1 represents the (k + 1)-th data, h L represents the L-th data, h L+1 represents the (L + 1)-th data, represents the N2-th data, L represents the sliding window length parameter, and 1 < L < N2, k is N2 - L + 1; the matrix H is the trajectory matrix; S2.2: Perform singular value decomposition on the trajectory matrix. At this time, the matrix H can be rewritten as: Where: d = rank (H) represents the intrinsic component with non-zero eigenvalues, H i represents the elementary matrix, λ i represents the singular values ​​in descending order, U i and V i denote the relevant left and right eigenvectors, V i T represents the right eigenvector V i The transposed matrix of S2.3: According to the size of the singular value, the corresponding singular vectors are arranged in descending order to extract the dominant component of the trajectory matrix H, that is, the matrix H is reconstructed by combining the first r basic matrices: Where: represents the reconstructed signal, r represents the number of features used for signal reconstruction, the residual R represents the noise, and i represents the current i-th basic matrix.

5. The wind power prediction method based on dominant component analysis and IOOA-KELM according to claim 1 is characterized by: The step 4 comprises the following steps: Step S4.1: Introduce Logistic chaos initialization to reduce the random fluctuation of population initialization. Assume that the osprey population is N3, the upper and lower bounds of the search space are ub and lb respectively, and the initialization formula of the Logistic chaos mapping is expressed as: x i =α·x i-1 ·(1-x i-1 ),0≤x0≤1,i=1,2,...,N3 Where: x i represents the i-th iteration value, x i-1 represents the i-1th iteration value, α∈[3.57,4], x0 represents the initial value, N3 represents the number of osprey populations, and the chaotic value x i The formula for mapping to the search space is: Where: Represents the initial value of population optimization; Step S4.2: Introduce the Levy flight strategy to perturb the population when updating the local development position to balance global search and local development; the Levy flight formula is expressed as follows: In the formula: α represents a constant; x and y are two variables that obey the normal distribution; The first stage search and fishing position update formula of the improved OOA algorithm is as follows: Where: represents the updated new position of the ith osprey in the jth dimension in the first stage, l represents Levy flight, x i.j represents the position of the i-th osprey in the j-th dimension before the update, SF ij represents the i-th fish selected in the j-th dimension, r i.j and I i.j Respectively represent random numbers between the interval 0 and 1 and random numbers between the interval 1 and 2; The position update formula for the second stage of hunting and eating is as follows: Where: represents the updated new position of the i-th osprey in the j-th dimension in the second stage; ub and lb represent the upper and lower bounds of the search space respectively; r represents a random number between 0 and 1.

6. The wind power prediction method based on dominant component analysis and IOOA-KELM according to claim 1 is characterized by: In step 5, the KELM model predicts all subsequences separately. For each subsequence x i , the goal of KELM is to find a function mapping relationship to predict y i ; The calculation formula is as follows: h new =[K(x new1 ,x1),…,K(x newm ,x m )] Where: h new represents the kernel function value, x1 represents the first subsequence; x new1 and x newm Respectively represent the new input subsequences corresponding to the first subsequence and the mth subsequence; K(x new1 ,x1) represents the kernel function; represents the predicted value; β represents the output weight between the hidden layer and the output layer.

7. The wind power prediction method based on dominant component analysis and IOOA-KELM according to claim 6 is characterized by: The step 5 comprises the following steps: S5.1: using the Logistic chaotic map in step S4.1 to initialize the population, and using the regularization coefficient C and the kernel parameter γ of the KELM model as the dimension of the osprey position in the improved osprey optimization algorithm IOOA; in wind power prediction, the regularization coefficient C is used to control the complexity of the KELM model, and the kernel parameter γ determines the width of the radial basis function kernel in the KELM; S5.2: Calculate the fitness value F of each osprey, where the fitness value F represents the error between the wind power value predicted by the KELM model and the actual value, and use the root mean square error as a fitness evaluation index to determine the initial fish school position FP corresponding to the osprey; S5.3: According to the position update strategy of the osprey at different stages, the position of each osprey is updated, and the updated position is used as the new regularization coefficient C and kernel parameter γ of the KELM model for verification, and the corresponding fitness value is calculated; S5.4: Compare the fitness values ​​of the current osprey position and the local optimal position of the osprey to obtain the optimal position X of the osprey best ; S5.5: Determine whether the current number of iterations has reached the maximum number of iterations. If so, output the optimal osprey position X best , and use it as the regularization coefficient C and kernel parameter γ of the KELM model, otherwise the number of iterations is increased by 1, and S5.2 to S5.4 are repeated.

8. The wind power prediction method based on dominant component analysis and IOOA-KELM according to claim 7 is characterized by: In S5.3, the position update strategy of the Osprey algorithm includes a positioning phase and a feeding phase: In the positioning stage, for each osprey, the positions of other ospreys with better fitness values ​​in the search space are considered to be underwater fish; at this time, the position formula of the fish corresponding to each osprey is expressed as follows: FP i ={X k |k∈{1,2,…,N3}∧F k <F i }∪{X best } Where: FP i represents the location set of the fish of the i-th osprey, X k represents the position of the fish corresponding to the kth osprey, k represents the kth osprey, N3 represents the number of osprey populations, and F k and F i represent the objective function values ​​corresponding to the kth and ith ospreys, respectively, X best indicates the best Osprey position; The osprey will randomly detect the position of one of the fish and attack it. Based on the simulation of the osprey's movement towards the fish, the osprey's position update formula is expressed as follows: Where: represents the new position of the i-th osprey in the j-th dimension in the first stage, x i,j represents the initial position of the i-th osprey in the j-th dimension in the first stage, r i,j Represents a random number in the interval [0,1], SF i,j represents the preferred fish of the i-th osprey in the j-th dimension, I i,j represents a random number in the set {1,2}; Where: lb j represents the lower bound of the j-th dimension, ub j represents the upper bound of the j-th dimension; Where: X i represents the position of the i-th osprey after the update, represents the position of the i-th osprey in the first stage, F i P1 Indicates its corresponding fitness value, F i represents the objective function value corresponding to the i-th osprey; In the feeding phase, after the osprey catches the fish, it will take it to a suitable location for feeding. The osprey's position update formula is as follows: Where: represents the position of the i-th osprey in the j-th dimension after the second stage update, r represents a random number in the interval [0,1], t represents the number of iterations, N3 represents the number of osprey populations, m represents the maximum dimension, and T represents the maximum number of iterations; Where: represents the position of the i-th osprey in the second stage, F i P2 Represents the corresponding objective function value.

9. The wind power prediction method based on dominant component analysis and IOOA-KELM according to claim 1 is characterized by: In step 6, the predicted values ​​of each subsequence are superimposed using the following formula: Where: i represents the number of subsequences, I represents the total number of subsequences, represents the predicted value corresponding to the i-th subsequence, represents the prediction result of the residual sequence, Represents the final prediction result obtained by superposition.

10. A wind power prediction system based on dominant component analysis and IOOA-KELM is characterized by: The test system includes: A data acquisition module collects an original power data set of a wind turbine generator set, wherein the original data set includes instantaneous power data of the wind turbine generator set with a sampling period of 15 minutes; The data preprocessing module decomposes the original power data set into multiple subsequences based on ICEEMDAN, and uses singular spectrum analysis SSA to extract the dominant components of each subsequence; The model training module uses the subsequence after data processing as the input of the IOOA-KELM prediction model for training to obtain a wind power prediction model; The prediction module inputs the subsequence of the wind turbine to be predicted that has undergone data preprocessing into the trained wind power prediction model for prediction.

Images