Wind power prediction method with directional focus attention mechanism

Abstract

The invention relates to a wind power prediction method with a directional focus attention mechanism, and the method is technically characterized in that data are preprocessed through principal component analysis, variable redundancy is reduced, and the efficiency and precision are improved; proposing an EDS-PKO algorithm combined with a dynamic strategy, wherein the EDS-PKO algorithm comprises dynamic parameter adjustment and a two-stage updating strategy; dFAT is introduced into the LSTM network, and attention weights of different time steps are dynamically adjusted; compared with PSO, WOA and GWO algorithms, the EDS-PKO algorithm has the advantages that RMSE values of the optimized LSTM-DFAT model are increased by 42.59%, 46.55% and 16.21% respectively, R2 values are increased by 16.04%, 18.59% and 6.96% respectively, MAPE values are increased by 3.97%, 4.41% and 1.36% respectively, and superiority is shown; a new means is provided for wind power prediction, and efficient utilization of wind energy is facilitated.

Description

Technical Field

[0001] The present invention belongs to the field of new energy output prediction, and in particular to a wind power prediction method with a directional focus attention mechanism; Background Art

[0002] Under the background of the global energy structure transformation and sustainable development strategy, wind energy, as a clean and renewable energy form, has become an important way for countries to alleviate energy shortages, reduce greenhouse gas emissions, and promote the development of green and low-carbon economy; Wind power prediction, as one of the key technologies for efficient grid connection of wind power and power grid dispatching, has inestimable value for improving the flexibility, stability, and economic benefits of power system operation; With the rapid progress of technologies such as big data, artificial intelligence, and cloud computing, the accuracy and efficiency of wind power prediction have been significantly improved; However, due to the inherent randomness and intermittency of wind power itself, wind power prediction still faces many challenges; In addition, under the condition of large-scale wind power grid connection, uncertainty and randomness also pose a threat to the stable operation of the power system; Therefore, constructing an efficient and accurate wind power prediction model is a necessary way to promote the effective utilization of clean wind energy resources and achieve the sustainable development of the power industry;

[0003] Previous studies have made significant progress in the field of wind power prediction; However, it still faces several challenges; First, although most wind power prediction models have made progress in dealing with non-linear data, the accuracy of predicting highly non-linear and dynamically changing wind power data is still affected; Second, the randomness of model parameters in traditional methods not only affects their adaptability to different data sets, but also weakens the generalization ability of the model, seriously affecting the stability of the model in practical applications; Finally, although the attention mechanism can explore more representative features in the data, when facing highly uncertain, strongly seasonal, and randomly fluctuating wind power data, they may ignore key information, thus possibly reducing the performance of the model when combining the attention mechanism; Therefore, in order to address these challenges, it is necessary to propose a directional focus attention enhanced deep learning network, aiming to improve prediction accuracy, reduce prediction errors, and provide more reliable data support for power grid dispatching; Summary of the Invention

[0004] The purpose of the present invention is to overcome the deficiencies of the prior art, use the principal component analysis method to reduce the data dimension of the wind power data set, enhance the efficiency and accuracy of the prediction model, retain information to the greatest extent, propose a directional focus attention enhanced deep learning network, improve prediction accuracy, reduce prediction errors, propose a variegated kingfisher algorithm based on an enhanced dynamic strategy and a directional focus attention mechanism for constructing a dynamic focusing mechanism, realize extremely small prediction errors under bad weather conditions, and be able to extract key information for wind power prediction under bad weather conditions;

[0005] The present invention solves its technical problems by adopting the following technical solutions:

[0006] A wind power prediction method with a directional focus attention mechanism, comprising the following steps:

[0007] Step 1, perform data dimensionality reduction processing on the original wind power meteorological data by using the principal component analysis method (PCA); extract the key feature information of the wind power time series;

[0008] The principle of PCA is as follows:

[0009] Assume that there is a data set X with n samples and m features:

[0010]

[0011] In the formula, X ∈ R n×m , the original variable index is x 1 , x 2 , …, x m , the new variable index after conversion is d 1 , d 2 , …, d t (t ≤ m), p nm represents the sample data set;

[0012] The new variable is usually expressed as a linear combination of the original variables:

[0013]

[0014] In the formula, d i is called the principal component of the original variable index x 1 , x 2 , …, x m , d ij is the coefficient of the original index variable x j (j = 1, 2, …, m) on the principal component d i (i = 1, 2,..., n);

[0015] The specific steps of PCA are as follows:

[0016] 1) Calculate the correlation matrix R = (r jk ) p×p :

[0017]

[0018] Suppose there is a data set, which constitutes a matrix with j rows and k columns, where represents the average value of the j - row samples; represents the average value of the k - column samples; r jk is the original variable x jand x k The correlation coefficient between

[0019] 2) Solve the characteristic equation |R-λE|=0 (where E is the same matrix), get the initial value, and arrange it in descending order as λ 1 ≥λ 2 ≥…≥λ p ≥0;

[0020] 3) Calculate the eigenvalue λ i The corresponding eigenvector

[0021]

[0022] In PCA, the selection of principal components is based on the contribution ratio ACR:

[0023]

[0024] Among them, the value of m (m≤p) corresponding to ACR≥80% was considered as the number of principal components;

[0025] Step 2: Introduce DFAT into the LSTM network to dynamically adjust the attention weights at different time steps to ensure that the model focuses on key information in a complex and changing environment.

[0026] The principle of LSTM is as follows:

[0027] F t =sigmoid(A f x t +B f h t-1 +b f ) (7)

[0028] I t =sigmoid(A i x t +B i h t-1 +b i ) (8)

[0029] C t =F t ×C t-1 +I t ×tanh(A c x t +B c h t-1 +b c ) (9)

[0030] O t =sigmoid(A o xt +B o h t-1 +b o ) (10)

[0031] h t =O t tanh(C t ) (11)

[0032] In the formula, F t , I t , O t represent the forget gate, input gate and output gate respectively; x t represents the input data of the LSTM at time t; h t-1 represents the hidden state at time t-1; A f , B f , A i , B i , Ac, B c , A o , B o are the weight parameters of the corresponding memory state gates; b i , b c , b 0 are the biases of the corresponding memory state gates; sigmoid and tanh are activation functions;

[0033] The LSTM is combined with the directional focus attention mechanism (DFAT). The attention mechanism calculates the similarity between the query vector and each element in the input vector, and obtains the attention weights of different elements through normalization; as shown in formulas (12), (13), and (14):

[0034]

[0035] Y = W·H (14)

[0036] In the formula, h i represents the i th -th element of the output vector of the LSTM network; Q represents the query vector; W i represents the attention weight of the i th -th element; W represents the attention weight vector; H represents the LSTM output vector; Y represents the output vector of the attention layer;

[0037] The mathematical model of the directional focus gating control unit is as follows:

[0038] D i = sigmoid(A i (1 - W i ) + b i ) if var(W i > T cir) (15)

[0039] In the formula, D i represents the updated directional focusing attention weight; A i and b i are the weight parameter and bias of the directional focusing gate; T cir is the threshold for determining bad weather conditions determined according to historical wind power generation data (when the variance of the attention score exceeds this threshold, it is regarded as bad weather);

[0040] Step 3: Based on the original PKO algorithm, introduce dynamic parameter adjustment, introduce a flight factor, and a two-stage update strategy;

[0041] The basic strategy of the PKO algorithm:

[0042] X i (t + 1) = X i (t) + α * T × (X j (t) - X i (t)), α ∈ 2 * randn(1, dim) (16)

[0043] In the formula, N is the total number of the population; i and j are natural numbers between 1 and N and not equal to each other; randn is a random number following a normal distribution; Dim represents the dimension of the problem considered;

[0044] The calculation method of the T parameter in the perching strategy is as follows:

[0045]

[0046] Crest = 2 * pi * rand (18)

[0047] In the formula, Max_iter is the maximum number of iterations; rand is a random value between 0 and 1;

[0048] The calculation method of the parameter T in the hovering strategy is as follows:

[0049]

[0050] In the formula, PKO_Fitness represents the physical fitness value of the kingfisher, BF (BF = 8) is the bounce factor;

[0051]

[0052] In the formula, is the best position in the current group; r and s are control parameters; randn(0, 1) is a random number following a normal distribution with a mean of 0 and a standard deviation of 1;

[0053]

[0054] Wherein, BF is the beating coefficient; Max_iter is the maximum number of iterations;

[0055]

[0056] Wherein, w is the flight factor;

[0057] In the exploration stage, individuals mainly conduct extensive searches to discover more potential optimal solutions:

[0058]

[0059] In the development stage, individuals mainly conduct local searches to improve the accuracy of the current solution:

[0060]

[0061] Step 4, initialize the EDS-PKO-LSTM-DFAT model parameters, including the EDS-PKO algorithm parameters and the LSTM-DFAT model parameters;

[0062] Step 5, adopt the fitness R between the actual wind power value and the predicted wind power value 2 Define the fitness function of the EDS-PKO algorithm;

[0063] Step 6, adopt the enhanced dynamic strategy-based kingfisher optimizer (EDS-PKO) to optimize the model parameters and obtain the optimal hyperparameter configuration;

[0064] Steps for the optimal hyperparameter configuration:

[0065] 1) Take the selected Z principal component equations as the multi-objective functions of the improved EDS-PKO algorithm, and use the Pareto front to solve to obtain S excellent solution sets common to the Z objective functions;

[0066] 2) Standardize the decision layer data;

[0067] 3) Define the distance between the i-th standardized data and the maximum standardized data as, and define the distance between the i-th standardized data and the minimum standardized data as; Calculate the score of the i-th data before non-standardization;

[0068] 4) Standardize the scores and sort them from largest to smallest; The score with the largest value is used as the optimal solution of the Pareto front and the optimal hyperparameter configuration;

[0069] Step 7, based on the optimal hyperparameter configuration, use the LSTM-DFAT model to predict the wind power of the test set and obtain the wind power prediction result;

[0070] The above-mentioned wind power prediction method with a directional focus attention mechanism, the principal component analysis method, the long short-term memory neural network, and the spotted kingfisher algorithm are existing methods;

[0071] The above-mentioned wind power prediction method with a directional focus attention mechanism, the wind power, etc. are well-known to those skilled in the art;

[0072] The advantages and positive effects of the present invention are:

[0073] 1. A composite model combining an evolutionary algorithm, a directional focus attention mechanism (DFAT), and deep learning is constructed, which can extract key feature information from time series under highly nonlinear conditions;

[0074] 2. The spotted kingfisher algorithm based on an enhanced dynamic strategy (EDS-PKO) is proposed, which integrates a dynamic parameter adjustment mechanism and a two-stage search mode;

[0075] 3. A dynamic focus mechanism is innovatively constructed in the attention mechanism, and a directional focus attention mechanism is introduced; the attention scores of different elements are changed through a directional focus gate unit to ensure focusing on the most critical information in various environments; BRIEF DESCRIPTION OF THE DRAWINGS

[0076] The present invention will be further described below in conjunction with the drawings and embodiments;

[0077] Figure 1 It is a schematic diagram of the short-term wind power prediction process based on the EDS-PKO-LSTM-DFAT model;

[0078] Figure 2 It is a graph of the wind power prediction results of each model under three normal weather conditions;

[0079] Figure 3 It is a graph of the prediction results of each prediction model under bad weather; DETAILED DESCRIPTION OF THE INVENTION

[0080] Embodiment

[0081] In order to improve the efficiency of wind power prediction, the present invention proposes a wind power prediction method with a directional focus attention mechanism, uses PCA to preprocess data, uses DFAT to enhance key information recognition, uses EDS-PKO to optimize parameters, and integrates an LSTM network to capture time features. The short-term wind power prediction process based on the EDS-PKO-LSTM-DFAT model is as Figure 1 shown;

[0082] Step 1, perform dimensionality reduction processing on the original wind power meteorology; extract key feature information;

[0083] Step 2, let the model focus on key information in a complex and ever-changing environment;

[0084] Step 3, introduce dynamic parameter adjustment, introduce flight factors and update strategies;

[0085] Step 4, initialize the model parameters;

[0086] Step 5, define the fitness function of the algorithm using the fitness between the actual wind power value and the predicted wind power value;

[0087] Step 6, optimize the model parameters based on the enhanced dynamic strategy algorithm to obtain the optimal hyperparameter configuration;

[0088] Step 7, based on the optimal hyperparameter configuration, perform wind power prediction on the test set to obtain the wind power prediction result;

[0089] Step 1, perform data dimensionality reduction processing on the original wind power meteorological data using the principal component analysis method (PCA); extract the key feature information of the wind power time series;

[0090] Use the principal component analysis method to perform dimensionality reduction processing on the meteorological data. First, assume that there is a data set (1) with N samples and M features, and the linear combination of the original variables is expressed as new variables (2); calculate the correlation matrix using equations (3)(4), solve the characteristic equation to obtain the initial values and arrange them in descending order, calculate the eigenvalues and eigenvectors using equations (4)(5), calculate the contribution rate of each principal component using equation (6), and select the principal components with a contribution rate greater than 80% as the number of principal components. The principal component analysis method maximally retains the information while reducing the data dimensionality;

[0091] Step 2, introduce DFAT into the LSTM network, and ensure that the model focuses on key information in a complex and ever-changing environment by dynamically adjusting the attention weights of different time steps;

[0092] Establish the LSTM network formula. Equations (7)(8)(9)(10)(11) represent that LSTM introduces the concepts of forget gate, input gate, output gate, and cell state on the basis of long short-term memory. It continuously updates the "long-term memory", that is, the cell state CT, through the forget gate and the input gate; the output gate integrates the information from the current time step and the "short-term memory" from the previous time step, that is, the hidden state HT-1, to calculate the output HT of the LSTM network at the current time step; equations (12)(13)(14) represent that the attention mechanism calculates the similarity between the query vector and each element in the input vector, and obtains the attention weights of different elements through normalization; establish the directional focusing gating control unit using equation (15);

[0093] Step 3: Based on the original PKO algorithm, introduce dynamic parameter adjustment, introduce a flight factor, and a two-stage update strategy;

[0094] First, establish the PKO algorithm formula. Use Equation (16) to randomly generate the initial population and distribute it in the search space to increase the diversity of the population and ensure the possibility of finding the global optimal solution. Use Equations (17), (18), (19), and (20) to establish the roosting strategy, and use Equation (21) to establish the update formula for the position of each individual;

[0095] Then, introduce dynamic parameter adjustment, introduce a flight factor, and a two-stage update strategy. Use Equation (22) to adjust the hunting strategy of the pied kingfisher according to different environments and prey, and use Equation (23) to adjust the speed and direction flight trajectory of the pied kingfisher when hunting in flight. To improve the stability and diversity of the algorithm, we introduce a two-stage update strategy and adopt different update methods in the exploration and exploitation stages. Use Equation (24) to represent that an individual conducts extensive searches in the exploration stage to discover more potential optimal solutions, and use Equation (25) to represent local searches in the individual development stage to improve the accuracy of the current solution;

[0096] Step 4: Initialize the parameters of the EDS-PKO-LSTM-DFAT model, including the parameters of the EDS-PKO algorithm and the LSTM-DFAT model

[0097] In this embodiment, the initialized parameters of the EDS-PKO algorithm and the LSTM-DFAT model are shown in Table 1;

[0098] Table 1 Parameter settings

[0099]

[0100] Step 5: Use the goodness-of-fit R between the actual wind power value and the predicted wind power value 2 Define the fitness function of the EDS-PKO algorithm

[0101] Step 6: Use the enhanced dynamic strategy-based pied kingfisher optimizer (EDS-PKO) to optimize the model parameters and obtain the optimal hyperparameter configuration;

[0102] Steps for the optimal hyperparameter configuration:

[0103] 1) Use the selected Z principal component equations as the multi-objective functions of the improved EDS-PKO algorithm and solve them using the Pareto front to obtain S excellent solution sets common to the Z objective functions;

[0104] 2) Standardize the decision layer data;

[0105] 3) Define the distance between the i-th normalized data and the maximum normalized data, and define the distance between the i-th normalized data and the minimum normalized data; calculate the score of the i-th data before normalization;

[0106] 4) Normalize the scores and sort them from largest to smallest; the score with the largest value is used as the optimal solution of the Pareto frontier and the optimal hyperparameter configuration;

[0107] Step 7, based on the optimal hyperparameter configuration, use the LSTM-DFAT model to predict the wind power of the test set to obtain the wind power prediction result;

[0108] In this embodiment, CNN, RNN, LSTM, and LSTM-Attention are selected as comparison models to evaluate the superiority of the EDS-PKO-LSTM-DFAT model in dealing with conventional meteorological conditions; all models use the same historical data length, and the wind power generation data of the past 9 days are used as input features to predict the wind power generation of the next day;

[0109] The specific implementation steps are as follows:

[0110] 1) Train and test each model to ensure that the training set and test set are divided according to the time series to avoid problems related to future information leakage;

[0111] 2) Select the root mean square error (RMSE), R-variance (R 2 ) and mean absolute percentage error (MAPE) as key indicators to comprehensively evaluate the prediction accuracy of the model;

[0112] 3) Analyze each model to determine its prediction accuracy;

[0113] The parameter settings of each prediction model are shown in Table 2;

[0114] Table 2 Parameters of the prediction model

[0115]

[0116]

[0117] Combined with the prediction model parameters, predict the wind power of each model under three normal weather conditions, and the prediction results of each model are as Figure 2 shown, Figure 2(a), (b), and (c) respectively show the comparison of wind power prediction results of multiple models under different meteorological conditions; through comparative analysis, compared with other models, the prediction curve generated by the DFAT-LSTM model shows a higher degree of coincidence with the actual wind power curves under various meteorological conditions; this result indicates that the DFAT-LSTM model demonstrates more excellent prediction performance in wind power prediction; specifically, by introducing deep learning algorithms and specific data processing techniques, this model realizes the accurate capture and efficient analysis of wind power data, thus still being able to maintain a high prediction accuracy under complex and changeable meteorological conditions;

[0118] The evaluation results of the prediction performance of each model, specifically manifested as prediction errors, are detailed in Table 3;

[0119] Table 3 Model prediction error results

[0120]

[0121] Through in-depth analysis of the data in Table 3, we can clearly observe that compared with traditional models, the DFAT-LSTM model shows significantly improved prediction accuracy under various different weather conditions; specifically, under the three typical normal weather conditions of sunny, cloudy, and light wind, the DFAT-LSTM model shows excellent adaptability; although under sunny weather conditions, its R 2 value is slightly lower than that of the LSTM-Attention model, but in all other prediction performance indicators, the DFAT-LSTM model comprehensively surpasses traditional models; particularly, the root mean square error (RMSE) of the DFAT-LSTM model always remains below 0.8, the R 2 value exceeds 90%, and the mean absolute percentage error (MAPE) is lower than 7%, and these indicators are significantly better than traditional models such as RNN, LSTM, and CNN; this result not only strongly verifies the high effectiveness and significant advantages of the DFAT-LSTM model in predicting wind power generation under normal weather conditions, especially in improving prediction accuracy and enhancing generalization ability, but also provides a more accurate and reliable decision-making basis for power grid scheduling and energy management, with important practical application value;

[0122] Under the optimal hyperparameters found by EDS search, the LSTM-DFAT model is used to predict wind power for the test set, and the parameter settings of each algorithm are shown in Table 4;

[0123] Table 4 Algorithm parameter settings

[0124]

[0125] The results predicted by the model are as Figure 2As shown, from Figure 2 it can be seen that under the optimal hyperparameters searched and determined by the EDS algorithm, the prediction results of the LSTM-DFAT model show a high degree of consistency with the actual wind power generation curve. This result fully demonstrates the accuracy and effectiveness of the EDS-PKO-LSTM-DFAT model in predicting wind power generation when dealing with adverse weather conditions;

[0126] Table 5 Comparison of Evaluation Metrics

[0127]

[0128] As can be seen from Table 5, compared with other prediction models, the EDS-PKO-LSTM-DFAT prediction model proposed in this paper shows excellent performance in various evaluation metrics; in terms of the R 2 metric, the EDS-PKO-LSTM-DFAT model has achieved significant improvements of 16.04%, 18.59% and 6.96% compared with the comparison models respectively; while in terms of the root mean square error (RMSE) metric, the model has achieved significant improvements of 42.59%, 46.55% and 16.21% respectively; especially under adverse weather conditions, the mean absolute percentage error (MAPE) of the EDS-PKO-LSTM-DFAT model is as low as 4.55%, which indicates that the prediction error of the model is extremely small under adverse weather conditions and has extremely high prediction accuracy; the model can accurately capture and extract the key information of wind power prediction under adverse weather conditions through the combination of effective feature extraction and deep learning algorithms, providing strong technical support for accurate prediction and efficient management in the field of wind power generation;

[0129] In the above embodiments, the principal component analysis method, the long short-term memory neural network, and the spotted kingfisher algorithm are existing methods;

[0130] It should be emphasized that the embodiments described in the present invention are illustrative. Therefore, the present invention is not limited to the embodiments described in the specific implementation manners. Any other implementation manners obtained by those skilled in the art according to the technical solutions of the present invention also belong to the scope of protection of the present invention.

Claims

1. A wind power forecasting method with a directional focus attention mechanism, characterized in that The following steps are involved: Step 1: reduce the dimension of the original wind power meteorology and extract key feature information; Step 2: Let the model focus on key information in a complex and changing environment; Step 3, introduce dynamic parameter adjustment, flight factor and update strategy; Step 4, initialize model parameters; Step 5, defining the fitness function of the algorithm using the degree of fit between the actual wind power value and the predicted wind power value; Step 6: Optimize model parameters based on the enhanced dynamic strategy algorithm to obtain the optimal hyperparameter configuration; Step 7: Based on the optimal hyperparameter configuration, the wind power prediction is performed on the test set to obtain the wind power prediction result; Step 1: Use principal component analysis (PCA) to reduce the dimension of the original wind power meteorological data; extract the key feature information of the wind power time series; The principle of PCA is as follows: Suppose there is a dataset X with n samples and m features: Where X∈R n×m , the original variable indices are x1,x2,…,x m , the new variable index after conversion is d1, d2, …, d t (t≤m), p nm represents a sample dataset; New variables are often expressed as linear combinations of the original variables: Where d i are called primitive variables with indices x1, x2, …, x m The principal components of ij is the original index variable x j (j=1,2,…,m) in the principal component d i coefficients on (i=1,2,...,n); The specific steps of PCA are as follows: 1) Calculate the correlation matrix R = (r jk ) p×p : Suppose there is a data set that forms a matrix with j rows and k columns, where Represents the average value of the j-row samples; represents the average value of k columns of samples; r jk is the original variable x j and x k The correlation coefficient between 2) Solve the characteristic equation |R-λE|=0 (where E is the same matrix), get the initial value, and arrange it in descending order as λ1≥λ2≥…≥λ p ≥0; 3) Calculate the eigenvalue λ i The corresponding eigenvector In PCA, the selection of principal components is based on the contribution ratio ACR: Among them, the value of m (m≤p) corresponding to ACR≥80% was considered as the number of principal components; Step 2: Introduce DFAT into the LSTM network to dynamically adjust the attention weights at different time steps to ensure that the model focuses on key information in a complex and changing environment. The principle of LSTM is as follows: F t =sigmoid(A f x t +B f h t-1 +b f ) (7) I t =sigmoid(A i x t +B i h t-1 +b i ) (8) C t =F t ×C t-1 +I t ×tanh(A c x t +B c h t-1 +b c ) (9) O t =sigmoid(A o x t +B o h t-1 +b o ) (10) h t =O t fishy(C) t ) (11) In the formula, F t ,I t ,O t Respectively represent the forget gate, input gate and output gate; x t represents the input data of LSTM at time t; h t-1 represents the hidden state at time t-1; A f ,B f ,A i ,B i ,Ac,B c ,A o ,B o is the weight parameter of the corresponding storage state gate; b i ,b c , b0 is the bias of the corresponding storage state gate; sigmoid and tanh are activation functions; LSTM is combined with the Directed Focus Attention Mechanism (DFAT). The attention mechanism calculates the similarity between the query vector and each element in the input vector, and obtains the attention weights of different elements through normalization; as shown in formulas (12), (13), and (14): Y=W·H (14) In the formula, h i Represents the i-th output vector of the LSTM network th elements; Q represents the query vector; W i Represents the i th The attention weight of each element; W represents the attention weight vector; H represents the LSTM output vector; Y represents the output vector of the attention layer; The mathematical model of the directional focusing gating control unit is as follows: D i =sigmoid(A i (1-W i )+b i )if var(W i >T cir ) (15) Where D i Represents the updated directional focus attention weight; A i and b i are the weight parameters and bias of the directional focusing gate; T cir is the threshold for determining severe weather conditions based on historical wind power generation data (when the variance of the attention score exceeds this threshold, it is considered severe weather); Step 3, based on the original PKO algorithm, dynamic parameter adjustment, flight factor and two-stage update strategy are introduced; The basic strategy of the PKO algorithm: X i (t+1)=X i (t)+α*T×(X j (t)-X i (t)),α∈2*randn(1,dim) (16) In the formula, N is the total population; i and j are natural numbers between 1 and N and are not equal to each other; randn is a random number following a normal distribution; Dim represents the dimension of the problem under consideration; The calculation method of T parameter in the habitat strategy is as follows: Crest=2*pi*rand (18) Where Max_iter is the maximum number of iterations; rand is a random value between 0 and 1; The calculation method of parameter T in the hover strategy is as follows: Where PKO_Fitness represents the physical fitness of the kingfisher, and BF (BF = 8) is the bounce factor; In the formula, is the best position in the current population; r and s are control parameters; randn(0,1) is a random number that follows a normal distribution with a mean of 0 and a standard deviation of 1; Where BF is the beating coefficient; Max_iter is the maximum number of iterations; Where w is the flight factor; In the exploration phase, individuals primarily conduct extensive searches to discover more potential optimal solutions: During the development phase, individuals primarily conduct local searches to improve the accuracy of the current solution: Step 4, initialize the EDS-PKO-LSTM-DFAT model parameters, including EDS-PKO algorithm parameters and LSTM-DFAT model parameters; Step 5: Use the fit R between the actual wind power value and the predicted wind power value 2 Define the fitness function of the EDS-PKO algorithm; Step 6, using the enhanced dynamic strategy-based Pied Kingfisher Optimizer (EDS-PKO) to optimize the model parameters and obtain the optimal hyperparameter configuration; Optimal hyperparameter configuration steps: 1) The selected Z principal component equations are used as the multi-objective functions of the improved EDS-PKO algorithm, and the Pareto preamble is used to solve them to obtain the S excellent solution sets common to the Z objective functions; 2) Standardize the decision-making data; 3) Define the distance between the i-th standardized data and the maximum standardized data as, define the distance between the i-th standardized data and the minimum standardized data as; calculate the score of the i-th data after unstandardization; 4) The scores are standardized and sorted from large to small; the score with the largest value is taken as the optimal solution of the Pareto predicate and the optimal hyperparameter configuration; Step 7: Based on the optimal hyperparameter configuration, the LSTM-DFAT model is used to predict the wind power of the test set to obtain the wind power prediction result.

2. A wind power forecasting method with a directional focus attention mechanism according to claim 1, characterized in that: The optimal hyperparameter configuration steps are as follows: 1) The selected Z principal component equations are used as the multi-objective functions of the improved EDS-PKO algorithm, and the Pareto preamble is used to solve them to obtain the S excellent solution sets common to the Z objective functions; 2) Standardize the decision-making data; 3) Define the distance between the ith standardized data and the maximum standardized data as, and define the distance between the ith standardized data and the minimum standardized data as; Calculate the score of the i-th data after it is not standardized; 4) The scores are standardized and sorted from large to small; the score with the largest value is taken as the optimal solution of the Pareto predicate and the optimal hyperparameter configuration.

Images