VMD-PCA-LSTM ultra-short-term wind power prediction method based on pelicans optimization algorithm

Abstract

The invention discloses a pelican optimization algorithm-based VMD-PCA-LSTM ultra-short-term wind power prediction method. The pelican optimization algorithm-based VMD-PCA-LSTM ultra-short-term wind power prediction method comprises a data preprocessing module, a power sequence modal decomposition module, an optimization algorithm module and a prediction model module. Firstly, abnormal data of a wind power plant are cleaned through a quartile method, missing data are interpolated, an energy difference method is adopted to select a VMD decomposition sub-modal number, and an original power sequence of complex fluctuation is decomposed; secondly, principal component extraction is carried out on high-dimensional meteorological data by adopting PCA, so that the mining difficulty of feature data is reduced; and finally, introducing a pelican optimization algorithm to carry out automatic optimization on four-dimensional hyper-parameters of a Dropout discarding rate, a learning rate, a first hidden layer unit number and a second hidden layer unit number of the LSTM prediction model so as to accelerate model convergence and improve fitting precision. The pelican optimization algorithm is simple in structure and easy to implement, has good expandability, and can optimize parameters of the LSTM model, so that the precision of wind power prediction is effectively improved.

Description

Technical Field

[0001] The present invention belongs to the field of ultra-short-term wind power prediction, and relates to a VMD-PCA-LSTM ultra-short-term wind power prediction method based on the Pelican optimization algorithm. Background Art

[0002] With the development of economy, the global energy demand is increasing. The use of traditional energy produces a large amount of greenhouse gases and air pollutants, which accelerates climate change and environmental damage. Wind energy, as a clean, low-carbon, renewable and widely distributed energy, has been widely recognized and applied. However, due to the volatility and intermittency of wind power generation, it poses a serious hidden danger to the safe operation of the power system. Accurate wind power forecasting can help grid dispatchers arrange spinning reserve capacity in advance and make dispatching decisions.

[0003] Wind power prediction models are mainly divided into physical models and data-based statistical models. The physical model is constructed by meteorological and geographical parameters, involving temperature, terrain and obstacle information. The data-based statistical model uses historical data of wind farms to find the nonlinear relationship between wind power and historical series, thereby improving the prediction accuracy. Common nonlinear prediction models include support vector machines, random forests, artificial neural networks, etc. Recurrent neural networks have strong memory capabilities, variable length sequence processing capabilities and parameter sharing advantages when processing time series data, and can effectively capture patterns and trends in time series. Long short-term memory and gated recurrent unit neural networks have further helped recurrent series networks to lead in the field of time series. However, in actual application scenarios such as high-frequency and high-noise ultra-short-term wind power prediction, the prediction accuracy of a single model is often difficult to meet application requirements. Therefore, it is of great research significance to design a high-precision, stable and reliable ultra-short-term wind power prediction method. Summary of the invention

[0004] In order to overcome the shortcomings of the prior art, the present invention provides a VMD-PCA-LSTM ultra-short-term wind power prediction method based on the Pelican optimization algorithm. The method combines four modules of data preprocessing, modal decomposition, optimization algorithm and prediction model, expands the advantages and disadvantages of the complementary modules, and improves the accuracy and stability of ultra-short-term prediction.

[0005] To achieve the above technical objectives, the present invention provides the following technical solutions: a VMD-PCA-LSTM ultra-short-term wind power prediction method based on the Pelican optimization algorithm, specifically comprising the following steps:

[0006] S1: Obtain historical power data of the wind farm, as well as historical meteorological data such as wind speed, temperature, and humidity. Clean and interpolate the collected data to obtain preprocessed data;

[0007] S2: VMD decomposition is used to decompose the historical work sequence, and the energy difference method is used to determine the number of sub-modes after VMD decomposition;

[0008] S3: PCA is used to reduce the dimensionality of high-dimensional meteorological features, extract principal components, and preliminarily mine meteorological features;

[0009] S4: The power sequence of each power sub-mode after decomposition P = [p1, p2...p t ], p i ∈R 1×1 , combined with the dimension-reduced meteorological data M = [m1, m2...m t ],m i ∈R t×n After feature concatenation, multivariate feature input is formed. Input the long short-term memory network, capture the temporal relationship and long-term dependency information of wind power and meteorological data through the LSTM loop structure, and establish a prediction model;

[0010] S5: Using POA’s efficient optimization capability, we optimized the four parameters of the two hidden layer units, Dropout rate, and learning rate of the LSTM-based prediction model.

[0011] S6: The prediction results of each sub-modal prediction model are superimposed and reconstructed to achieve ultra-short-term wind power prediction. The reconstruction results need to be denormalized. The present invention uses root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE) as evaluation indicators of model prediction effect.

[0012]

[0013] Where: N is the sample size, y i is the actual value of sample i, is the predicted value of sample i.

[0014] Compared with the prior art, the present invention has the following beneficial effects:

[0015] (1) The VMD decomposition based on the energy difference method can extract the deep features of the power sequence and effectively separate its high and low frequency components. It improves the endpoint effect problem of the traditional EMD decomposition, solves the phase lag problem that is common in single prediction models, and solves the problem of needing to set the number of decomposition sub-modes a priori.

[0016] (2) The present invention uses PCA to extract the principal components of complex and redundant high-dimensional meteorological features. While ensuring sufficient and effective feature extraction, it reduces the fitting pressure of subsequent models and effectively reduces the occurrence of dimensionality disasters.

[0017] (3) The POA algorithm used in the present invention simulates the behavioral strategies of pelicans in attacking and hunting, and effectively balances the two stages of the algorithm approaching prey (exploration stage) and flying on the water surface (exploitation stage). At the same time, the algorithm has the characteristics of simple structure, fast optimization speed, and high optimization accuracy.

[0018] (4) In the context of ultra-short-term power prediction, it is difficult to select the hyperparameters of the LSTM prediction model. The accuracy and efficiency of empirical values ​​or manual trial methods are low. The POA algorithm is used to intelligently optimize the model hyperparameters, which can improve the mapping ability of the LSTM model, accelerate model convergence, and effectively improve the accuracy and stability of the prediction results. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 This is a general framework diagram of the ultra-short-term wind power prediction model of the present invention;

[0020] Figure 2 This is the power-wind speed scatter plot before and after removing outliers using the quartile method;

[0021] Figure 3 This is a comparison chart before and after abnormal data cleaning;

[0022] Figure 4 Each sub-mode diagram of VMD decomposition based on the energy difference method;

[0023] Figure 5 This is the network structure diagram of the long short-term memory neural network;

[0024] Figure 6 Iteration curve of loss function in the process of optimizing network hyperparameters for IMF1's Pelican optimization algorithm;

[0025] Figure 7 This is a comparison chart of the prediction results of the POA-VMD-PCA-LSTM prediction model and other benchmark models;

[0026] Figure 8 It is a partial graph of the prediction results of the POA-VMD-PCA-LSTM prediction model and other benchmark models. DETAILED DESCRIPTION

[0027] The technical solution of the present invention will be described in detail below in conjunction with the accompanying drawings, but the protection scope of the present invention is not limited to the following contents.

[0028] The high volatility and intermittency of power series make it difficult to predict directly. The present invention uses the energy difference method to determine the number of sub-modes of variational mode decomposition, and uses VMD to decompose the original nonlinear power series into more stable sub-modes. Through principal component analysis, the principal components of high-dimensional meteorological characteristics are extracted to reduce the complexity and redundancy of high-dimensional meteorological characteristics. Finally, the POA algorithm is introduced to optimize the hyperparameters of the long short-term memory neural network prediction model, which helps the model converge faster and improve the prediction accuracy. Based on the above, a VMD-PCA-LSTM ultra-short-term wind power prediction method based on the Pelican optimization algorithm is proposed. Figure 1 As shown, the following steps are included:

[0029] S1: Obtain historical power data of the wind farm, as well as historical meteorological data such as wind speed, temperature, and humidity. Clean and interpolate the collected data to obtain preprocessed data;

[0030] S1.1: Based on the correlation between wind power and wind speed, the quartile method is used to clean abnormal data in the data collection process. The comparison before and after data cleaning is as follows: Figure 2 As shown;

[0031] S1.2: Since there is a strong correlation between adjacent data in the time series, the cleaned and missing values ​​are filled with the mean of the nearest and equally spaced data before and after. The wind direction angle data is sinusoidally transformed and its sine value is used as the wind direction input data. In order to facilitate the convergence of model training and reduce the impact of dimension on the prediction results, the power and meteorological data are standardized.

[0032] S2: Figure 3 The VMD decomposition is used to decompose the historical work sequence, and the energy difference method is used to determine the number of sub-modes after VMD decomposition;

[0033] S2.1: Use energy interpolation method to determine the number of sub-modes of VMD decomposition;

[0034] Furthermore, the energy value and energy difference under each decomposition mode number are calculated according to formulas (1) and (2), and the decomposition layer number K with the smallest energy difference is selected as the optimal decomposition mode number.

[0035]

[0036] S2.2: Based on the K value obtained by the energy interpolation method, VMD is used to decompose the original power sequence into K subsequences. VMD is a non-recursive variational mode decomposition method, which can decompose the non-stationary wind power sequence into several sub-modal sequences with certain sparse properties, and can effectively avoid problems such as modal aliasing caused by decomposition methods such as EMD.

[0037] (1) The main framework of this method is to solve the variational problem. Its variational constraint expression is:

[0038]

[0039] Where: δ(t) is the Dirac function, u k (t) is the value of modal component k at time t, ω k is the frequency center of each component.

[0040] (2) By adding the penalty factor α and the Lagrangian operator λ, the problem is transformed into an unconstrained variational problem.

[0041]

[0042] (3) Perform time-frequency domain transformation and use the alternating direction multiplier method (ADMM) to calculate u k ,ω k Iterate and update until the convergence accuracy is reached or the maximum number of iterations is reached.

[0043]

[0044]

[0045] Where n is the number of iterations, τ is the update factor, and ε is the convergence accuracy.

[0046] S3: PCA is used to reduce the dimension of high-dimensional meteorological features (wind speed and direction at 10 meters, wind speed and direction at 30 meters, wind speed and direction at 50 meters, wind speed and direction at hub height, temperature, atmospheric pressure, and humidity), extract the principal components, and preliminarily mine meteorological features;

[0047] Furthermore, the implementation process of PCA includes the following steps:

[0048] (1) Assuming there are m sets of meteorological data samples for power prediction, each sample contains n meteorological features, then the entire data set can be represented by a matrix as follows:

[0049]

[0050] (2) The original matrix is ​​normalized to generate the normalized matrix X * As shown in formula (10):

[0051]

[0052] (3) Calculate the covariance matrix Q:

[0053]

[0054] (4) According to the characteristic equation, we can obtain the characteristic roots λ1,λ2,…,λ n , solve for the unit eigenvector e1,e2,…,e corresponding to each eigenvalue n . Sort the eigenvalues ​​in order of size and calculate the contribution rate of each eigenvalue;

[0055]

[0056] S4: The power sequence of each power sub-mode after decomposition P = [p1, p2...p t ], p i ∈R 1×1 , combined with the dimension-reduced meteorological data M = [m1, m2...m t ],m i ∈R t×n , after feature concatenation, a multivariate feature input is formed. Input the long short-term memory network, and capture the temporal relationship and long-term dependency information of wind power and meteorological data through the LSTM loop structure, where t is the time window length, which is 6 in the present invention. Establish a prediction model;

[0057] LSTM is an improved recurrent neural network developed specifically to solve the long-term memory problem of RNN gradient vanishing. LSTM replaces the single hidden layer structure of the original RNN with the LSTM unit. Among them, the LSTM unit is a four-layer hidden layer structure, which is constructed into three control gates, namely the forget gate, input gate and output gate. Its internal structure is as follows Figure 4 As shown in Figure 2, the three-gate structure can be like the hippocampus of the human brain, processing long-term memory and short-term memory information by operating the created long-term memory line Ct.

[0058] S5: Using POA’s efficient optimization capability, we optimized the Dropout rate, learning rate, and the number of units in the two hidden layers of the LSTM-based prediction model.

[0059] S5.1 Pelican Optimization Algorithm Update Pelican position update includes the following steps:

[0060] The pelican position update mainly includes three stages: population initialization stage, moving towards prey stage, and skimming over the water surface stage;

[0061] (1) In the population initialization stage, the positions of the individuals in the pelican population are determined by the population position matrix in equation (13). Each row of the matrix represents a candidate solution, and each column of the matrix represents the solution of each dimension of the problem variable.

[0062]

[0063] Initially, the population members are randomly initialized using formula (14) according to the lower and upper bounds of the problem:

[0064] x ij = l j +α·(u j -l j ) i=1,2,...,n; j=1,2,...,m (14)

[0065] (2) In the stage of moving towards the prey, the pelican will determine the location of the prey and then move to the determined area. By simulating this strategy of the pelican, the search space can be scanned, thereby improving the exploration ability of POA in discovering different areas of the search space. The strategy of the pelican moving towards the location of the prey is shown in formula (15):

[0066]

[0067] Where P j is the position of the prey in the jth dimension. If the objective function value of the pelican improves at the new position, the position is accepted. In this update, called effective update, the algorithm will not move to non-optimal areas. This process is shown in formula (16).

[0068]

[0069] (3) When the pelicans fly over the water, they spread their wings and hunt on the water, using their throat pouches to collect prey. This process improves the algorithm's local search capabilities. The new position of the pelican can be described by the following mathematical model.

[0070]

[0071] Where t is the current iteration number and T is the maximum iteration number. Similarly, effective update is performed through equation (16).

[0072] S5.2 The specific steps of the Pelican optimization algorithm parameter optimization process are as follows:

[0073] (1) As shown in formula (12), the mean square error function (RMSE) is used as the loss function of the model, and the Adam algorithm is used to reversely update the model gradient. The number of iterations of the model is set to 50 times, and the batch size is set to 64;

[0074]

[0075] (2) The initial population size of the Partridge Optimization Algorithm is set to 20, the maximum number of iterations is set to 50, the network structure parameters of the LSTM are encoded as the population position of the Partridge Optimization Algorithm, and the search space size is 4 dimensions, namely the Dropout rate, the learning rate, the number of units in the first hidden layer, and the number of units in the second hidden layer;

[0076] (3) Set four hyperparameter optimization ranges and initialize the population position. Use the partitioned training set for optimization and the partitioned validation set to calculate the loss function value.

[0077] (4) In a further embodiment, after reaching the maximum number of iterations, the POA algorithm ends and returns the optimized hyperparameters and weights corresponding to the global optimal position; these optimized hyperparameters and weights will be used for subsequent model training and prediction to improve the performance of the LSTM model in the wind power prediction task. Figure 5 It is the loss function iteration curve in the IMF1 optimization process.

[0078] S6: The prediction results of each sub-modal prediction model are superimposed and reconstructed to achieve ultra-short-term wind power prediction. The reconstruction results need to be denormalized. The present invention uses root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE) as evaluation indicators of model prediction effect.

[0079] Example

[0080] The present invention provides an embodiment to verify the proposed prediction method. The example data comes from a wind farm in Xinjiang, China. The wind farm is located in one of the nine wind zones in Xinjiang and is rich in wind resources. The installed capacity of the wind farm is 200MW. A data set for the winter of 2019 was selected for the experiment. The data set contains actual power generation and wind speed (10 meters, 30 meters, 50 meters and wind turbine hub height), wind direction (10 meters, 30 meters, 50 meters and wind turbine hub height), atmospheric pressure, temperature, humidity, and a total of 11-dimensional historical meteorological data. The data set sampling interval is 15min. The data from January 1 to March 31 are selected as winter data, of which March 20 to March 28 are divided into a validation set, March 28 to March 31 are test sets, and the rest are training sets. The training set, validation set, and test set contain 7584, 768, and 288 data, respectively.

[0081] The simulation software used in this paper is Python 3.10.4, and the model is built based on the pytorch1.12.1 framework. The CPU is i7-10750H CPU, the GPU is NVIDIA GeForce GTX 1650Ti, 2.69GHz main frequency, 16GB running memory, and the operating system is Windows 10 (64-bit).

[0082] In order to verify the performance of the proposed model, LSTM and three ablation combination models are used to compare with the prediction method proposed in the present invention. They are PCA-LSTM, VMD-LSTM, and PSO-VMD-PCA-GRU. Due to the simple structure of a single model, the model hyperparameters are determined by multiple manual trials. If there is no combination model for optimizing with an intelligent algorithm, the grid search method is used to select hyperparameters. Among them, the PCA-LSTM model is used to verify the effect of the feature on the prediction effect after the dimension is reduced by principal component analysis. VMD-LSTM is used to reflect the effect of the modal decomposition module. PSO-VMD-PCA-GRU is used to verify the superiority of the optimization ability of the POA algorithm.

[0083] Figure 6 The comparison curves between several prediction models and measured values ​​are shown. Table 1 shows the three corresponding error evaluation indicators. It can be seen from the figure that POA-VMD-PCA-LSTM has significantly improved the prediction effect compared with the single LSTM and the undecomposed PCA-GRU model. The prediction method of the present invention has the best prediction effect and has good fitting performance at the locations with high prediction difficulty such as the extreme points of the measured curve. There is obvious prediction lag in the single model.

[0084] As can be seen from Table 1, the three evaluation indicators of the POA-VMD-PCA-LSTM model have achieved the best. Compared with the traditional LSTM model, the three evaluation indicators are reduced by 66.90%, 65.16%, and 59.97%, respectively, and the prediction accuracy has been significantly improved. The prediction results of PCA-LSTM and VMD-LSTM compared with LSTM verify the applicability of principal component analysis and variational mode decomposition, indicating that the two methods can help the model capture detailed features and learn effectively. Compared with the traditional PSO algorithm, the three evaluation indicators of the prediction model optimized by the POA algorithm are reduced by 34.61%, 37.58%, and 39.96%, respectively, proving that the POA algorithm has obvious advantages in search efficiency and optimization ability, and can better handle complex tasks such as wind power prediction and find the optimal solution.

[0085] Table 1 Comparison of three evaluation indicators for prediction results of each model

[0086]

[0087] The present invention proposes a VMD-PCA-LSTM ultra-short-term wind power prediction method based on the Pelican optimization algorithm. Through example analysis and prediction model comparison, the proposed model has the following advantages: VMD decomposition based on the energy difference method can effectively extract the deep detail features of wind power, improve the modal aliasing and endpoint effect problems of traditional EMD decomposition, improve the prediction accuracy of the combined prediction model, and solve the general phase lag problem of a single prediction model; the improvement of the model prediction accuracy by the PCA method may be relatively limited, but its main advantage is that it can effectively extract the principal component information of the input features, thereby reducing the scale of the internal structure of the model, thereby accelerating the convergence speed of the model and improving the learning efficiency; for different prediction tasks, the LSTM model hyperparameter selection is difficult, and the empirical value and manual trial methods are cumbersome and have limited accuracy. The POA algorithm can overcome the above problems and improve the model prediction accuracy.

[0088] The above is only a preferred embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any equivalent replacement or change made by any technician in the technical field of the present invention, under the premise of following the technical solution and inventive concept of the present invention, should be included in the protection scope of the present invention.

Claims

1. The VMD-PCA-LSTM ultra-short-term wind power forecasting method based on the Pelican optimization algorithm is characterized by: The following steps are involved: S1: Obtain historical power data of the wind farm, clean and interpolate the collected data, and obtain pre-processed data; S2: After determining the number of sub-modes after VMD decomposition using the energy difference method, the historical power sequence is decomposed using VMD decomposition to obtain several sub-modal power sequences; S3: PCA is used to reduce the dimensionality of high-dimensional meteorological features, extract principal components, and preliminarily mine meteorological features; S4: The power sequence of each power sub-mode after decomposition P = [p1, p2...p t ], p i ∈R 1×1 , combined with the dimension reduction meteorological features M = [m1, m2...m t ],m i ∈R t×n After feature concatenation, multivariate feature input is formed and input into the long short-term memory network. The temporal relationship and long-term dependency information of wind power and meteorological data are captured through the LSTM loop structure to establish a prediction model; S5: Using POA’s efficient optimization capability, we optimized the four parameters of the two hidden layer units, Dropout rate, and learning rate based on the LSTM prediction model. S6: The prediction results of each sub-modal prediction model are superimposed and reconstructed to achieve ultra-short-term wind power prediction. The reconstructed results need to be denormalized and the root mean square error (RMSE), mean absolute error (MAE), and determination coefficient (R-Square, R 2 ), as an evaluation index of the model prediction effect.

2. The VMD-PCA-LSTM ultra-short-term wind power prediction method based on the Pelican optimization algorithm according to claim 1 is characterized in that: The preprocessing of the collected data includes the following steps: S1.1: Based on the correlation between wind power and wind speed, the quartile method is used to clean abnormal data in the data collection process; S1.2: Fill the cleaned and missing values ​​with the mean of the nearest and equally spaced data before and after them, perform a sine transform on the wind direction angle data, use its sine value as the wind direction input data, and standardize the power and meteorological data.

3. The VMD-PCA-LSTM ultra-short-term wind power prediction method based on the Pelican optimization algorithm according to claim 1 is characterized in that: The VMD decomposition based on the energy difference method includes the following steps: S2.1: Use energy interpolation method to determine the number of sub-modes of VMD decomposition; According to formulas (1) and (2), the energy value and energy difference under each decomposition mode number are calculated, and the decomposition layer number K with the smallest energy difference is selected as the optimal decomposition mode number: S2.2: Based on the K value obtained by the energy interpolation method, VMD is used to decompose the original power sequence into K subsequences: (1) The main framework of this method is to solve the variational problem, and its variational constraint expression is: Where: δ(t) is the Dirac function, uk(t) is the value of modal component k at time t, and ωk is the frequency center of each component; (2) By adding the penalty factor α and the Lagrangian operator λ, the problem is transformed into an unconstrained variational problem; (3) Perform time-frequency domain transformation and use the alternating direction multiplier method (ADMM) to calculate u k ,ω k Perform iterative updates until the convergence accuracy or the maximum number of iterations is reached; Where n is the number of iterations, τ is the update factor, and ε is the convergence accuracy.

4. The VMD-PCA-LSTM ultra-short-term wind power prediction method based on the Pelican optimization algorithm according to claim 1 is characterized in that: The implementation process of PCA includes the following steps: (1) Assuming there are m sets of meteorological data samples for power prediction, each sample contains n meteorological features, then the entire data set can be represented by a matrix as follows: (2) The original matrix is ​​normalized to generate the normalized matrix X * As shown in formula (10): (3) Calculate the covariance matrix Q: (4) According to the characteristic equation, we can obtain the characteristic roots λ1,λ2,…,λ n , solve for the unit eigenvector e1,e2,…,e corresponding to each eigenvalue n , sort the eigenvalues ​​in order of size and calculate the contribution rate of each eigenvalue; 5. The VMD-PCA-LSTM ultra-short-term wind power prediction method based on the Pelican optimization algorithm according to claim 1 is characterized in that: The LSTM calculation process is as follows: f t =σ(W f ·[h t-1 ,x t ]+b f ) (13) I t =σ(W i ·[h t-1 ,x t ]+b i ) (14) C t =f t *C t-1 +I t *tanh(W c ·[h t-1 ,x t ]+b c ) (15) the t =σ(W o ·[h t-1 ,x t ]+b o ) (16) h t =o t *tanh(C t ) (17) Where: [h t- 1,x t ] are the time series input at the current moment and the hidden state at the previous moment, W and b are the weights and biases of the hidden layers of each gate structure, σ(·) is the Sigmoid function, and tanh(·) is the hyperbolic tangent function.

6. The VMD-PCA-LSTM ultra-short-term wind power prediction method based on the Pelican optimization algorithm according to claim 1 is characterized in that: The Pelican optimization algorithm updates the Pelican position by following these steps: The pelican position update mainly includes three stages: population initialization stage, moving towards prey stage, and skimming over the water surface stage; (1) In the population initialization stage, the positions of the individuals in the pelican population are determined by the population position matrix in equation (18). Each row of the matrix represents a candidate solution, and each column of the matrix represents the solution of each dimension of the problem variable. Initially, the population members are randomly initialized according to the lower and upper bounds of the problem using formula (19): x ij =l j +α·(u j -l j ) i=1,2,...,n; j=1,2,…,m (19) (2) In the stage of moving towards the prey, the pelican will determine the location of the prey and then move to the determined area. By simulating this strategy of the pelican, the search space can be scanned, thereby improving the exploration ability of POA in discovering different areas of the search space. The strategy of the pelican moving towards the location of the prey is shown in formula (20): Where P j is the position of the prey in the jth dimension. If the objective function value of the pelican improves at the new position, the position will be accepted. In this update, called effective update, the algorithm will not move to the non-optimal area. This process is shown in formula (22). (3) When the pelicans fly over the water, they spread their wings and hunt on the water, using their throat pouches to collect prey. This process improves the local search capability of the algorithm. The new position of the pelican can be described by the following mathematical model. In the formula, t is the current iteration number, T is the maximum iteration number, and similarly, effective update is performed through formula (21).

7. The VMD-PCA-LSTM ultra-short-term wind power prediction method based on the Pelican optimization algorithm according to claim 1 is characterized in that: The process of optimizing model hyperparameters by the Pelican Optimizer includes the following steps: (1) The mean square error function (RMSE) is used as the loss function of the model, and the Adam algorithm is used to reversely update the model gradient. The number of model iterations is set to 50 and the batch size is set to 64; (2) The initial population size of the Partridge Optimization Algorithm is set to 20, the maximum number of iterations is set to 30, the network structure parameters of the LSTM are encoded as the population position of the Partridge Optimization Algorithm, and the search space size is 4-dimensional, which are the Dropout rate, the learning rate, the number of units in the first hidden layer, and the number of units in the second hidden layer; (3) Set four hyperparameter optimization ranges and initialize the population position. Use the partitioned training set for optimization and the partitioned validation set to calculate the loss function value.

8. The VMD-PCA-LSTM ultra-short-term wind power prediction method based on the Pelican optimization algorithm according to claim 1 is characterized in that: The wind power prediction method comprises four modules: data preprocessing, power sequence modal decomposition, optimization algorithm and prediction model.

Images