Wind power short-term prediction method based on feature selection and improved HO optimization

Abstract

The invention discloses a wind power short-term prediction method based on feature selection and improved HO optimization. The method comprises the following steps: S1, collecting wind turbine generator data and local meteorological data of a wind power plant; s2, performing normalization preprocessing on the data, and dividing the data into a training set and a test set according to a ratio of 9: 1; s3, screening characteristics with high correlation with the wind power by adopting a maximum information coefficient (MIC); s4, optimizing a hyper-parameter of the Transform-BiLSTM model by using an improved Heima algorithm (IHO); and S5, verifying the optimized model on a test set, and outputting a prediction result by adopting a sliding time window strategy. The method aims at solving the problem that in existing wind power prediction, a traditional optimization algorithm is insufficient in combination model hyper-parameter optimization capacity.

Description

TECHNICAL FIELD

[0001] The application relates to a wind power short-term prediction method based on feature selection and improved HO optimization. BACKGROUND

[0002] Wind power is affected by multiple factors such as seasonal changes, weather conditions (such as wind speed, temperature, and air pressure), and wind turbine operating parameters (such as speed and blade angle), and has strong randomness and uncertainty, which poses a serious challenge to the safe and stable operation of the power system, the improvement of wind power consumption capacity, and the formulation of power market transaction strategies. Therefore, high-precision wind power short-term prediction technology is a key support for ensuring efficient power grid dispatching and improving the economic benefits of wind farms. Currently, scholars have proposed various models to improve prediction accuracy. Among them, the Transformer model achieves global temporal dependency modeling through a multi-head self-attention mechanism, and performs well in complex scenarios. The BiLSTM model captures bidirectional temporal features through a bidirectional structure and is suitable for processing volatile data. Although the Transformer-BiLSTM combined model can balance global and local features, the number of hyperparameters increases, and traditional parameter tuning methods are difficult to adapt. To solve the hyperparameter optimization problem, heuristic algorithms such as particle swarm optimization (PSO), grey wolf optimization (GWO), and sparrow search algorithm (SSA) are widely used, but these algorithms generally have slow convergence speed and are prone to local optimization. The Hippopotamus Optimization Algorithm (HO) has strong evolutionary ability and fast search speed, but its global exploration and local development capabilities are imbalanced, and it is still prone to local optimization when faced with high-dimensional complex problems, resulting in poor optimization of model hyperparameters and difficulty in meeting the high-precision requirements of wind power prediction. Therefore, there is an urgent need for a method that can effectively optimize the hyperparameters of the combined model and improve prediction accuracy. SUMMARY

[0003] The application aims to solve the problem of insufficient optimization of combined model hyperparameters by traditional optimization algorithms in existing wind power prediction, and provides a wind power short-term prediction method based on feature selection and improved HO optimization.

[0004] To solve the above problems, the technical scheme of the application is as follows:

[0005] A wind power short-term prediction method based on feature selection and improved HO optimization, comprising the following steps:

[0006] S1: Collect wind turbine data and local weather data of a wind farm;

[0007] S2: Normalize the data and divide it into a training set and a test set in a ratio of 9:1;

[0008] S3: Use the maximum information coefficient (MIC) to screen features that are highly correlated with wind power;

[0009] S4: Optimize the hyperparameters of the Transformer-BiLSTM model using the Improved Hippo Algorithm (IHO);

[0010] S5: Validate the optimized model on the test set and output the prediction results using a sliding time window strategy.

[0011] Furthermore, the data mentioned in step S1 includes wind power, average wind speed, maximum wind speed, minimum wind speed, blade angle, wind direction, rotational speed, nacelle location, outdoor temperature, outdoor humidity, air pressure, hub temperature, and nacelle temperature. This comprehensive data collection can more accurately reflect the various factors affecting wind power, providing a rich and effective data foundation for subsequent prediction models. This allows the models to better learn the relationship between various factors and wind power, thereby improving the accuracy of predictions.

[0012] Furthermore, the normalization preprocessing described in step S2 is used to eliminate the dimensional differences between features, avoiding the excessive influence of certain features with large dimensions on model training. This ensures that each feature has equal importance in model training, which helps the model converge faster and learn the relationship between features and wind power more accurately. The training set is used for model training, and the test set is used for model validation. Dividing the data into training and test sets provides a reasonable sample allocation method for model training and validation, enabling the model to learn patterns in the training set and validate its generalization ability in the test set, thus ensuring the scientificity and reliability of model performance evaluation.

[0013] Furthermore, the Maximum Information Coefficient (MIC) described in S3 is used to select average wind speed, maximum wind speed, minimum wind speed, rotational speed, and wind direction as highly correlated features by calculating the maximum information number between variables and the grid partitioning. This effectively measures linear and nonlinear correlations without requiring assumptions about data distribution, thus accurately identifying key features highly correlated with wind power, such as average wind speed and maximum wind speed. Eliminating redundant features not only reduces the computational complexity of the model and the risk of overfitting but also allows the model to focus on key influencing factors, improving its predictive efficiency and accuracy.

[0014] The screening method includes Discretizing the correlation between two variables into a two-dimensional space using mutual information (MIC) and applying a grid. The maximum information number (MIC) is calculated using both mutual information and the grid partitioning method. The formula for calculating MIC is:

[0015]

[0016] In the formula: A and B are random variables; p(a,b) is the joint probability density of A and B; p(a) and p(b) are the marginal probability densities of A and B, respectively; MI(A,B) is the mutual information of A and B; c and d are the number of grid cells in the a and b directions, respectively; B(n′) is a hyperparameter; and n′ is the number of samples.

[0017] Furthermore, the improvements to the Improved Hippo Algorithm (IHO) described in step S4 include: introducing a Bernoulli-Tent hybrid chaotic mapping during the population initialization phase, using random numbers and thresholds to control the switching between the two mappings to balance search capabilities. This enhances population diversity. This diversity allows the algorithm to explore different regions more broadly in the search space, effectively avoiding getting trapped in local optima, improving the algorithm's global search capability, and thus making it more likely to find the optimal combination of hyperparameters for the Transformer-BiLSTM model.

[0018] Furthermore, the improved Hippo algorithm (IHO) also includes an adaptive weighting strategy introduced during the exploration phase, where the weights gradually decrease as the number of iterations increases, as shown in the formula:

[0019]

[0020] In the formula, t is the current iteration number, and Max_iter is the maximum iteration number.

[0021] An adaptive weighting strategy is introduced during the exploration phase, causing the weights to gradually decrease as the number of iterations increases. In the early stages of iteration, larger weights help the algorithm perform a broader global search and quickly locate the region where the optimal solution may exist. As the iterations progress, the weights gradually decrease, allowing the algorithm to search more precisely in local regions, accelerating the convergence speed, and improving the efficiency and accuracy of the algorithm in the hyperparameter optimization process.

[0022] Furthermore, the Improved Hippo Algorithm (IHO) also includes an improved Levy distribution strategy, which adjusts the step size through a perturbation factor. The perturbation is large in the early stages of iteration to escape local optima, and weakens in the later stages to accelerate convergence. The perturbation factor formula ν is:

[0023] An improved Levy distribution strategy is introduced, adjusting the step size through a perturbation factor. In the early stages of iteration, a larger perturbation factor enhances the algorithm's ability to escape local optima, enabling it to explore a wider search space. As the number of iterations increases, the perturbation factor gradually weakens, allowing the algorithm to converge more stably to the global optimum. This balances the algorithm's search capabilities at different stages, further improving the performance of the Improved Hippo Algorithm (IHO) in hyperparameter optimization.

[0024] Furthermore, the hyperparameters described in S4 include the number of heads in the Transformer self-attention mechanism, the number of neurons in the BiLSTM hidden layer, the maximum number of iterations, the initial learning rate, and the regularization coefficient. By optimizing these hyperparameters, the Transformer-BiLSTM model can better adapt to the characteristics of wind power prediction tasks, uncover potential patterns in the data, and thus significantly improve the model's prediction performance and generalization ability.

[0025] Furthermore, in the Transformer-BiLSTM model, the Transformer encoder includes a position encoding layer, a multi-head self-attention mechanism, and a feedforward neural network. Position encoding distinguishes feature locations using sin / cos functions; the BiLSTM captures the sequential dependencies of time-series data through a bidirectional structure. This helps the model learn the sequential relationships in time-series data; the bidirectional structure of BiLSTM can capture the sequential dependencies of time-series data, making full use of the information in the data. The combination of these two elements enables the model to learn the characteristics and patterns of wind power data more comprehensively and accurately, providing a solid model foundation for high-precision prediction.

[0026] Furthermore, the prediction results described in S5 are expressed using the mean absolute error (MAE), root mean square error (RMSE), and fitting coefficient R. 2 Evaluation, including

[0027]

[0028] In the formula: p is the sample size; y i′ This represents the actual value of the i′ sample. d is the predicted value for the i′th sample; i′ This represents the average of the actual values ​​for the i′th sample. These metrics provide a comprehensive and objective evaluation of the model's predictive performance from different perspectives. They offer quantitative indicators for assessing the model's accuracy, stability, and fit to real-world data, helping researchers and engineers intuitively understand the model's strengths and weaknesses, thereby facilitating further improvement and optimization.

[0029] The beneficial effects of this invention are as follows:

[0030] (1) The present invention uses the maximum information coefficient method to screen characteristic variables such as meteorological and wind turbine operating parameters, select characteristic variables that are strongly correlated with wind power, and improve the prediction accuracy of wind power prediction model.

[0031] (2) This invention overcomes the shortcomings of HO by integrating Bernoulli-Tent hybrid chaotic mapping, adaptive weight strategy and improved Levy distribution strategy to improve IHO, significantly improves the convergence speed, enhances the global search capability and effectively improves the ability to escape local optima. Using IHO to fine-tune the model hyperparameters can improve the hyperparameter fit of the wind power prediction model.

[0032] (3) Through comparative experiments of prediction models, the wind power prediction of IHO-Transformer-BiLSTM is closer to the true value and has lower error than the optimized Transformer-BiLSTM and Transformer-BiLSTM models of GWO, SSA and HO, and has better prediction accuracy. Attached Figure Description

[0033] The invention will be further described below with reference to the accompanying drawings:

[0034] Figure 1 This is an overall flowchart of the method of the present invention.

[0035] Figure 2 This is a screening diagram of wind power-related features for this invention.

[0036] Figure 3 This is a comparison chart of the test functions of this invention.

[0037] Figure 4 This is a comparison chart of wind power prediction based on the present invention.

[0038] Figure 5 This is a comparison chart of wind power prediction and evaluation indicators of the present invention.

[0039] Figure 6 The optimized models of the four algorithms of this invention were evaluated in the test set for MAE, RMSE, and R. 2 The comparison chart. Detailed Implementation

[0040] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0041] like Figure 1 As shown, a short-term wind power prediction method based on feature selection and an improved Hippo algorithm-optimized Transformer-BiLSTM is implemented according to the following steps:

[0042] S1: Research and collect data on wind turbine units at the wind farm and local meteorological bureau data, including wind power, average wind speed, maximum wind speed, minimum wind speed, blade angle, wind direction, speed, nacelle location, wind power, outdoor temperature, outdoor humidity, air pressure, hub temperature, and nacelle temperature, in that order. Figure 2 D1-D13 in the middle.

[0043] S2: To address the issue of differences in the units of measurement in the original data, preprocessing is required. Multidimensional data is normalized to eliminate the impact of differences in the units of measurement between features on model training. The dataset is then divided into training and test sets in a 9:1 ratio.

[0044] S3: MIC uses the following method to screen out features such as meteorological and wind turbine operating parameters that are highly correlated with wind power:

[0045] The maximum information coefficient (MIC) is a correlation algorithm used to measure the strength of the association between variables, specifically whether it is linear or nonlinear. MIC assesses the functional and statistical relationships between variables without making any assumptions about the data distribution. It discretizes the correlation between two variables into a two-dimensional space using a grid. The maximum information coefficient is calculated using both mutual information and the grid partitioning method. The formula for calculating MIC is as follows:

[0046]

[0047] In the formula: A and B are random variables; p(a,b) is the joint probability density of A and B; p(a) and p(b) are the marginal probability densities of A and B, respectively; MI(A,B) is the mutual information of A and B; c and d are the number of grid cells in the a and b directions, respectively; B(n′) is a hyperparameter; and n′ is the number of samples.

[0048] Including all meteorological and wind turbine parameter features during the experiment may lead to feature redundancy, increasing the risk of model overfitting. Conversely, insufficient selection of key feature variables can result in missing information, affecting the model's prediction accuracy. Therefore, feature selection is performed using Micro-Matching (MIC), such as... Figure 2 As shown in the figure. Among them, the average wind speed, maximum wind speed, minimum wind speed, rotational speed, and wind direction have a strong correlation with wind power, so these characteristics were selected for subsequent experiments.

[0049] Furthermore, in S4, the improved HO algorithm is used to optimize five hyperparameters of the Transformer-BiLSTM model: the number of heads in the Transformer self-attention mechanism, the number of neurons in the BiLSTM hidden layer, the maximum number of iterations, the initial learning rate, and the number of regularization coefficients, as detailed below. Figure 3 .

[0050] In the S4.1 population initialization stage, a Bernoulli-Tent hybrid chaotic map is introduced. While the Bernoulli map provides powerful global search capabilities, it lacks sufficient fine-tuning, whereas the Tent map offers more refined local exploration capabilities, compensating for Bernoulli's shortcomings. Therefore, a composite chaotic sequence, Bernoulli-Tent, fusing Bernoulli and Tent sequences, is introduced in the IHO initialization stage. First, an initial point e0 is randomly generated using a uniform distribution within the interval [0,1]. Then, a chaotic map switching rule is introduced: a random number r is generated within [0,1], and a parameter l is used as a threshold to control the switching between the two maps, ensuring search balance. Its mathematical expression can be represented as:

[0051]

[0052] In the hippopotamus exploration phase (S4.2), the algorithm's global search capability is strong when the initial weights are relatively large. As the number of iterations increases, the weight coefficients gradually decrease. Introducing an adaptive weight strategy allows for a finer search around the optimal solution, accelerating the convergence speed. The adaptive weight formula is:

[0053]

[0054] In the formula: X_P1(i,:) is the position of the male hippopotamus; w(t) is the weighting factor; X(i,:) is the position of the i-th candidate solution; r1 is a uniform random number in [0,1]; D hippo I1 represents the position of the dominant male hippopotamus; I1 is a uniform random number in the range [1,2]; Max_iter is the maximum number of iterations.

[0055] As the number of iterations increases, individual hippos cluster towards the optimal position. The greater the weight, the faster the individual moves, thus improving the convergence speed of the algorithm.

[0056] S4.3 The Levy distribution strategy used in traditional HO can generate random step sizes, improve population diversity, and enable the algorithm to escape local optima. However, in the later stages of iteration, using the Levy distribution strategy to represent the predator mutation position can actually reduce the convergence speed of the population. Therefore, this invention introduces an improved Levy distribution strategy, adding a perturbation factor ν to the original predator posture. In the early stages of algorithm iteration, the perturbation is larger, and the algorithm has a stronger ability to escape local optima. As the number of iterations increases, the perturbation effect gradually weakens. The mathematical model of the perturbation factor ν is shown in Equation (6), and the overall position update model is shown in Equation (7), where x best This represents the optimal population position for the current iteration number.

[0057]

[0058] In the formula: X_P2(i,:) is the location of the predator invading the hippopotamus's territory; RL is a random vector with a Levy distribution; It is the XOR operator; Let be the predator's position in the search space; ζ be a uniformly random number in the range [2,4]; c1 be a uniformly random number in the range [1,1.5]; d1 be a uniformly random number in the range [2,3]; g be a uniformly random number in the range [-1,1]; D be the distance between the hippopotamus and the predator; r2 be a 1×m uniformly random vector; n be the hippopotamus population size; and m be the number of decision variables. F represents the target value for the hippopotamus during the defense exploration phase. i This represents the objective function value.

[0059] S4.4 Using the above steps, construct the improved Hippo algorithm to optimize the Transformer-BiLSTM model, including the number of Transformer self-attention mechanism heads, the number of BiLSTM hidden layer neurons, the maximum number of iterations, the initial learning rate, and the number of regularization coefficients.

[0060] The Transformer-BiLSTM model is as follows:

[0061] The encoder of the Transformer model mainly consists of a positional encoding layer, a multi-head self-attention mechanism, and a feedforward neural network. The multi-head self-attention mechanism uses parallel computation to correlate features at each time point with features at other time points in the sequence, thereby constructing a more accurate global representation. This global representation can effectively capture long-term dependencies in wind power prediction, thus improving the ability to identify key moments and the accuracy of prediction. The standard Transformer model is a full encoder-decoder structure. This invention focuses on the Transformer's encoding layer, effectively simplifying the computational path of the model while fully retaining the powerful modeling capabilities of the multi-head attention mechanism for temporal features, providing a more lightweight and efficient solution for wind power prediction tasks.

[0062] Position encoding can represent the relative positional information between different data, and its mathematical expression can be expressed as:

[0063]

[0064]

[0065] In the formula: pos represents the position of the feature in the feature sample; h represents the dimension of the position encoding, where h is greater than 2q and 2q+1. The cosine function is used when the position of the feature in the sample is odd, while the sinine function is used when the position of the feature in the sample is even.

[0066] The multi-head attention mechanism is the core of this model, mainly consisting of the query matrix Q, the key matrix K, and the value matrix V. The multi-head attention mechanism employs scaled dot product operations, treating the input encoding as a set of KV pairs, with the output being a weighted sum of V. The weight of each value is determined by the dot product of Q and K. The expression for the attention mechanism is as follows: The multi-head attention mechanism combines multiple dot product attention modules using the concatenation operation. The model calculates the correlation between data and assigns weights, and its mathematical expression can be represented as:

[0067]

[0068] In the formula: Softmax is the activation function; d K Let K be the dimension, where Its function is to scale the attention score of the dot product; head q = (i = 1, 2, ..., n1) is the output of the qth attention head, and n1 is the total number of attention heads.

[0069] BiLSTM fully leverages the sequential information in time series data by simultaneously processing both forward and reverse sequences, comprehensively capturing the long-term dependencies in wind power data. Addressing the intermittent and volatile nature of wind power data, BiLSTM utilizes its bidirectional structure to effectively mine data features and process these complex time series data. Its mathematical expression can be represented as:

[0070] A i =f1(ω1x i +ω2A i-1 (12);

[0071] B i =f2(ω4x i +ω3A i+1 (13);

[0072] Y i =f3(ω5A i +ω6B i (14);

[0073] In the formula: f1, f2, and f3 are the activation functions of each layer; x1, x2, x3, and x t Input data for the structural response at different times; A1, A2, A3, and A t The hidden states of the forward LSTM; B1, B2, B3, and B t The hidden states of the backward LSTM are: Y1, Y2, Y3, and Y... tThe corresponding output data; ω1, ω2, ω3, ω4, ω5 and ω6 are the weights of different levels.

[0074] S5: Output the prediction results. Validate the optimized combined model on the test set. Use a sliding time window prediction strategy to output the wind power prediction curve. The specific method is as follows:

[0075] The final prediction results are presented through a comparison of the predicted curve and the actual wind power output, the mean absolute error (MAE), the root mean square error (RMSE), and the fitting coefficient R. 2 Demonstrates the model's predictive performance. Its mathematical expression can be represented as:

[0076]

[0077] In the formula: p is the sample size; y i′ This represents the actual value of the i′ sample. d is the predicted value for the i′th sample; i′ is the average of the actual values ​​of the i′ sample.

[0078] Experiment 1: Convergence and convergence speed comparison of the IHO algorithm with the HO algorithm, Sparrow Search Algorithm (SSA), and Grey Wolf Algorithm (GWO). To ensure the feasibility of the algorithm comparison experiment, three unimodal functions and three multimodal functions were selected for verification. All test functions were CEC2005 benchmark functions. The maximum number of iterations for each algorithm was 500, the population size was 30, and the dimension was 5. The benchmark functions used are shown in Table 1.

[0079] Table 1 Test Function Table

[0080]

[0081] Depend on Figure 4 As can be seen, compared with the GWO, SSA, and HO algorithms, the IHO algorithm proposed in this study has significantly lower optimal fitness values ​​for the six test functions and the fastest convergence speed, exhibiting better hyperparameter optimization characteristics. The IHO algorithm, through Bernoulli-Tent hybrid chaotic mapping, adaptive weighting strategy, and improved Levy distribution strategy, demonstrates stronger search capabilities, faster convergence speed, and avoids getting trapped in local optima compared to other algorithms.

[0082] Experiment 2: Comparison of wind power prediction under different algorithm optimizations.

[0083] Depend on Figure 5It can be seen that HO has a better fitting curve than GWO and SSA, but all three are inferior to the model optimized by IHO. The IHO algorithm, due to its fusion of three strategies, has stronger global search and the ability to escape local optima. The model optimized by the IHO algorithm has a prediction curve closer to the true value, a smoother fitting curve, and a more reasonable fit between model hyperparameters. The four optimized models performed well in the test set in terms of MAE, RMSE, and R... 2 like Figure 6 .Depend on Figure 6 It can be seen that the MAE, RMSE, and R of IHO-Transformer-BiLSTM 2 The scores were 28.01, 45.12, and 0.9894, respectively. Their MAE was reduced by 47.31%, 37.28%, and 27.49% compared to the optimized models using GWO, SSA, and HO algorithms, respectively; their RMSE was reduced by 39.54%, 27.28%, and 19.36%, respectively; and their R... 2 Compared to the other three algorithms, the optimized model also shows improvement, demonstrating the better robustness and reliability of the IHO optimized model proposed in this paper.

[0084] The embodiments described in this specification are merely examples of implementations of the inventive concept. The scope of protection of this invention should not be considered as limited to the specific forms stated in the embodiments. The scope of protection of this invention also extends to equivalent technical means that can be conceived by those skilled in the art based on the inventive concept.

Claims

1. A short-term wind power forecasting method based on feature selection and improved HO optimization, characterized in that, Includes the following steps: S1: Collect wind turbine data and local meteorological data from the wind farm; S2: Perform normalization preprocessing on the data and divide it into training set and test set in a 9:1 ratio; S3: Use the maximum information coefficient (MIC) to screen features that are highly correlated with wind power; S4: Optimize the hyperparameters of the Transformer-BiLSTM model using the Improved Hippo Algorithm (IHO); S5: Validate the optimized model on the test set and output the prediction results using a sliding time window strategy.

2. The short-term wind power prediction method based on feature selection and improved HO optimization according to claim 1, characterized in that, The data mentioned in step S1 includes wind power, average wind speed, maximum wind speed, minimum wind speed, blade angle, wind direction, rotational speed, nacelle position, outdoor temperature, outdoor humidity, air pressure, hub temperature, and nacelle temperature.

3. The short-term wind power prediction method based on feature selection and improved HO optimization according to claim 1, characterized in that, The normalization preprocessing in step S2 is used to eliminate the dimensional differences between features. The training set is used for model training, and the test set is used for model validation.

4. The short-term wind power prediction method based on feature selection and improved HO optimization according to claim 1, characterized in that, The Maximum Information Coefficient (MIC) described in S3 is determined by calculating the maximum information number between variables and the mesh partitioning. Mean wind speed, maximum wind speed, minimum wind speed, rotational speed, and wind direction are selected as highly correlated features. The selection method includes discretizing the correlation between two variables into a two-dimensional space using MIC and mesh partitioning. The maximum information number is calculated jointly using mutual information and mesh partitioning. The formula for calculating MIC is: In the formula: A and B are random variables; p(a,b) is the joint probability density of A and B; p(a) and p(b) are the marginal probability densities of A and B, respectively; MI(A,B) is the mutual information of A and B; c and d are the number of grid cells in the a and b directions, respectively; B(n′) is a hyperparameter; and n′ is the number of samples.

5. The short-term wind power prediction method based on feature selection and improved HO optimization according to claim 1, characterized in that, The improvement to IHO in step S4 includes: introducing a Bernoulli-Tent hybrid chaotic mapping during the population initialization stage, using random numbers and thresholds to control the switching between the two mappings to balance search capabilities, and introducing a composite chaotic sequence Bernoulli-Tent that fuses Bernoulli and Tent sequences during the IHO initialization stage. The introduction method includes first randomly generating an initial point e0 uniformly distributed within the interval [0,1], and then introducing a chaotic mapping switching rule, that is, generating a random number r in [0,1] and using parameter l as a threshold to control the switching between the two mappings to ensure the balance of the search. Its mathematical expression can be expressed as: In the formula: λ is a random number in [0,1]; e n and e n+1 All are random numbers in the range [0,1], e n+1 This represents the size of each individual in the population.

6. The short-term wind power prediction method based on feature selection and improved HO optimization according to claim 1, characterized in that, The IHO also includes an adaptive weighting strategy introduced during the exploration phase, where the weights gradually decrease as the number of iterations increases. The adaptive weighting formula is: X_P1(i,:)=w(t)×X(i,:)+r1×(D hippo -I1X(i,:)); In the formula, X_P1(i,:) represents the position of the male hippopotamus; w(t) is the weighting factor; X(i,:) represents the position of the i-th candidate solution; r1 is a uniformly random number in the range [0,1]; D hippo I1 represents the position of the dominant male hippopotamus; I1 is a uniform random number in the range [1,2]; Max_iter is the maximum number of iterations.

7. The short-term wind power prediction method based on feature selection and improved HO optimization according to claim 1, characterized in that, The Improved Hippo Algorithm (IHO) also incorporates an improved Levy distribution strategy, adjusting the step size through a perturbation factor. The perturbation is large in the early iterations to escape local optima, and weakens in the later iterations to accelerate convergence. The perturbation factor formula ν is: In the formula: τ is a random number in the range [0,1].

8. The short-term wind power prediction method based on feature selection and improved HO optimization according to claim 1, characterized in that, The hyperparameters mentioned in S4 include the number of heads in the Transformer self-attention mechanism, the number of neurons in the BiLSTM hidden layer, the maximum number of iterations, the initial learning rate, and the regularization coefficient.

9. A short-term wind power prediction method based on feature selection and improved HO optimization according to claim 1, characterized in that, In the Transformer-BiLSTM model, the Transformer encoder includes a position encoding layer, a multi-head self-attention mechanism, and a feedforward neural network. The position encoding distinguishes feature positions through the sin / cos function; the BiLSTM captures the sequential dependencies of time-series data through a bidirectional structure.

10. A short-term wind power prediction method based on feature selection and improved HO optimization according to claim 1, characterized in that, The prediction results described in S5 are expressed using mean absolute error (MAE), root mean square error (RMSE), and fitting coefficient R. 2 Evaluation, including In the formula: p is the sample size; y i′ This represents the actual value of the i′ sample. d is the predicted value for the i′th sample; i′ is the average of the actual values ​​of the i′ sample.

Images