Dual-stage wind power generation capacity prediction method and device considering meteorological factors

Abstract

The invention relates to a two-stage wind power generation capacity prediction method and device considering meteorological factors, and the method comprises the steps: obtaining historical power generation capacity data and meteorological data, and carrying out the importance sorting of meteorological characteristics; predicting a generating capacity time sequence at a future moment by utilizing a wind generating capacity prediction model based on the ARIMA based on the historical generating data; subtracting a real value from the predicted value of the generating capacity time sequence to obtain a residual sequence, inputting the residual sequence and the sorted meteorological features into a residual prediction model based on BiLSTM and an attention mechanism, and outputting a residual predicted value of the model; and adding the power generation capacity time sequence predicted value after reverse normalization and the residual predicted value to obtain a corrected wind power generation capacity predicted value. Compared with the prior art, the method is suitable for a multi-time-scale scene, and a high-precision prediction basis is provided for power generation plan making, power grid dispatching optimization, power market transaction and wind power plant operation and maintenance decision making.

Description

Technical Field

[0001] This invention relates to the field of power system power generation prediction, and in particular to a two-stage wind power generation prediction method and apparatus based on historical power generation data and meteorological data. Background Technology

[0002] As the global energy structure transitions towards cleaner and lower-carbon energy, wind power, as a core component of renewable energy, has seen its installed capacity and electricity generation continuously increase. However, wind energy is highly volatile and intermittent, and its power generation is comprehensively affected by meteorological factors such as wind speed, wind direction, temperature, and air pressure, as well as the operating status of equipment. This significantly increases the difficulty of grid dispatching, power supply and demand balancing, and wind farm operation and maintenance decisions. Therefore, high-precision wind power generation forecasting has become a key technological support for ensuring the stable operation of the power system and improving wind power absorption.

[0003] Existing wind power generation forecasting methods are mainly divided into two categories: statistical learning methods based on historical data and physical models based on meteorological factors. However, each method has significant limitations: statistical learning methods rely excessively on historical data patterns and are poorly adaptable to sudden weather changes; while physical models can reflect the impact of meteorological factors, they struggle to capture the nonlinear correlations implicit in historical data such as equipment aging and maintenance status. Furthermore, forecasting needs vary significantly across different time scales, and existing models struggle to balance forecasting accuracy and computational efficiency across multiple scenarios, especially under extreme weather conditions or data noise interference, resulting in large error fluctuations.

[0004] Furthermore, with the large-scale development of wind farms and the advancement of power market reforms, grid dispatching requires more refined power generation forecasts to support intraday unit combination optimization. Wind farms also need to participate in power trading and formulate operation and maintenance plans through accurate forecasts. Traditional single-stage forecasting models can no longer meet the multi-dimensional and highly dynamic application needs. Against this backdrop, constructing a model that integrates historical power generation data and meteorological data, and optimizes feature extraction and forecasting accuracy in stages, has become an important direction for overcoming existing technological bottlenecks.

[0005] For example, the energy consumption prediction method disclosed in Chinese patent application CN202411453193.6 relies solely on historical energy consumption data for prediction, neglecting external environmental factors (such as meteorological and economic indicators). This results in insufficient robustness of the model to sudden changes and also suffers from technical problems such as unreasonable model order selection, leading to overfitting or underfitting. Chinese patent application CN202210406662.3 discloses a method for predicting water conservancy facility deformation by integrating seasonal ARIMA and BiLSTM, but this method does not consider environmental variables such as temperature and water level. The prediction accuracy is limited by the data dimensionality, and it also suffers from limited residual correction capabilities. Summary of the Invention

[0006] The present invention addresses the shortcomings of existing technologies by providing a two-stage wind power generation prediction model and device based on historical power generation data and meteorological data.

[0007] The objective of this invention can be achieved through the following technical solutions: As a first aspect of the present invention, a two-stage wind power generation prediction method considering meteorological factors is provided, comprising the following steps: Acquire historical power generation data and meteorological data, and rank the meteorological features by importance; Based on historical power generation data, an ARIMA-based wind power generation prediction model is used to predict the power generation time series for future moments. The residual sequence is obtained by subtracting the actual value from the predicted value of the power generation time series. The residual sequence and the sorted meteorological features are then input into the residual prediction model based on BiLSTM and attention mechanism, and the model's residual prediction value is output. The corrected wind power generation forecast is obtained by adding the inverse normalized time series forecast value to the residual forecast value.

[0008] As a preferred technical solution, the meteorological features are ranked by the importance of the features using a random forest algorithm. The meteorological features include wind speed and direction, temperature, air pressure, and humidity.

[0009] As a preferred technical solution, the ARIMA-based wind power generation prediction model includes: an autoregressive part, which uses past observations to predict future values; a differencing part, which converts non-stationary power generation time series data into stationary time series data through differencing; and a moving average part, which considers a linear combination of errors and adjusts the current prediction based on previous prediction errors.

[0010] As a preferred technical solution, the ARIMA-based wind power generation prediction model performs stationarity and white noise tests on the power generation time series when predicting wind power generation.

[0011] As a preferred technical solution, the ARIMA-based wind power generation prediction model uses a unit root test to check the stationarity of the time series after differencing. If the test is rejected, the series is further differencing until the series becomes stationary.

[0012] As a preferred technical solution, the residual prediction model based on BiLSTM and attention mechanism includes an input layer, a two-layer BiLSTM layer, an attention layer, and a fully connected layer connected in sequence; a batch normalization layer is added after the input layer and the attention layer; a random deactivation layer is introduced after the BiLSTM layer to prevent overfitting; the fully connected layer includes two layers: the output dimension of the first fully connected layer is determined by the sliding window size and the number of meteorological features, and the output dimension of the second fully connected layer is 1; The first normalization layer normalizes the meteorological characteristic data of the input layer to eliminate differences in characteristic dimensions; The dual-layer BiLSTM layer captures the bidirectional time dependence of meteorological sequences, enhancing the capture of complex time series patterns; The random deactivation layer randomly discards some feature nodes from the output of the two-layer BiLSTM layer during training; The attention layer assigns different weights to the temporal features output by the random deactivation layer, strengthening the meteorological time segments that are key to the prediction results and weakening the influence of irrelevant features. The first fully connected layer integrates the high-dimensional temporal features output by the attention layer with the temporal information corresponding to the sliding window, and maps them into low-dimensional dense features. The second fully connected layer further maps the integrated features output by the first fully connected layer, outputting the final single-value prediction result.

[0013] As a preferred technical solution, the method jointly optimizes the autoregressive order and moving average order of the ARIMA model, as well as the number of neurons in the two BiLSTM layers stacked sequentially in the bilayer BiLSTM structure.

[0014] As a preferred technical solution, the joint optimization adopts a particle swarm optimization algorithm, using the autoregressive order and moving average order of the ARIMA model, as well as the number of neurons in the first and second hidden layers of the bilayer BiLSTM residual prediction model, as particle position information. The optimization process is as follows: The position and velocity of each particle are initialized, and the particles constitute the initial population in the search space; Calculate the fitness value, individual extreme value, and global extreme value for each particle at its current position; Update the velocity and position of each particle based on individual and global extrema; Repeat the above process until the fitness value no longer changes, thus obtaining the optimal combination of parameters for the model.

[0015] As a second aspect of the present invention, a two-stage wind power generation prediction device considering meteorological factors is provided, including a memory, a processor, and a program stored in the memory, wherein the processor executes the program to implement the two-stage wind power generation prediction method considering meteorological factors as described above.

[0016] As a third aspect of the invention, a storage medium is provided having a program stored thereon, which, when executed, implements the two-stage wind power generation prediction method considering meteorological factors as described above.

[0017] Compared with the prior art, the present invention has the following beneficial effects: 1) The dual-stage wind power generation prediction method proposed in this invention can effectively isolate non-critical interference factors by merging historical power generation data and meteorological data in stages, accurately capture the intrinsic relationship between power generation trends and meteorological factors, and significantly improve prediction reliability; it can flexibly adapt to the needs of different types of wind farms and multiple time scale scenarios, covering diverse applications from short-term scheduling to long-term planning; it has strong anti-interference capabilities and can maintain stable performance under complex operating conditions, while helping to improve wind power utilization efficiency and promote the efficient and low-carbon operation of the energy system.

[0018] 3) This invention effectively solves the problems of blind parameter optimization, data-driven limitations, and insufficient residual prediction in existing patents by using PSO to collaboratively optimize ARIMA and BiLSTM parameters and fuse meteorological data for residual prediction. It significantly improves prediction accuracy, robustness, and computational efficiency, and has important engineering application value. Attached Figure Description

[0019] Figure 1 This is a flowchart of a two-stage wind power generation prediction method based on historical power generation data and meteorological data provided by the present invention.

[0020] Figure 2 This is a schematic diagram of the prediction model structure based on BiLSTM and attention mechanism of the present invention.

[0021] Figure 3 This refers to the original power generation data in an embodiment of the present invention.

[0022] Figure 4 This is the two-stage predicted value of power generation in an embodiment of the present invention. Detailed Implementation

[0023] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0024] Example 1 This invention provides a two-stage wind power generation prediction method based on historical power generation data and meteorological data. For example... Figure 1 As shown, the specific operating steps are as follows: S1. Data analysis and preprocessing: Obtain historical power generation data, normalize the data, and lay a unified data foundation for subsequent feature analysis and model training. S2. Construct a wind power generation prediction model based on historical power generation data, use the PSO algorithm to determine the order of the ARIMA prediction model, and realize the time series prediction of power generation. S3. Construct a wind power generation prediction model based on numerical meteorological data, and obtain the residual sequence by subtracting the actual value from the predicted value of the ARIMA model. S4. The meteorological characteristics and residual sequence are fed into the two-layer BiLSTM model optimized by the PSO algorithm. The residual sequence predicted by the BiLSTM model is added to the ARIMA prediction value to obtain the corrected power generation prediction value.

[0025] The data analysis and preprocessing in step S1 are as follows: First, set the sliding window size to obtain continuous time series, use linear interpolation to fill in missing values, and normalize the dataset; then, divide the dataset into training and test sets.

[0026] Specifically, data for the target time period is extracted from the database and aligned with timestamps to form a dataset; the proportion of missing values ​​is calculated, and if the missing rate of a single feature is less than 5%, it is filled with the average of the preceding and following times; if the missing rate is ≥ 5%, the corresponding time period is marked and investigated in conjunction with the equipment operation and maintenance logs.

[0027] Based on the 3σ principle: Calculate the mean of each feature. with standard deviation Marking out The values ​​within the range represent potential anomalies; Combining box plot method: Calculate characteristic interquartile range ,mark or The value; Business rule verification: Remove abnormally low power generation during equipment maintenance and downtime periods, and mark the power generation corresponding to instantaneous abnormally high wind speeds caused by extreme weather.

[0028] Random errors: Linear interpolation is used to complete the data (based on fitting three normal data points before and after the outlier); Systematic errors: Outlier data segments are directly removed, while the corresponding timestamps are retained to avoid affecting the model's trend learning; Extreme weather-related anomalies: These are labeled and retained separately for the model's adaptive training in extreme scenarios.

[0029] This invention uses the Min-max normalization method to normalize power generation data: In the formula: The values ​​to be taken for the input sample; The maximum value of the sample; This is the minimum value in the sample.

[0030] To accurately identify key influencing factors, improve model usability, and provide clear basis for business decisions, avoiding model redundancy or deviation from core requirements due to blindly introducing meteorological data, a random forest algorithm is used to rank the importance of meteorological features such as wind speed and direction, temperature, air pressure, and humidity, denoted as follows: .

[0031] In the formula: It is the impurity of the parent node (using the Gini coefficient). It is the number of child node samples. This is the total number of samples in the parent node, with a default binary split (v=2, left / right child nodes).

[0032] In the formula: It is the total number of trees. It is the first The prediction of the number of trees, It is an indicator function (1 if the condition is met, 0 otherwise).

[0033] In the formula: It is the target meteorological characteristic. It is Split nodes, It is a node Sample size That is the total number of samples.

[0034] The results show that wind power generation is mainly affected by wind speed, while other characteristics are of lower importance. Wind speeds at 10, 30, 50, and 70 m were selected as important meteorological characteristics, while other meteorological characteristics had a relatively small impact and were ignored.

[0035] S2. Wind power generation prediction based on historical data; The mean and variance of wind power generation time series data change over time, making them non-stationary. Therefore, this invention uses the ARIMA model and historical power generation data from wind farms to predict future power generation. The ARIMA model consists of three parts: an autoregressive part that uses past observations to predict future values; a differencing part that transforms the non-stationary power generation time series data into a stationary one to eliminate trend and seasonal effects; and a moving average part that considers a linear combination of errors, adjusting the current forecast based on previous prediction errors.

[0036] The ARIMA model uses differencing to process non-stationary time series data, transforming non-stationary power generation time series data into stationary time series data. Wind power typically exhibits autocorrelation, and the moving average component of the model helps account for the impact of random errors, further improving prediction quality. The structural expression is: In the formula: Let be the autoregressive order of the AR model; Let be the order of the moving average in the MA model; The degree of difference; and These are the autoregressive coefficient and the moving average coefficient, respectively. It is a white noise sequence; This is a non-stationary observation sequence.

[0037] When using the ARIMA model to predict wind power generation, it is necessary to perform stationarity and white noise tests on the power generation time series. The difference operation uses first-order and second-order differences: In the formula: , These are first-order and second-order difference operators, respectively; This is time series data on power generation; This is the shift operator.

[0038] The differencing sequence needs to be tested for stationarity using a unit root test, the formula of which is: In the formula: For constant terms; This is a trend item; For coefficient terms; The slope coefficient; This is the residual term.

[0039] If the test is rejected If the sequence has a unit root, it means that the sequence is non-stationary and needs to be further differencing until the sequence becomes stationary.

[0040] S3. Wind power generation prediction based on numerical weather data; While the ARIMA model can effectively predict wind power generation in the short term, it has limitations when fitting nonlinear time series data, and the differencing process may lead to information loss. To improve the model's nonlinear fitting ability and capture the variation of power generation with weather data, this invention constructs a prediction model based on numerical meteorological data to correct the ARIMA model's power generation predictions. The numerical meteorological data mainly includes temperature, humidity, air pressure, wind, precipitation, cloud cover, visibility, and sunshine duration.

[0041] Compared to traditional attention mechanisms, temporal attention mechanisms place greater emphasis on the relative positions of time points within a sequence, enabling the model to focus on historical data that influences predictions. In this way, temporal attention enhances the model's ability to capture long-term dependencies, thereby improving prediction accuracy.

[0042] In the time attention mechanism, each time step is assigned a weight that reflects its importance to the current time step. The weight is calculated based on the correlation between the state of the current time step and the states of past time steps. The formula for calculating the weight is: In the formula: (query) means query; (key) is the key.

[0043] Because the complex patterns in wind power generation are deeply hidden, traditional LSTM or BiLSTM models are insufficient to achieve ideal prediction performance in long-sequence power generation prediction tasks. Therefore, deeper models are needed to effectively predict power generation.

[0044] The network structure designed in this invention, combining BiLSTM and attention mechanisms, is as follows: Figure 2 As shown: The network structure consists of five components: a batch normalization (BN) layer, a BiLSTM layer, a dropout layer, an attention layer, and a fully connected layer. The fully connected layer comprises two layers: the output dimension of the first fully connected layer (Dense1) is determined by the sliding window size and the number of meteorological features, while the output dimension of the second fully connected layer (Dense2) is 1. To further optimize the model's performance, a BN layer is added after the input layer and the attention layer, and a random deactivation layer (Dropout) is introduced to prevent overfitting.

[0045] Batch normalization (BN) layers normalize the input meteorological feature data, eliminate differences in feature dimensions, accelerate network training convergence, and alleviate the gradient vanishing problem. The dual-layer BiLSTM layer captures the bidirectional time dependence of meteorological sequences. The dual-layer structure further enhances the feature extraction capability and strengthens the capture of complex temporal patterns. During training, the Dropout layer randomly discards some feature nodes from the BiLSTM output to prevent the network from over-relying on specific features, avoid overfitting, and improve the model's generalization ability. The attention layer assigns different weights to the temporal features of the Dropout output, strengthening the meteorological time segments that are key to the prediction results and weakening the influence of irrelevant features; The first fully connected layer (Dense1) integrates the high-dimensional temporal features output by the attention layer with the temporal information corresponding to the sliding window, mapping them into low-dimensional dense features to prepare for the final prediction. The second fully connected layer (Dense2) further maps the integrated features output by Dense1, outputting the final single-value prediction result. The output feature dimension is fixed at 1, corresponding to a single prediction target.

[0046] The residual sequence is obtained by subtracting the actual value from the ARIMA model prediction. Then, the residual sequence and meteorological features are input into the prediction model based on BiLSTM and attention mechanism. At the same time, the PSO algorithm is used to optimize the number of hidden layer neurons and output the residual prediction value of the model. The inverse normalized ARIMA prediction value is added to the residual prediction value of BiLSTM-Attention to obtain the corrected wind power generation prediction value.

[0047] S4, Particle Swarm Optimization (PSO) Hyperparameter Optimization; Specifically, when using the likelihood function for parameter estimation in an ARIMA model, it is necessary to determine the model's order. Traditional methods typically employ the AIC and BIC criteria to determine the autoregressive order of the ARIMA model. and moving average order However, sometimes the result obtained is not globally optimal. The two stacked BiLSTM layers in the two-layer BiLSTM structure are the core hidden layers responsible for temporal feature extraction in the model. This invention uses the particle swarm optimization algorithm to integrate the ARIMA model... and And the number of neurons in the first hidden layer of the BiLSTM model. Number of neurons in the second hidden layer As the particle position information in the PSO algorithm, the optimal parameter combination is obtained after PSO parameter tuning.

[0048] F= In the formula: This represents the fitness value of the particle. E This is the error index value; To minimize, avoid E The mathematical error is that the denominator is 0 when the result is equal to 0.

[0049] In the PSO algorithm, each particle has an initial position and velocity, forming the initial population in the search space. The fitness value, individual extreme value, and global extreme value of each particle's current position are calculated. The PSO algorithm updates each particle's velocity and position based on the individual and global extreme values. This process is repeated until the fitness value no longer changes significantly. (Velocity) and location The updated formula is: In the formula: For particles i In time t speed; For particles i In time t Location; Inertial weight; , All are acceleration constants; , All are random numbers uniformly distributed within [0,1]. , Particles i The individual optimal position and the global optimal position.

[0050] Initialize parameters, determine the range of values ​​for the number of particles, number of iterations, inertia weight, and particle positions, and initialize the position of each particle. and speed The process involves calculating the fitness values ​​of particles, finding individual and swarm extrema, and updating the velocity and position of each particle. Then, the particle fitness values ​​are calculated again, the individual and swarm extrema are updated, and it is determined whether the termination condition is met. Finally, the optimal combination of parameters for the model is determined.

[0051] Save the trained improved PSO-ARIMA-BiLSTM model, use the divided test dataset for model testing, and finally output the model's prediction results.

[0052] Example 2 As another embodiment of the present invention, this embodiment provides a specific experimental example of power generation prediction. For example... Figure 3 As shown, the data period is 1 week. The load is reduced on weekends compared to weekdays. At the same time, the load is higher at the beginning of the month than at other times of the month. The data from the first 28 days is used to train and the prediction result for the 29th day is obtained. The data from the first 29 days is used to train and the prediction result for the 30th day is obtained. And so on, the prediction results for the 7 days from the 29th day to the 35th day are obtained. In the experiment, there are both cases of load abrupt change and prediction of load abrupt change.

[0053] The optimal number of hidden layer neurons in the BiLSTM was obtained after optimization using the PSO algorithm. The residual sequence data and numerical meteorological feature data were used as inputs to the second-stage BiLSTM-Attention model, which outputs predicted values ​​for the residual sequence. The predicted values ​​from the ARIMA model and the BiLSTM-Attention model were then added to obtain the corrected power generation prediction. The prediction results are as follows: Figure 4 As shown.

[0054] The prediction accuracy of various models on training datasets of different sizes is shown in Table 1.

[0055] Table 1. Prediction results of this invention and various benchmark models It can be seen that the PSO-BiLSTM proposed in this invention is the best performing of the five models, and it outperforms the prediction models based on single historical power generation data and single numerical meteorological data in terms of prediction accuracy.

[0056] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0057] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A two-stage wind power generation prediction method considering meteorological factors, characterized in that the steps include... include: Acquire historical power generation data and meteorological data, and rank the meteorological features by importance; Based on historical power generation data, an ARIMA-based wind power generation prediction model is used to predict the power generation time series for future moments. The residual sequence is obtained by subtracting the actual value from the predicted value of the power generation time series. The residual sequence and the sorted meteorological features are then input into the residual prediction model based on BiLSTM and attention mechanism, and the model's residual prediction value is output. The corrected wind power generation forecast is obtained by adding the inverse normalized time series forecast value to the residual forecast value.

2. The two-stage wind power generation prediction method considering meteorological factors according to claim 1, characterized in that, The meteorological features are ranked by importance using a random forest algorithm. The meteorological features include wind speed and direction, temperature, air pressure, and humidity.

3. The two-stage wind power generation prediction method considering meteorological factors according to claim 1, characterized in that, The ARIMA-based wind power generation prediction model includes: an autoregressive component that uses past observations to predict future values; a differencing component that converts non-stationary power generation time series data into stationary time series data through differencing; and a moving average component that considers a linear combination of errors and adjusts the current prediction based on previous prediction errors.

4. The two-stage wind power generation prediction method considering meteorological factors according to claim 3, characterized in that, The ARIMA-based wind power generation prediction model performs stationarity and white noise tests on the power generation time series when predicting wind power generation.

5. The two-stage wind power generation prediction method considering meteorological factors according to claim 3, characterized in that, The ARIMA-based wind power generation prediction model uses a unit root test to check the stationarity of the differencing time series. If the test is rejected, the series is further differencing until it becomes stationary.

6. The two-stage wind power generation prediction method considering meteorological factors according to claim 1, characterized in that, The residual prediction model based on BiLSTM and attention mechanism includes an input layer, a two-layer BiLSTM layer, an attention layer, and a fully connected layer connected in sequence; a batch normalization layer is added after the input layer and the attention layer; a random deactivation layer is introduced after the BiLSTM layer to prevent overfitting; the fully connected layer includes two layers: the output dimension of the first fully connected layer is determined by the sliding window size and the number of meteorological features, and the output dimension of the second fully connected layer is 1. The first normalization layer normalizes the meteorological characteristic data of the input layer to eliminate differences in characteristic dimensions; The dual-layer BiLSTM layer captures the bidirectional time dependence of meteorological sequences, enhancing the capture of complex time series patterns; The random deactivation layer randomly discards some feature nodes from the output of the two-layer BiLSTM layer during training; The attention layer assigns different weights to the temporal features output by the random deactivation layer, strengthening the meteorological time segments that are key to the prediction results and weakening the influence of irrelevant features. The first fully connected layer integrates the high-dimensional temporal features output by the attention layer with the temporal information corresponding to the sliding window, and maps them into low-dimensional dense features. The second fully connected layer further maps the integrated features output by the first fully connected layer, outputting the final single-value prediction result.

7. The two-stage wind power generation prediction method considering meteorological factors according to claim 1, characterized in that, The method jointly optimizes the autoregressive order and moving average order of the ARIMA model, as well as the number of neurons in the two sequentially stacked BiLSTM layers in the two-layer BiLSTM structure.

8. The two-stage wind power generation prediction method considering meteorological factors according to claim 7, characterized in that, The joint optimization employs a particle swarm optimization algorithm, using the autoregressive order and moving average order of the ARIMA model, as well as the number of neurons in the first and second hidden layers of the bilayer BiLSTM residual prediction model, as particle position information. The optimization process is as follows: The position and velocity of each particle are initialized, and the particles constitute the initial population in the search space; Calculate the fitness value, individual extreme value, and global extreme value for each particle at its current position; Update the velocity and position of each particle based on individual and global extrema; Repeat the above process until the fitness value no longer changes, thus obtaining the optimal combination of parameters for the model.

9. A two-stage wind power generation prediction device considering meteorological factors, comprising a memory, a processor, and a program stored in the memory, characterized in that, When the processor executes the program, it implements the two-stage wind power generation prediction method considering meteorological factors as described in any one of claims 1-8.

10. A storage medium having a program stored thereon, characterized in that, When the program is executed, it implements the two-stage wind power generation prediction method that takes meteorological factors into account as described in any one of claims 1-8.

Images