Wind power prediction method based on non-overlapping multi-scale division and interactive modeling

Abstract

The invention is applied to the technical field of new energy power generation prediction, and discloses a wind power prediction method based on non-overlapping multi-scale division and interactive modeling, and the method comprises the specific steps: firstly, carrying out the space-time embedding of historical wind power data, and obtaining the space information; secondly, a multi-scale token length set is set, non-overlapping multi-scale partitioning is carried out on the embedded sequence, and tokens of different scales are generated so as to reduce redundancy caused by overlapping; then, constructing a hierarchical interactive encoder, extracting features layer by layer from a small scale to a large scale, and embedding the small-scale potential representation of the previous layer into the input of the current layer in the large-scale encoding process to realize the guidance of the large scale by the small-scale features; and finally, fusing and mapping each scale feature output to obtain a future wind power prediction value, and training and optimizing the model by adopting mean square error loss. Verification on actual wind power plant data shows that the method can effectively improve the accuracy of wind power prediction.

Description

Technical Field

[0001] This invention relates to the field of new energy power generation prediction technology, specifically a wind power prediction method based on non-overlapping multi-scale partitioning and interactive modeling. Background Technology

[0002] Faced with the depletion of fossil fuels and the deterioration of the ecological environment, clean energy has become a core strategy for countries to cope with the energy crisis and protect the environment. As a renewable and pollution-free energy source, wind energy has important value in replacing traditional energy sources, alleviating energy shortages, and reducing environmental pollution. The Global Wind Energy Association's "Global Wind Energy Report 2024" shows that 117 gigawatts of new wind power capacity were added globally in 2024, and the global installed wind power capacity has exceeded 1 terawatt. However, the inherent volatility and intermittency of wind power output can easily lead to unstable output power of wind farms, posing a severe challenge to the optimized scheduling of the power system and grid connection. Therefore, accurate prediction of wind power output is crucial to ensuring the stable and safe operation of the power system and reducing the risks brought about by the uncertainty of wind energy. Wind power forecasting aims to predict the output power value or range of wind farms in the future, providing a scientific basis for power system dispatch. Existing forecasting methods can be roughly divided into three categories: physical methods, statistical methods, and machine learning methods. Physical methods are based on fluid dynamics models, combined with numerical weather forecast data such as wind speed and air pressure, as well as geographical information such as ground roughness and altitude, to predict wind power. They are usually suitable for medium- and long-term forecasts. Statistical methods identify patterns in historical data through probability theory and mathematical statistics. Typical methods mainly include the Autoregressive Moving Average (ARMA) model and the Seasonal Autoregressive Moving Average (SARMA) model. In recent years, machine learning has received widespread attention from renewable energy forecasting researchers due to its strong robustness and portability, and has gradually become an important method to address the challenges of wind power fluctuation forecasting. The Transformer model, with its multi-head attention mechanism and encoder structure, significantly improves the ability to model time series features and has also gradually become an important method in wind power forecasting tasks. Most existing wind power prediction methods rely on fixed time scales for data modeling. They perform well under stable weather conditions, but their prediction accuracy drops significantly when faced with complex climate fluctuations and sudden weather events. To address this challenge, researchers have proposed multi-scale prediction methods that improve model robustness by fusing data from multiple time resolutions. However, current multi-scale methods still have two major drawbacks: (1) data redundancy, where cross-scale information duplication leads to increased computation and storage overhead; and (2) insufficient interaction, where features are processed independently between scales and there is a lack of collaborative mechanisms, which limits the ability to capture nonlinear spatiotemporal correlations. Therefore, a wind power forecasting method based on non-overlapping multi-scale and hierarchical interaction modeling (Multi-Scale Hierarchical Interaction Transformer, MSHIT) is proposed. This model captures short-term patterns and long-term trends through multi-scale modeling, significantly improving the comprehensiveness and accuracy of the forecast. Specifically, MSHIT adopts a hierarchical architecture, modeling time dependencies layer by layer from small time scales to large time scales. At the same time, it integrates complementary information from multiple scales through a hierarchical interaction mechanism to enhance the model's forecasting performance and robustness. In addition, the non-overlapping multi-scale partitioning strategy effectively avoids data redundancy and significantly improves computational efficiency. Experiments show that MSHIT not only performs excellently in wind power forecasting tasks, solving the problems of data redundancy and insufficient interaction in traditional multi-scale methods, but also provides new technical ideas for other time series forecasting tasks. Summary of the Invention

[0003] The purpose of this invention is to provide a wind power prediction method based on non-overlapping multi-scale partitioning and interactive modeling, so as to solve the problems of large data redundancy, insufficient feature interaction between different scales, and poor prediction accuracy and stability in the existing wind power prediction methods proposed in the background art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a wind power prediction method based on non-overlapping multi-scale partitioning and interactive modeling, the prediction method comprising the following steps: Step 1: Acquisition of historical wind power data and construction of dataset: Collect historical operation data of wind farms, preprocess the historical data, and construct a wind power prediction modeling dataset according to the prediction duration requirements; Step 2, Spatiotemporal Feature Embedding Process: Spatiotemporal feature embedding processing is performed on the modeling dataset. While maintaining the time order and variable correlation, joint feature mapping is performed on the multivariate time series. Step 3: Non-overlapping multi-scale time series partitioning: Set multiple time scale parameters to perform multi-scale partitioning on the embedded time series; Step 4: Multi-scale hierarchical interactive feature modeling: Construct a multi-scale hierarchical interactive Transformer network, model in order of increasing time scale, and introduce a cross-scale interaction mechanism in the modeling process of larger time scales, so that the features extracted at smaller time scales can participate in and guide the feature learning at larger time scales. Step 5: Multi-scale feature fusion and prediction output: The feature representations obtained at each time scale are fused to form a unified multi-scale feature expression. Based on the fused features, wind power prediction results for one or more future times are generated. Step 6: Model training and prediction result output: Train the model with historical wind power real values. After the model training is completed, evaluate the wind power prediction performance using test data.

[0005] Furthermore, in step 1, the historical wind farm operation data includes two years of historical data, including time-series data on wind power, wind speed, and wind direction. The historical operation data is then divided according to chronological order. , , ,in , For characteristic number, For time step, The predictor variable is the actual value of wind power, and the data is divided into segments according to time sequence. , Two years of data, As a training set The data from the first quarter is the validation set. The data for the remaining three quarters are for the test set. , , .

[0006] Furthermore, in step 2, the temporal embedding method focuses on feature extraction at the time step, which can easily limit the modeling ability of time series at multiple scales. In contrast, we adopt a spatiotemporal embedding method, which can maintain the integrity of the spatial (variable) dimension and the temporal (time step) dimension. Then, by dividing non-overlapping tokens, we achieve the learning of multi-scale temporal dependencies, thereby effectively solving the problem of spatiotemporal information processing in wind power prediction. Assuming the input time series data is... , It is the length of the time series. The feature dimension at each time point, after spatiotemporal embedding, is:

[0007] in, It is a convolutional network, obtained after the convolution operation. It is the embedded feature map. It is the number of channels in the embedded feature map.

[0008] Furthermore, in step 3, the multi-scale data partitioning method can better model these complex time dependencies, thereby significantly improving the performance of the prediction model. To address the data redundancy problem during scale partitioning, this invention employs non-overlapping partitioning: there is no overlap between each token, thus ensuring that adjacent tokens are completely independent. Tokenization divides the time series into segments of length [missing information]. of Non-overlapping blocks;

[0009] Historical Time Series Length is The token length is The non-overlapping block method divides the time series into independent data blocks, reducing data redundancy and computational load. The output of the multi-scale non-overlapping time series partitioning module is as follows: ; Each token contains only an independent time block, prompting the model to extract features within each time window. This approach helps improve the robustness and generalization ability of the model in time series data processing, and is particularly suitable for prediction tasks that need to capture long-term dependencies.

[0010] Furthermore, in step 4, within the hierarchical interaction network, we employ multiple encoders to extract features from data at different scales. The input to each encoder is data that has undergone spatiotemporal embedding and non-overlapping partitioning. , This indicates the level at which the Encoder operates, where The Encoder primarily utilizes the classic Transformer-encode multi-head attention mechanism to process the input data. The calculation process of the attention mechanism is shown below:

[0011]

[0012]

[0013] in, , and These represent the query features, key features, and value features generated from the input features, respectively, used to characterize the correlation relationships between different positions in a time series. This is the dimension used to represent the attention features, and it's used to scale the results of feature correlation calculations to prevent excessively large values ​​in the high-dimensional feature space from causing issues. Vanishing gradients improve the stability of model training;

[0014] This represents the process of weighted aggregation of value features based on feature correlation. Through this attention mechanism, the Encoder can adaptively highlight the time segment features that are more important to the current prediction task and suppress redundant information, thereby obtaining an effective temporal feature representation at the current scale. The output of the Encoder serves as the intermediate feature representation of the hierarchical interactive network for subsequent multi-scale feature fusion and wind power prediction.

[0015] Furthermore, in step 5, the feature representations obtained at different time scales are fused to form a unified multi-scale feature expression. Based on the fused features, wind power prediction results for one or more future time periods are generated to simultaneously consider the short-term fluctuation characteristics and long-term trends of the wind power sequence. Following the hierarchical feature extraction method described above, the latent scale information learned at each level is relatively independent. To achieve information interaction between different time scales, a cross-scale guidance mechanism is introduced in the encoder: the encoder output at the small-scale level is embedded into the encoder input at the large-scale level, thereby enabling small-scale features to guide large-scale modeling. In the layer encoder (when) (Time), spatiotemporal embedding features of this layer Compared with the output of the previous encoder layer The data is concatenated along the channel dimension and then fused using a 1x1 convolutional layer to obtain the input representation at the current scale.

[0016] in, This represents a concatenation operation along the channel dimension. Conv is a 1x1 convolutional layer that fuses features and reduces the number of channels from 2c to c. This operation ensures that the fine dynamics captured in shorter time scales can guide modeling in larger time scales, effectively improving the model's cross-scale modeling capability. At this time, the current scale level is the smallest, so there is no need for scale interaction. Spatiotemporal embedding is used directly as the encoder input.

[0017] Finally, The latent representations of the 1-level encoder are concatenated and passed through a convolutional layer to generate the predicted value, as shown below:

[0018]

[0019] in, The convolutional layer represents the concatenation of potential outputs from each level, mapping them to predicted values ​​to generate the prediction result. This process of inferring predictions from multi-scale latent representations can effectively utilize cross-scale temporal dependencies, thereby ultimately improving prediction accuracy.

[0020] Furthermore, in step 6, the optimal structure is trained and tuned using the mini-batch stochastic gradient descent algorithm. After convergence, the model performance is obtained on the test set, and supervised learning is used to train the model and optimize the parameters.

[0021] After the model is trained, the prediction accuracy and stability of the prediction model under different prediction durations are evaluated using a test set, and the final wind power prediction results are output.

[0022] Compared with existing technologies, the beneficial effects of this invention are: the wind power prediction method based on non-overlapping multi-scale partitioning and interactive modeling: A hierarchical interactive prediction framework with multi-scale non-overlapping data partitioning and cross-scale guidance is proposed. This framework reduces computational redundancy while enhancing the multi-scale feature fusion capability, thereby improving the accuracy and efficiency of wind power prediction. The specific steps are as follows: First, historical wind power data is spatiotemporally embedded to obtain spatial information. Secondly, a multi-scale token length set is set, and non-overlapping multi-scale block division is performed on the embedded sequence to generate tokens of different scales, so as to reduce redundancy caused by overlap. Then, a hierarchical interactive encoder is constructed to extract features layer by layer from small scale to large scale, and during the large-scale encoding process, the latent representation of the previous small scale layer is embedded into the input of the current layer to realize the guidance of small-scale features on large scale. Finally, the feature outputs at each scale are fused and mapped to obtain the predicted future wind power value. The mean squared error loss is used to train and optimize the model. The present invention has been verified on actual wind farm data, which shows that the method can effectively improve the accuracy of wind power prediction and has good engineering application value. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the implementation framework of the MSHIT invention; Figure 2 This is a diagram of the hierarchical interactive network structure of the present invention. Detailed Implementation

[0024] 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.

[0025] Please see Figures 1-2 This invention provides a technical solution: a wind power prediction method based on non-overlapping multi-scale partitioning and interactive modeling.

[0026] Example 1 This embodiment uses real wind farm operation data to verify the wind power prediction method based on non-overlapping multi-scale hierarchical interactive modeling proposed in this invention; Step 1: Data Acquisition and Modeling Dataset Construction. In this embodiment, the actual wind power data comes from the International New Energy Laboratory dataset. The operating data of a wind turbine with a rated capacity of 16MW is selected, and the data time span is from 2011 to 2012. The original data sampling interval is 5 minutes. After downsampling the original data, the sampling interval is adjusted to 15 minutes, and approximately 70,000 valid samples are obtained. Historical wind speed, wind direction, and wind power are selected as model input variables, and future wind power is used as the predicted output variable. In this embodiment, the historical input window length of the model is fixed at 4 hours, that is, each sample consists of multivariate time series data from the past 4 hours, which is used to predict wind power at different future time points. According to the time order, the modeling dataset is divided into three parts: training set, validation set, and test set, as shown in Table 1. Table 1 shows the dataset partitioning used in the embodiments.

[0027] Step 2: Spatiotemporal feature embedding processing. The modeling dataset obtained in Step 1 is subjected to spatiotemporal feature embedding processing. In this embodiment, each input sample corresponds to 4 hours of historical data, which includes wind speed, wind direction and power information of multiple consecutive time steps. The time dimension and variable dimension information are jointly mapped by a convolutional neural network. While keeping the original time order unchanged, the features of the multivariate time series are enhanced so that the relevant information between wind speed, wind direction and power can be more fully expressed in the feature space, providing a unified feature representation for subsequent multi-scale modeling. Step 3: Non-overlapping multi-scale time series partitioning. For a fixed 4-hour historical input window, this embodiment sets three time scales: 1 hour, 2 hours, and 4 hours, and performs multi-scale partitioning on the embedded time series: 1-hour scale: used to characterize the rapid fluctuations in wind power over a short period of time; 2-hour scale: used for modeling trends of change over a medium time range; 4-hour scale: used to extract overall change patterns within a complete historical window; At each time scale, the time series is divided using a non-overlapping time block method, that is, the 4-hour historical series is divided into several data blocks of different lengths but independent of each other, ensuring that there is no overlapping area between adjacent data blocks; Step 4: Multi-scale hierarchical interactive coding modeling. Construct a multi-scale hierarchical interactive coding network to extract features from data blocks at different time scales. The network structure is organized according to the time scale from small to large. At the 1-hour scale, the focus is on extracting the local fluctuation characteristics of wind power. At the 2-hour and 4-hour scale layers, a cross-scale interaction mechanism is introduced to fuse the features extracted from the smaller time scale layer with the input features at the current scale, so that short-term detailed information can guide the modeling of medium- and long-term trends. Step 5: Multi-scale feature fusion and prediction output. The feature representations output by the coding layers at each time scale are fused to form a unified multi-scale feature representation. Based on the fused feature representation, a prediction output module is constructed to generate wind power prediction results 1 hour, 2 hours, and 4 hours in advance, respectively. Step 6: Model training and performance evaluation. The model is trained using supervised learning, with historical wind power values ​​as the supervision signal. The model parameters are iteratively optimized. After the model training is completed, the prediction performance of the model is evaluated using a test set under three prediction durations: 1 hour, 2 hours, and 4 hours in advance.

[0028] This invention compares the following methods: (1) LSTM: Using LSTM networks to predict wind power; (2) Transformer: Using Transformer networks to predict wind power; (3) iTransformer: A redesigned Transformer architecture that improves the performance of time series forecasting without modifying the basic components by applying attention and feedforward networks to the inverted dimension. (4)Formertime: A multi-scale time series prediction method that effectively solves the computational efficiency bottleneck of traditional Transformer when processing multi-scale data by parallelizing the Temporal Reduction Attention layer and the multi-resolution feature extraction module. (5)MSGNet: A multi-scale time series prediction method that captures complex correlations between sequences at multiple time granularities through an adaptive scale selection mechanism. This model demonstrates excellent generalization ability and prediction accuracy in multivariate prediction tasks. (6)MTST: A multi-scale time series forecasting method that uses a multi-branch architecture to allow simultaneous learning of time series representations at different scales, thereby improving the performance of time series forecasting. (7) MSHIT: The multi-scale method proposed in this invention;

[0029] To ensure fairness in the comparison, we thoroughly tuned the main hyperparameters of all the above models and selected their optimal configurations for testing on the validation set. The experimental method of this invention uses the coefficient of determination. The root mean square error (RMSE) is a performance evaluation index, and its calculation formula is: ; ; in, The number of samples in the test set. , These are the predicted value and the actual value of the i-th sample, respectively; The effectiveness and superiority of the proposed method were verified using real-world wind farm datasets. The experimental results are shown in Table 2. Table 2 Comparison of prediction performance of different prediction methods

[0030] This experiment compared the prediction performance of various methods at three different time points (1 hour, 2 hours, and 4 hours). The experimental results are shown in Table 1. Prediction performance was evaluated using two metrics: RMSE and R². Overall, different methods showed significant differences in performance at different time points. Some methods, such as LSTM, performed reasonably well in short-term prediction (1 hour), but their performance gradually declined over time, especially in long-term prediction (4 hours), where RMSE increased significantly and R² decreased significantly, indicating that this method has weak long-term prediction ability. Furthermore, while Transformer and its variant iTransformer maintained performance stability to some extent, they did not exhibit optimal performance at all time points. While some multi-scale methods (such as Formertime, MSGNet, and MTST) perform reasonably well at certain time points, their overall prediction accuracy and stability are relatively low, and their performance declines significantly over time. The method proposed in this invention demonstrates superior prediction performance at most time points, with relatively low RMSE and relatively high R². From the perspective of the impact of time points, the performance of most methods declines as the prediction time increases. This may be due to factors such as accumulated errors in long-term predictions, data instability, and the model's insufficient ability to capture long-term dependencies. Our proposed method can overcome these difficulties to some extent and maintain good prediction performance and stability.

[0031] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A wind power prediction method based on non-overlapping multi-scale partitioning and interactive modeling, characterized in that: Includes the following steps: Step 1: Acquisition of historical wind power data and construction of dataset: Collect historical operation data of wind farms, preprocess the historical data, and construct a wind power prediction modeling dataset according to the prediction duration requirements; Step 2, Spatiotemporal Feature Embedding Process: Spatiotemporal feature embedding processing is performed on the modeling dataset. While maintaining the time order and variable correlation, joint feature mapping is performed on the multivariate time series. Step 3: Non-overlapping multi-scale time series partitioning: Set multiple time scale parameters to perform multi-scale partitioning on the embedded time series; Step 4: Multi-scale hierarchical interactive feature modeling: Construct a multi-scale hierarchical interactive Transformer network, model in order of increasing time scale, and introduce a cross-scale interaction mechanism in the modeling process of larger time scales, so that the features extracted at smaller time scales can participate in and guide the feature learning at larger time scales. Step 5: Multi-scale feature fusion and prediction output: The feature representations obtained at each time scale are fused to form a unified multi-scale feature expression. Based on the fused features, wind power prediction results for one or more future times are generated. Step 6: Model training and prediction result output: Train the model with historical wind power real values. After the model training is completed, evaluate the wind power prediction performance using test data.

2. The wind power prediction method based on non-overlapping multi-scale partitioning and interactive modeling according to claim 1, characterized in that: In step 1, the historical wind farm operation data includes two years of historical data, including time-series data on wind power, wind speed, and wind direction. The historical operation data is divided according to chronological order. , , ,in , For characteristic number, For time step, The predictor variable is the actual value of wind power, and the data is divided into segments according to time sequence. , Two years of data, As a training set The data from the first quarter is the validation set. The data for the remaining three quarters are for the test set. , , .

3. The wind power prediction method based on non-overlapping multi-scale partitioning and interactive modeling according to claim 1, characterized in that: In step 2, the temporal embedding method focuses on feature extraction at the time step, which can limit the modeling ability of time series at multiple scales. In contrast, we use a spatiotemporal embedding method, which maintains the integrity of both spatial and temporal dimensions. Furthermore, by partitioning non-overlapping tokens, we learn the temporal dependencies at multiple scales, effectively solving the problem of spatiotemporal information processing in wind power forecasting. Assuming the input time series data is... , It is the length of the time series. The feature dimension at each time point, after spatiotemporal embedding, is: ； in, It is a convolutional network, obtained after the convolution operation. It is the embedded feature map. It is the number of channels in the embedded feature map.

4. The wind power prediction method based on non-overlapping multi-scale partitioning and interactive modeling according to claim 1, characterized in that: In step 3, the multi-scale data partitioning method can better model these complex time dependencies, thereby significantly improving the performance of the prediction model. To solve the data redundancy problem during scale partitioning, this invention adopts non-overlapping partitioning: there is no overlap between each token, thus ensuring that adjacent tokens are completely independent. Tokenization divides the time series into lengths of [missing information]. of Non-overlapping blocks; ； Historical Time Series Length is The token length is The non-overlapping block method divides the time series into independent data blocks, reducing data redundancy and computational load. The output of the multi-scale non-overlapping time series partitioning module is as follows: ； Each token contains only an independent time block, prompting the model to extract features within each time window. This approach helps improve the robustness and generalization ability of the model in time series data processing, and is particularly suitable for prediction tasks that need to capture long-term dependencies.

5. The wind power prediction method based on non-overlapping multi-scale partitioning and interactive modeling according to claim 1, characterized in that: In step 4, within the hierarchical interaction network, we employ multiple encoders to extract features from data at different scales. The input to each encoder is data that has undergone spatiotemporal embedding and non-overlapping partitioning. , This indicates the level at which the Encoder operates, where The Encoder primarily utilizes the classic Transformer-encode multi-head attention mechanism to process the input data. The calculation process of the attention mechanism is shown below: ； ； ； in, , and These represent the query features, key features, and value features generated from the input features, respectively, used to characterize the correlation relationships between different positions in a time series. This is the dimension used to represent the attention features, and it's used to scale the results of feature correlation calculations to prevent excessively large values ​​in the high-dimensional feature space from causing issues. Vanishing gradients improve the stability of model training; This represents the process of weighted aggregation of value features based on feature correlation. Through this attention mechanism, the Encoder can adaptively highlight the time segment features that are more important to the current prediction task and suppress redundant information, thereby obtaining an effective temporal feature representation at the current scale. The output of the Encoder serves as the intermediate feature representation of the hierarchical interactive network for subsequent multi-scale feature fusion and wind power prediction.

6. The wind power prediction method based on non-overlapping multi-scale partitioning and interactive modeling according to claim 1, characterized in that: In step 5, the feature representations obtained at different time scales are fused to form a unified multi-scale feature expression. Based on this fused feature, wind power prediction results for one or more future time periods are generated to simultaneously consider the short-term fluctuation characteristics and long-term trends of the wind power sequence. Following the hierarchical feature extraction method described above, the latent scale information learned at each level is relatively independent. To achieve information interaction between different time scales, a cross-scale guidance mechanism is introduced in the encoder: the encoder output of the small-scale level is embedded into the encoder input of the large-scale level, thereby enabling small-scale features to guide large-scale modeling. In the layer encoder (when) (Time), spatiotemporal embedding features of this layer Compared with the output of the previous encoder layer The data is concatenated along the channel dimension and then fused using a 1x1 convolutional layer to obtain the input representation at the current scale. ； in, This represents a concatenation operation along the channel dimension. Conv is a 1x1 convolutional layer that fuses features and reduces the number of channels from 2c to c. This operation ensures that the fine dynamics captured in shorter time scales can guide modeling in larger time scales, effectively improving the model's cross-scale modeling capability. At this time, the current scale level is the smallest, so there is no need for scale interaction. Spatiotemporal embedding is used directly as the encoder input. Finally, The latent representations of the 1-level encoder are concatenated and passed through a convolutional layer to generate the predicted value, as shown below: ； ； in, The convolutional layer represents the concatenation of potential outputs from each level, mapping them to predicted values ​​to generate the prediction result. This process of inferring predictions from multi-scale latent representations can effectively utilize cross-scale temporal dependencies, thereby ultimately improving prediction accuracy.

7. The wind power prediction method based on non-overlapping multi-scale partitioning and interactive modeling according to claim 1, characterized in that: In step 6, the optimal structure is trained and tuned using the mini-batch stochastic gradient descent algorithm. After convergence, the model performance is obtained on the test set. Supervised learning is then used to train the model and optimize the parameters. After the model is trained, the prediction accuracy and stability of the prediction model under different prediction durations are evaluated using a test set, and the final wind power prediction results are output.

Images