A combined LSTM-Transformer model for medium- and long-term wind power forecasting and its system

Abstract

This invention discloses a method and system for a combined LSTM-Transformer model for medium- and long-term wind power prediction, comprising: acquiring and preprocessing meteorological data of wind farms; using a generative adversarial network to capture the intrinsic features of the original dataset and generating synthetic historical weather data samples to augment the existing dataset; performing empirical mode decomposition on the augmented dataset to obtain intrinsic mode components at different time scales, and performing variational mode decomposition on the augmented dataset to obtain wind turbine power trend sequences and periodic sequences; constructing a SwinLSTM network model, training the SwinLSTM network model using data obtained from signal decomposition, and predicting the comprehensive environmental data of wind turbine locations and the wind power at wind turbine locations, respectively; constructing an adaptive weight adjustment method based on deep reinforcement learning flexible action evaluation, and adaptively adjusting the theoretical wind power and predicted wind power to obtain the final predicted wind power.

Description

Technical Field

[0001] This invention relates to the field of wind power prediction technology, and in particular to a method and system for a combined LSTM-Transformer medium- and long-term wind power prediction model. Background Technology

[0002] With the continued growth of global demand for clean energy, wind power, as an important component of renewable energy, is gradually taking a dominant position in energy production. However, the characteristics of wind power, including its uncertainty and volatility, pose a challenge to the stable operation of the power system. Against this backdrop, accurately predicting wind power output has become one of the key factors for optimizing grid dispatch and achieving efficient energy management.

[0003] While traditional forecasting methods have achieved some success, they still exhibit limitations when faced with the complex influence of wind speed, wind direction, and other environmental factors. Therefore, leveraging advanced data analysis techniques and machine learning algorithms is becoming a new direction for improving the accuracy of wind power forecasting. A major challenge in using deep learning methods for forecasting is the lack of sufficient datasets. Data-driven methods are prone to overfitting to the training dataset, which directly reduces the prediction accuracy of neural networks. This is especially true for medium- to long-term forecasts, where the longer forecast timeframe requires a larger dataset, and insufficient data significantly impacts the accuracy of wind farm power generation predictions. Generative Adversarial Networks (GANs) provide a solution for data-driven algorithms by enhancing the training dataset, effectively avoiding overfitting. Furthermore, GANs can repair and enhance data; by learning the distribution characteristics of the data, GANs can recover information from damaged or incomplete data, improving data integrity and quality. Using GAN networks to process data helps us explore the structure and characteristics of wind power datasets more deeply, thereby enhancing our understanding and analytical capabilities regarding the relationship between wind power and environmental data. Signal decomposition techniques also have a significant impact on wind power forecasting using deep learning. It can effectively process complex wind power time series data. Furthermore, signal decomposition methods help in the in-depth analysis of the dynamic changes and periodic characteristics within the wind power data. Building upon this, the most crucial aspect is the selection of the prediction model; a reasonable and effective prediction strategy can often significantly improve prediction accuracy.

[0004] Existing research on ensemble prediction models consists of two stages: first, selecting some sub-models, and then generating optimal weights for the sub-models based on a given training dataset. Most existing ensemble prediction methods use a fixed set of weights to merge the results of the sub-models. In recent years, reinforcement learning, as one of the hottest technologies in the field of artificial intelligence, has demonstrated powerful capabilities in multiple areas. Reinforcement learning agents, through trial and error and feedback mechanisms, can continuously optimize decision-making strategies, improving system performance and efficiency. Summary of the Invention

[0005] In view of the aforementioned existing problems, the present invention is proposed.

[0006] Therefore, in view of the problem that existing technologies have less consideration of historical meteorological factors and lower prediction accuracy in medium and long-term wind power forecasting, this invention provides an LSTM-Transformer medium and long-term wind power forecasting combined model method and system, which solves the problem of insufficient historical wind power meteorological data and improves prediction accuracy.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0008] In a first aspect, this invention provides a combined LSTM-Transformer model for medium- and long-term wind power prediction, comprising: acquiring and preprocessing meteorological data of wind farms; using a generative adversarial network to capture the intrinsic features of the original dataset and generating synthetic historical weather data samples to augment the existing dataset; performing empirical mode decomposition on the augmented dataset to obtain intrinsic mode components at different time scales, and performing variational mode decomposition on the augmented dataset to obtain wind turbine power trend sequences and periodic sequences; constructing a SwingLSTM network model, training the SwingLSTM network model using data obtained from signal decomposition, and predicting the integrated environmental data of wind turbine locations and the wind power at wind turbine locations, respectively; constructing an adaptive weight adjustment method based on deep reinforcement learning flexible action evaluation, and adaptively adjusting the theoretical wind power and predicted wind power to obtain the final predicted wind power.

[0009] As a preferred embodiment of the LSTM-Transformer long-term wind power forecasting combined model method described in this invention, the step of generating synthetic historical weather data samples to expand the existing dataset includes:

[0010] Embedded networks are used to capture features of wind power time series data and map them into a low-dimensional latent space, preserving key information of the time series.

[0011] Construct a generator G and a discriminator D, where G(z) represents the generator function that maps the latent space vector z sampled from the standard normal distribution to the data space, and D(x) represents the probability that the output x of the discriminator network comes from the training data rather than the generator.

[0012] The segmented data is input into DCGAN for training. Real wind power data is labeled as 1, and fake data is generated and labeled as 0 for adversarial training. The performance of the discriminator is measured by binary cross-entropy loss.

[0013] The generated data distribution is compared with the real data distribution, and the Kullback-Leibler divergence is used for evaluation. The model is then fine-tuned using real wind power generation data.

[0014] As a preferred embodiment of the LSTM-Transformer long-term wind power prediction combined model method described in this invention, the step of performing empirical mode decomposition on the augmented dataset to obtain intrinsic mode components at different time scales includes:

[0015] For an original wind power sequence x(i), find all the maximum points in the sequence and determine all the maximum points as the upper envelope, and find all the minimum points in the sequence and determine all the minimum points as the lower envelope. m(i) represents the average of the upper and lower envelopes. The first component is h1(i) = x(i) - m(i). In the next round of screening, h1(i) is regarded as the original data, and m1(i) is the average of the upper and lower envelopes of h1(i), thus determining the second component h2(i).

[0016] The screening process is repeated n times until h is reached. n (i) is an intrinsic mode function or residual component r n (i) is a monotonic function, which terminates the decomposition process;

[0017] Specify q1(i) = h1(i), q2(i) = h2(i), ..., q k (i)=h k x(i) is ultimately decomposed into n eigenmode components q. t (i) and a remaining component r n (i).

[0018] As a preferred embodiment of the LSTM-Transformer long-term wind power forecasting combined model method described in this invention, wherein: the step of performing variational mode decomposition on the augmented dataset to obtain the wind turbine power trend sequence and periodic sequence includes:

[0019] The wind power data f is decomposed into i components. The decomposition sequence is preserved as a modal component with a medium frequency and finite bandwidth. At the same time, the sum of the estimated bandwidths of each mode is minimized. The constraint condition is that the sum of all modes is equal to the original data. The constrained variational problem is transformed into an unconstrained variational problem.

[0020] The algorithm continuously updates each component and its frequency using the alternating direction multiplier iterative algorithm to obtain the saddle point of the unconstrained model, which is the optimal solution of the original problem. The algorithm re-estimates the repetition frequency based on the power spectrum repetition frequency of each component and initializes the parameters.

[0021] The execution cycle is n = n + 1. When the frequency ω > 0, each component is updated. The steps are repeated until the stopping condition is met.

[0022] As a preferred embodiment of the LSTM-Transformer long-term wind power forecasting combined model method described in this invention, the construction of the SwingLSTM network model includes:

[0023] The construction of the STB module includes using window-based multi-head self-attention technology to reduce the number of elements involved in a single computation, capturing contextual information by moving the window position at different levels, and capturing global spatial dependencies by inputting the input sequence into the normalization layer and multilayer perception mechanism.

[0024] The construction of a SwinLSTM Cell based on the STB module includes capturing temporal dependencies by introducing convolution operators into the input-to-state and state-to-state transitions, calculating similarity scores between all positions through a self-attention mechanism to capture global spatial dependencies, and merging the input gate, forget gate, and output gate into a single filter gate. Combining the temporal dependencies and global spatial dependencies, the system updates both the unit state and the hidden state to capture long-term and short-term temporal dependencies.

[0025] Constructing a SwinLSTM network model using the SwinLSTM Cell as the core frequency includes dividing wind power data at time t into non-overlapping data blocks, flattening the data blocks and inputting them into a patch embedding layer, and the SwinLSTM layer receiving the transformed image patch and hidden state H. t-1 and unit state C t-1 To generate hidden state H t and unit state C t H t The image is copied into two copies, one for rebuilding the layer, and the other for use with C. t Together they are used in the SwinLSTM layer for the next time step;

[0026] The SwinLSTM network model is used to process wind power data to obtain predicted wind turbine power generation.

[0027] As a preferred embodiment of the LSTM-Transformer long-term wind power forecasting combined model method described in this invention, it further includes:

[0028] Based on the wind power signal components obtained from variational mode decomposition, the normalized matrix of the original data is obtained, and the correlation coefficient matrix R = (r) of the normalized matrix is ​​calculated. ij ) m×n ;

[0029] Calculate the eigenvalue matrix and the eigenvectors corresponding to the eigenvalues, and calculate the contribution rate T. i and cumulative contribution rate η i According to the cumulative contribution rate η i Determine the number of principal components to be selected;

[0030] The dimensionality-reduced data is obtained by selecting the eigenvalues ​​of the principal components and their corresponding eigenvectors.

[0031] As a preferred embodiment of the LSTM-Transformer long-term wind power prediction combined model method described in this invention, the acquisition of the final predicted wind power includes:

[0032] Define basic parameters, including reward R, action value function Q(s,a), state value function V(s), and advantage function A; construct a SAC network model, and maximize the expected cumulative reward by optimizing the policy network and value function network, while maximizing the entropy of the action distribution; construct an Actor network model and a Critic network model, and formulate a Markov decision process to obtain a combined model prediction method based on deep reinforcement learning flexible action evaluation, and adaptively adjust the prediction weights of the sub-models.

[0033] Data from the actual physical systems associated with the wind farm are collected, and the theoretical wind turbine power generation is calculated based on the data predicted by the SwinLSTM network model.

[0034] By adaptively adjusting the weights of the theoretical wind turbine power generation and the predicted wind turbine power generation using the SAC algorithm, the future wind power generation can be predicted.

[0035] Secondly, this invention provides an LSTM-Transformer combined mid-to-long-term wind power forecasting model system, comprising:

[0036] The dataset processing module is used to acquire and preprocess meteorological data from wind farms, use adversarial generative networks to capture the inherent features of the original dataset, and generate synthetic historical weather data samples to expand the existing dataset.

[0037] The signal decomposition module is used to perform empirical mode decomposition on the augmented dataset to obtain modal components at different time scales, and to perform variational mode decomposition on the augmented dataset to obtain wind turbine power trend sequences and periodic sequences.

[0038] The training and prediction module is used to construct the SwinLSTM network model and train the SwinLSTM network model using the data obtained from signal decomposition to predict the comprehensive environmental data of the wind turbine location and the wind power at the wind turbine location, respectively.

[0039] The results acquisition module is used to construct an adaptive weight adjustment method based on deep reinforcement learning for flexible action evaluation, and to adaptively adjust the theoretical wind power and predicted wind power to obtain the final predicted wind power.

[0040] Thirdly, the present invention provides an electronic device, comprising:

[0041] Memory and processor;

[0042] The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions, which, when executed by the processor, implement the steps of the LSTM-Transformer medium- and long-term wind power prediction combined model method.

[0043] Fourthly, the present invention provides a computer-readable storage medium storing computer-executable instructions that, when executed by a processor, implement the steps of the LSTM-Transformer medium- and long-term wind power prediction combined model method.

[0044] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention provides a combined LSTM-Transformer model and system for medium- and long-term wind power prediction. It utilizes a deep convolutional generative adversarial network (DCGAN) to augment the original wind power dataset, capturing and preserving its intrinsic features. Considering the complexity of factors influencing wind power, it employs empirical mode decomposition (EMD) to process the augmented dataset, extracting intrinsic mode components at different time scales. Based on this, a SwinLSTM network is used to predict the comprehensive environmental data and wind power at wind turbine locations, effectively reducing prediction errors. Finally, a soft actor-critic (SAC) method based on deep reinforcement learning is introduced to achieve adaptive weight adjustment, dynamically optimizing the prediction weights of each sub-model to ensure the accuracy and flexibility of the prediction strategy. This research not only improves the technical level of wind power prediction but also provides strong support for the stable operation of power systems and energy management, contributing to the effective utilization and sustainable development of renewable energy. Attached Figure Description

[0045] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0046] Figure 1 This is a schematic diagram of the overall process logic of the LSTM-Transformer long-term wind power prediction combined model method according to an embodiment of the present invention;

[0047] Figure 2 This is a flowchart illustrating the LSTM-Transformer combined model method for medium- and long-term wind power prediction according to an embodiment of the present invention.

[0048] Figure 3 This is a background block diagram of the LSTM-Transformer combined model method for medium- and long-term wind power prediction according to an embodiment of the present invention;

[0049] Figure 4 This is a structural block diagram of the LSTM-Transformer long-term wind power prediction combined model method according to an embodiment of the present invention;

[0050] Figure 5 This is a schematic diagram of the DCGAN generator G in the LSTM-Transformer long-term wind power prediction combined model method according to an embodiment of the present invention.

[0051] Figure 6 This is a schematic diagram of the SwinTransformer block of the LSTM-Transformer long-term wind power prediction combined model method according to an embodiment of the present invention;

[0052] Figure 7 This is a schematic diagram of the SwinLSTM module of the LSTM-Transformer long-term wind power prediction combined model method according to an embodiment of the present invention.

[0053] Figure 8 This is a schematic diagram of the SwinLSTM-B module of the LSTM-Transformer long-term wind power prediction combined model method according to an embodiment of the present invention.

[0054] Figure 9 This is a comparison chart of the predicted and actual values ​​of the LSTM-Transformer long-term wind power prediction combined model method according to an embodiment of the present invention.

[0055] Figure 10This is a comparison chart of the predicted and actual values ​​using the original dataset in the LSTM-Transformer long-term wind power prediction combined model method according to an embodiment of the present invention.

[0056] Figure 11 This is a comparison chart of the undecomposed predicted and actual values ​​of the LSTM-Transformer long-term wind power prediction combined model method according to an embodiment of the present invention.

[0057] Figure 12 This is a bar chart comparing the errors of various methods in the LSTM-Transformer long-term wind power prediction combined model method described in one embodiment of the present invention. Detailed Implementation

[0058] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0059] Example 1

[0060] Reference Figures 1-8 As an embodiment of the present invention, a method for a combined mid-to-long-term wind power forecasting model using LSTM-Transformer is provided, such as... Figure 1 The specific steps shown are as follows:

[0061] S100: Acquire meteorological data from the wind farm and summarize and clean the data;

[0062] S200: Uses adversarial generative networks to capture the intrinsic features of the original dataset and generates synthetic historical weather data samples to augment the existing dataset;

[0063] S300: Perform Empirical Mode Decomposition on the augmented dataset to obtain intrinsic mode components at different time scales, and perform Variational Mode Decomposition on the augmented dataset to obtain wind turbine power trend sequences and periodic sequences;

[0064] S400: Construct a SwinLSTM network model and train the SwinLSTM network model using the data obtained from signal decomposition to predict the comprehensive environmental data of the wind turbine location and the wind power at the wind turbine location, respectively.

[0065] S500: It uses PCA to reduce the dimensionality of input variables, which improves prediction accuracy while maintaining the computational speed of the SwinLSTM network model, and overcomes the problem of overfitting.

[0066] S600: Construct an adaptive weight adjustment method for flexible action evaluation based on deep reinforcement learning, and train base modules to convergence based on the augmented dataset;

[0067] S700: Adaptively adjusts the theoretical wind power and predicted wind power to obtain the final predicted wind power, thus enabling the prediction of future wind power.

[0068] It should be noted that this invention provides a combined LSTM-Transformer model and system for medium- and long-term wind power forecasting. It utilizes a deep convolutional generative adversarial network (DCGAN) to augment the original wind power dataset, capturing and preserving its intrinsic features. Considering the complexity of factors influencing wind power, empirical mode decomposition (EMD) is used to process the augmented dataset, extracting intrinsic mode components at different time scales. Based on this, a SwinLSTM network is employed to predict both the comprehensive environmental data of wind turbine locations and wind power, effectively reducing prediction errors. Finally, a soft actor-critic (SAC) method based on deep reinforcement learning is introduced to achieve adaptive weight adjustment, dynamically optimizing the prediction weights of each sub-model to ensure the accuracy and flexibility of the prediction strategy. This research not only improves the technical level of wind power forecasting but also provides strong support for the stable operation of power systems and energy management, contributing to the effective utilization and sustainable development of renewable energy.

[0069] The flowchart of this invention is as follows Figure 2 As shown, a combined LSTM-Transformer model for medium- and long-term wind power forecasting based on signal decomposition is presented. The background diagram of this invention is as follows. Figure 3 As shown, the overall network model structure diagram constructed in this invention is as follows: Figure 4 As shown.

[0070] In this embodiment of the application, step S100 includes the following sub-steps A1-A3;

[0071] In A1: Set the outliers in the dataset to 0, select continuous sequence data of different lengths to construct training and test sets, and set the outliers as the predicted points of the test set;

[0072] In A2: The dataset is normalized by scaling the data to the range [-1, 1] or [0, 1] so that the neural network can process it better, as shown in the following formula:

[0073]

[0074] In A3: Identify and process noisy data that deviates significantly from the normal range.

[0075] It should be noted that step S100 above, by acquiring and cleaning the meteorological data of the wind farm, ensures that subsequent analysis and forecasting are based on high-quality and accurate data. This not only removes noise and outliers, improving the reliability and usability of the data, but also lays the foundation for generating more realistic synthetic data, thereby directly improving the performance and accuracy of the entire forecasting system.

[0076] In this embodiment of the application, step S200 includes the following sub-steps B1-B2;

[0077] In B1: An embedded network is used to capture the features of wind power time series data and map it to a low-dimensional latent space, preserving the key information of the time series;

[0078] In B2: such as Figure 5 The diagram shows the construction of a generator G and a discriminator D. G(z) represents the generator function that maps the latent space vector z sampled from the standard normal distribution to the data space, and D(x) represents the probability that the output x of the discriminator network comes from the training data rather than the generator.

[0079] In B3: The segmented data is input into DCGAN for training. Real wind power data is labeled as 1, and fake data is generated and labeled as 0 for adversarial training. The performance of the discriminator is measured using binary cross-entropy loss.

[0080] l(x,y)=L={l1,...,l N} T ,l n =-[y n *logx n +(1-y n )*log(1-x n )]

[0081] In B4: the distribution of generated data is compared with that of real data, Kullback-Leibler divergence is used for evaluation, the model is fine-tuned with real wind power generation data, and DCGAN can be used to generate the required historical weather time series data.

[0082] It should be noted that step S200 above effectively expands the existing dataset by using a Generative Adversarial Network (GAN) to capture the inherent features of the original dataset and generate synthetic historical weather data samples. This method not only increases the amount of data but also introduces more diverse meteorological conditions, allowing the model training to move beyond the limitations of the original data and learn from a wider range of data distributions and patterns. The resulting rich and diverse training data helps improve the model's generalization ability, enabling it to make more accurate and stable predictions under various meteorological conditions in actual operation, thereby enhancing the reliability and adaptability of the entire wind power prediction system.

[0083] In this embodiment of the application, step S300 includes:

[0084] Specifically, performing Empirical Mode Decomposition (EMD) on augmented datasets to obtain intrinsic mode components at different time scales includes:

[0085] For an original wind power sequence x(i), find all the maximum points in the sequence and determine all the maximum points as the upper envelope, find all the minimum points in the sequence and determine all the minimum points as the lower envelope, m(i) represents the average value of the upper envelope and the lower envelope, and the first component is h1(i) = x(i) - m(i);

[0086] In the next round of screening, h1(i) is regarded as the original data, and m1(i) is the average of the upper and lower envelopes of h1(i), thus determining the second component h2(i);

[0087] The screening process is repeated n times until h is reached. n (i) is an intrinsic mode function or residual component r n (i) is a monotonic function, which terminates the decomposition process;

[0088] Specify q1(i) = h1(i), q2(i) = h2(i), ..., q k (i)=h k x(i) is ultimately decomposed into n eigenmode components q. t (i) and a remaining component r n (i): The formula is expressed as:

[0089]

[0090] Specifically, variational mode decomposition is performed on the augmented dataset to obtain wind turbine power trend sequences and periodic sequences, including:

[0091] Let wind power data f be decomposed into i components. The decomposition sequence should be a finite bandwidth modal component with a medium frequency, while minimizing the sum of the estimated bandwidths of all modes. The constraint is that the sum of all modes is equal to the original data. The corresponding constraint variational expression is as follows, where K is the number of modes to be decomposed, {u k},、{ω k}, corresponding to the k-th modal component and the mid-frequency after decomposition, respectively, δ(i), is the Lagrange function, and * is the convolution operator. Then, the equation is solved by introducing the Lagrange multiplication operator λ;

[0092]

[0093] Transforming the constrained variational problem into an unconstrained variational problem, we obtain the augmented Lagrange expression as follows, where ζ is a quadratic penalty factor, which is used to reduce the interference of Gaussian noise;

[0094]

[0095] By continuously updating each component and its frequency using the alternating direction multiplier iterative algorithm, the saddle point of the unconstrained model, i.e., the optimal solution to the original problem, is obtained. All components can be obtained from the rate space by the following formula, where ω represents the rate. They correspond to The Lie transform of s(i), τ(i);

[0096]

[0097] ,yes After the remaining quantity is filtered by the dimensionality filter, the algorithm re-estimates the repetition rate based on the power spectrum repetition rate of each component and initializes it. and n;

[0098] The execution cycle is n = n + 1, and updates are performed when the frequency ω > 0. And update ω according to the following formula k ,

[0099]

[0100] Repeat the above steps until the iteration stopping condition is met.

[0101] It should be noted that step S300 above can separate long-term trends, periodic changes, and other nonlinear characteristics from the data, thereby providing a better understanding of the dynamic characteristics of wind power. By decomposing the raw signal into components with clear physical meaning, it not only helps to reveal the inherent patterns hidden behind the data but also provides more refined and targeted inputs for subsequent prediction models, enabling the models to more accurately capture the changing patterns of wind power, thus improving prediction accuracy and reliability.

[0102] In this embodiment of the application, step S400 includes the following sub-steps D1-D4;

[0103] In D1: such as Figure 6 The STB module is constructed as shown;

[0104] Specifically, this module employs window-based multi-head self-attention (W-MSA) technology, which effectively reduces the number of elements involved in a single computation by dividing the input sequence into smaller windows and applying the self-attention mechanism independently within these windows, thereby alleviating the computational burden.

[0105] Specifically, building upon W-MSA, SW-MSA further enables the model to capture broader contextual information while maintaining computational efficiency by shifting the positions of these windows at different levels. This mechanism not only enhances the model's ability to express global spatial dependencies but also allows for a more nuanced understanding of spatial relationships in images through the shifting of local windows, making it possible to achieve more accurate dense predictions.

[0106] Specifically, the kernel frequency equations of the Swin Transformer block are as follows, where LN represents Layer Normalization (LayerNorm), and MLP is a multilayer perceptron used for further data processing after the self-attention step. These equations describe how the Swin Transformer block processes the input sequence and captures global spatial dependencies through the self-attention mechanism;

[0107]

[0108]

[0109] In D2: such as Figure 7 The image shows a SwinLSTM Cell constructed based on the STB module.

[0110] Specifically, ConvLSTM overcomes the shortcomings of FC-LSTM in processing spatiotemporal data by introducing convolution operators into the input-to-state and state-to-state transitions. The key equation of ConvLSTM is as follows, where '*' represents convolution. Represents the Hadamard product. I t ,、F,、O t These are the input gate, the forget gate, and the output gate;

[0111] I t =σ(W xi *X t +W hi *H t-1 +B i )

[0112] F t =σ(W xf *X t +W hf *H t-1 +B f )

[0113] O t =σ(W xo *X t +W ho *H t-1+B o )

[0114]

[0115] Specifically, unlike convolution operators which extract local relevance, self-attention mechanisms capture global spatial dependencies by calculating similarity scores across all locations. Therefore, we remove all weights W and biases B from the above equation, resulting in the following equation, where I t =F t =O t Therefore, we will have three doors: I t F, O t The components are merged into a single gate, called the filter gate R. t ;

[0116] I t =F t =O t =σ(X) t +H t-1 )

[0117]

[0118] Specifically, in SwinLSTM, the cell state C t and hidden state H t The information level is updated to capture both long-term and short-term temporal dependencies. Simultaneously, the Swing Transformer block vertically learns the global spatial dependencies. The key equations of the SwingLSTM are as follows, where STB represents the Swing Transformer block and LP represents the linear projection;

[0119] R t =σ(STB(LP(X) t H t-1 )))

[0120] C t =R t ⊙(tanh(STB(LP(X t H t-1 )))+C t-1 )

[0121] H t =R t ⊙tanh(C t )

[0122] In D3: such as Figure 8 The image shows a SwinLSTM-B network model constructed using SwinLSTM Cell as the core frequency.

[0123] Specifically, the wind power data at time t is divided into non-overlapping data blocks of length δ, where δ is equal to 2 or 4. Therefore, the feature dimension of each patch is C·δ, where C represents the number of channels.

[0124] Specifically, the data blocks are flattened and fed into the patch embedding layer, which linearly maps the original features of the patches to arbitrary dimensions;

[0125] Specifically, the SwinLSTM layer receives the converted image patch and the hidden state H t-1 and unit state C t-1 To generate hidden state H t and unit state C t H t The image is copied into two copies, one for rebuilding the layer, and the other for use with C. t Together they are used in the SwinLSTM layer for the next time step;

[0126] In D4: The SwinLSTM network model is used to process wind power data to obtain the predicted wind turbine power output.

[0127] It should be noted that step S400 combines the excellent global spatial dependency capture capability of the Swing Transformer with the powerful performance of LSTM in processing time series data, thus enabling more accurate modeling of complex spatiotemporal relationships. By predicting environmental data and wind power separately, the model not only reduces prediction errors but also provides a more detailed understanding of the various factors affecting wind power output. Furthermore, using data preprocessed with Empirical Mode Decomposition and Variational Mode Decomposition as input further enhances the model's ability to learn features at different time scales, ensuring the accuracy and stability of the prediction results.

[0128] In this embodiment, step S500 uses PCA to reduce the dimensionality of the input variables, improving prediction accuracy while maintaining the computational speed of the SwinLSTM network model, and overcoming overfitting issues, including:

[0129] Based on the wind power signal components obtained from variational mode decomposition, the normalized matrix of the original data is obtained, and the correlation coefficient matrix R = (r) of the normalized matrix is ​​calculated. ij ) m×n ;

[0130] Calculate the eigenvalue matrix and the eigenvectors corresponding to the eigenvalues, and calculate the contribution rate T. i and cumulative contribution rate η i According to the cumulative contribution rate η i Determine the number of principal components to be selected, where the contribution rate T i and cumulative contribution rate η i The calculation formula is:

[0131]

[0132] The dimensionality-reduced data is obtained by selecting the eigenvalues ​​of the principal components and their corresponding eigenvectors.

[0133] It should be noted that step S500 effectively reduces the dimensionality of the data, eliminates redundant information and noise, and allows the model to focus on the most representative and explanatory features, thereby enhancing the model's generalization ability. The dimensionality-reduced data not only speeds up training and prediction but also ensures the model's efficiency and responsiveness when handling large-scale datasets. Simultaneously, it reduces complexity, avoids overfitting due to excessive parameters, and maintains the model's stable performance and accuracy on new data.

[0134] In this embodiment of the application, step S600 includes the following sub-steps F1-F5;

[0135] In F1: Define basic parameters;

[0136] Specifically, the reward R (Reward) is given for each interaction with the environment (each step), and the more rewards you get, the better.

[0137] Specifically, the following formula represents the action value function Q(s,a), whose inputs are the current state and the action to be performed, and whose output is the value that the action can bring. Therefore, a frequency-greedy approach is to select the action that maximizes Q(s,a) for execution. The dimension of Q(s,a) is equal to the dimension of the action space.

[0138]

[0139] Specifically, the following equation represents the state value function V(s), which is the expectation of the Q function. Because the integral of the expectation eliminates action A, the state value function V can be used to intuitively reflect the quality of a state. It is actually the weighted average of Q(s,a) over different values ​​of a.

[0140]

[0141] Specifically, the following formula is the advantage function A, whose value is equal to the action value function minus the state value function, and is in line with the action value Q(s,a) minus its baseline;

[0142] A π (s,a)=Q π (s,a)-V π (S)

[0143] In F2: A SAC network model is constructed, optimizing the policy network and value function network to maximize the expected cumulative reward while simultaneously maximizing the entropy of the action distribution. The optimization objective of the SAC algorithm is as follows:

[0144]

[0145] Here, α is a regularization coefficient used to control the importance of entropy. A larger α indicates stronger exploratory behavior, which helps accelerate subsequent policy learning; conversely, a smaller α increases the likelihood of the policy getting trapped in poor local optima. ρ π Let represent the state distribution of policy π, where policy entropy It is expressed as follows:

[0146]

[0147] In F3: Construct the Actor network model;

[0148] Specifically, the Actor network consists of a fully connected DNN network. Its input is the system state s observed by the Agent. t According to state s t Output the Gaussian distribution values ​​and standard deviation and perform sampling. The Actor network loss function obtained from the KL divergence can be expressed as follows, where v represents the network parameters of the Actor network;

[0149]

[0150] Specifically, since sampling actions based on a Gaussian distribution is non-differentiable, a reparameterization approach is used to rewrite the expectation of the action as the expectation of the noise. Therefore, the loss function of the Actor network can be rewritten as follows, where... b represents the expected value obtained from sampling. mean and b std These represent the output value and variance of the Actor network, respectively, and ⊙ represents the Hadamard product:

[0151]

[0152] In F4: Construct the Critic network model;

[0153] Specifically, the Critic network consists of two Q-networks and two target Q-networks. The Critic network receives action-state pairs (s) from the Actor network. t ,a tAs input, the fully connected DNN network outputs an expected value as an evaluation of the action-state pair. The Critic network selects a network with a smaller Q-value each time it uses a Q-network, thus mitigating the problem of excessively high Q-values. This target Q-network setting is used to improve stability during training. The loss function of any Q-network can be expressed as follows, where w represents the network parameters of the Q-network, w... - The network parameters represent the target Q-network, B represents the empirical replay pool, and γ is the discount factor.

[0154]

[0155] Specifically, the parameters of each target Q-network are updated using the Exponential Moving Average (EMA) method, given by the following formula, where 0 < ε << 1 is an update factor. The EMA method allows the target Q-network to follow the Q-network in a smooth manner, which helps reduce update discrepancies and effectively avoids unstable fluctuations during the learning process:

[0156] w - ←εw+(1-ε)w -

[0157] In F5: Develop a Markov Decision Process (MDP);

[0158] Specifically, a state-space model is constructed. Before the agent makes a decision, it observes meteorological data from weather forecasts and actual meteorological data within the wind farm, including forecast wind direction and speed and actual measured wind direction and speed within the wind farm. The state set is defined as follows, where... These represent the wind speed and wind direction in the weather forecast, respectively. These represent the actual wind speed and wind direction within the wind farm, respectively.

[0159]

[0160] Specifically, an action space model is constructed. The agent's decision-making involves assigning weights to the prediction results of the sub-models. The action is defined as follows, where w 1 ,w 2 These represent the weight allocation of the prediction results from the two sub-models:

[0161] A={w 1 ,w 2}

[0162] Specifically, a reward model is constructed. In the proposed method, the reward view is defined to reflect the effect of the weights. The predictive performance of each sub-model in the combined model is measured based on the value squared error. The reward view is defined as follows, where p 1 ,p 2p represents the prediction results of two sub-models. * This indicates the actual predicted power generation:

[0163]

[0164] It should be noted that step S600 above constructs an adaptive weight adjustment method based on deep reinforcement learning-based Soft Actor-Critic (SAC). By training base modules on the augmented dataset until convergence, it can dynamically optimize the prediction weights of each sub-model, ensuring the flexibility and accuracy of the prediction strategy. The advantage of this method is that it allows the model to adaptively adjust according to the constantly changing data environment, thereby better coping with the uncertainties and complexities in wind power prediction. By automatically learning the optimal weight allocation through deep reinforcement learning, not only is prediction accuracy improved, but the robustness and adaptability of the system are also enhanced, enabling the prediction model to maintain high performance under different conditions, effectively improving the reliability and practicality of wind power prediction.

[0165] In this embodiment of the application, step S700 includes the following sub-steps G1-G3;

[0166] In G1: Collect data on the actual physical systems related to the wind farm, including wind turbine operating data, meteorological data (such as wind speed, wind direction, temperature, humidity, etc.), terrain data, etc.

[0167] In G2: The theoretical wind turbine power generation is calculated based on wind speed, air pressure, and temperature data at the turbine location predicted by SwinLSTM. The motor power calculation method is as follows, where the air volume Q is in meters per second (m³). 3 / h, total pressure P is in Pa, power N is in kW, η is the total pressure efficiency of the fan, and K is the motor capacity coefficient;

[0168] N = (Q / 3600) * P / (1000 * η) * K

[0169] In G3: The SAC algorithm adaptively adjusts the weights of theoretical wind turbine power generation and predicted wind turbine power generation to predict future wind power output.

[0170] It should be noted that step S700 not only improves the accuracy of predictions but also enhances the model's adaptability to different environmental conditions, making predictions more reliable and stable. Furthermore, adaptive weight adjustment helps balance the relationship between model complexity and prediction performance, avoiding overfitting or underfitting problems, thereby providing strong support for the stable operation of the power system and energy management, and ensuring the effective utilization and sustainable development of wind power resources.

[0171] The above is an illustrative scheme of an LSTM-Transformer medium- and long-term wind power forecasting combined model method according to this embodiment. It should be noted that the technical solution of this LSTM-Transformer medium- and long-term wind power forecasting combined model system belongs to the same concept as the technical solution of the LSTM-Transformer medium- and long-term wind power forecasting combined model method described above. Details not described in detail in the technical solution of the LSTM-Transformer medium- and long-term wind power forecasting combined model system in this embodiment can be found in the description of the technical solution of the LSTM-Transformer medium- and long-term wind power forecasting combined model method described above.

[0172] This embodiment also provides an LSTM-Transformer medium- and long-term wind power forecasting combined model system, including:

[0173] The dataset processing module is used to acquire and preprocess meteorological data from wind farms, use adversarial generative networks to capture the inherent features of the original dataset, and generate synthetic historical weather data samples to expand the existing dataset.

[0174] The signal decomposition module is used to perform empirical mode decomposition on the augmented dataset to obtain modal components at different time scales, and to perform variational mode decomposition on the augmented dataset to obtain wind turbine power trend sequences and periodic sequences.

[0175] The training and prediction module is used to build the SwinLSTM network model. The SwinLSTM network model is trained using the data obtained from signal decomposition to predict the comprehensive environmental data of the wind turbine location and the wind power at the wind turbine location, respectively.

[0176] The results acquisition module is used to construct an adaptive weight adjustment method based on deep reinforcement learning for flexible action evaluation, and to adaptively adjust the theoretical wind power and predicted wind power to obtain the final predicted wind power.

[0177] The above-mentioned unit modules can be embedded in the processor of the computer device in hardware form or independent of it, or they can be stored in the memory of the computer device in software form, so that the processor can call and execute the corresponding operations of the above modules.

[0178] This embodiment also provides an electronic device, which includes a processor, a memory, a communication interface, a display screen, and an input device connected via a system bus. The processor of this computer device provides computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The communication interface of the computer device is used for wired or wireless communication with external terminals. Wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The computer program, when executed by the processor, implements an LSTM-Transformer medium- and long-term wind power forecasting combined model method. The display screen of the computer device can be a liquid crystal display (LCD) or an e-ink display. The input device of the computer device can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the computer device casing, or an external keyboard, touchpad, or mouse, etc.

[0179] This embodiment also provides a computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the method proposed in the above embodiments.

[0180] The storage medium proposed in this embodiment belongs to the same inventive concept as the method proposed in the above embodiments. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0181] Based on the above description of the implementation methods, those skilled in the art can clearly understand that the present invention can be implemented using software and necessary general-purpose hardware, and of course, it can also be implemented using hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of the various embodiments of the present invention.

[0182] Example 2

[0183] Reference Figures 9-12Based on the previous embodiment, this embodiment provides an application example of the LSTM-Transformer medium- and long-term wind power prediction combined model method and system to verify and explain the technical effects adopted in this method.

[0184] To verify the effectiveness of the proposed algorithm, this embodiment was tested on a dataset from a wind farm in Germany. The prediction time was one month, the prediction event resolution was 15 minutes, and the probabilistic prediction of wind power generation data for the next month was performed. Figure 9 This paper presents a comparison between the predicted and actual values ​​of medium- and long-term wind power forecasts using the proposed LSTM-Transformer model based on signal decomposition. Performance comparisons are then conducted using two evaluation metrics: mean absolute error (MAE) and root mean square error (RMSE). The proposed LSTM-Transformer combined model for medium- and long-term wind power forecasting is the algorithm presented in this paper. For comparison, [the following is provided] Figure 10 The predictions are trained using a dataset that has not been augmented with DCGAN.

[0185] This demonstrates that using DCGAN to capture the intrinsic features of the original dataset and then generating synthetic historical weather data samples to repair and expand the existing dataset significantly increases the amount and diversity of data in the dataset, improves the randomness during sample learning, thereby enhancing prediction accuracy and having a positive impact on model training.

[0186] For comparison, prediction results without signal decomposition and direct training are provided here. Figure 11 It can be seen that by decomposing EMD and VMD to predict environmental data and wind power data respectively, and then configuring weights to calculate the final prediction result, the robustness of the prediction can be enhanced. The prediction is more accurate when the wind power reaches its peak, thereby improving the prediction accuracy and having a positive impact on model training.

[0187] To verify the prediction accuracy of the method in this invention, it was compared with four other methods. The evaluation metrics selected were RMSE and MAE. The experimental results on the original dataset are shown in Table 1.

[0188] Table 1: Performance of Each Model and Method

[0189] method RMA% MSE% LSTM 22.87 20.07 Transformer 21.31 19.69 GRU 20.45 19.25 LSTM-CNN 21.29 18.96 LSTM-Transformer-Conb 19.95 17.11

[0190] As shown in Table 1, the proposed LSTM-Transformer combined model for medium- and long-term wind power prediction, based on signal decomposition, achieves the best performance, outperforming single-network optimization methods such as Transformer, GRU, and LSTM. Similarly, compared to the LSTM / CNN combined model, the proposed combined model, by integrating two prediction methods derived from signal decomposition and reinforcement learning training, can make near-optimal weight decisions, resulting in better prediction accuracy. The error histograms for the five schemes are shown below. Figure 12 As shown.

[0191] Considering the wind power scenario, this invention combines deep learning with signal decomposition technology, employing the DCGAN module to repair and expand existing datasets. Through signal decomposition, it proposes a combined LSTM-Transformer model for medium- and long-term wind power prediction, addressing both environmental and wind power data. This method aims to extract the inherent features and patterns of wind power and corresponding environmental data through signal decomposition, and then uses deep learning algorithms to train and optimize the model's prediction weights, resulting in a prediction model for both wind power and the wind power environment. Furthermore, an adaptive weight adjustment method based on Deep Reinforcement Learning Flexible Action Evaluation (SAC) is constructed, and this module is trained to convergence using the augmented dataset. Finally, the theoretical and predicted wind power are adaptively weighted to obtain the final predicted wind power, thus achieving accurate wind power prediction. This research not only advances wind power prediction technology but also provides strong support for the stable operation of power systems and energy management.

[0192] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and such modifications or substitutions should be covered within the scope of the claims of the present invention.

[0193] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of this application can be implemented in various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0194] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0195] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0196] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0197] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0198] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A combined LSTM-Transformer model for medium- and long-term wind power forecasting, characterized in that, include: Meteorological data from wind farms are acquired and preprocessed. Generative adversarial networks are used to capture the intrinsic features of the original dataset and generate synthetic historical weather data samples to augment the existing dataset. Empirical mode decomposition is performed on the augmented dataset to obtain intrinsic mode components at different time scales, and variational mode decomposition is performed on the augmented dataset to obtain wind turbine power trend sequences and periodic sequences. A SwinLSTM network model is constructed, and the SwinLSTM network model is trained using the data obtained from signal decomposition to predict the comprehensive environmental data of the wind turbine location and the wind power at the wind turbine location, respectively. An adaptive weight adjustment method based on deep reinforcement learning for flexible action evaluation is constructed, and the theoretical wind power and the predicted wind power are adaptively adjusted to obtain the final predicted wind power. The construction of the SwinLSTM network model includes: The construction of the STB module includes using window-based multi-head self-attention technology to reduce the number of elements involved in a single computation, capturing contextual information by moving the window position at different levels, and capturing global spatial dependencies by inputting the input sequence into the normalization layer and multilayer perception mechanism. The construction of a SwinLSTM Cell based on the STB module involves capturing temporal dependencies by introducing convolution operators into the input-to-state and state-to-state transitions, calculating similarity scores between all positions through a self-attention mechanism to capture global spatial dependencies, and merging the input gate, forget gate, and output gate into a single filter gate. Combining the temporal dependencies and global spatial dependencies, the unit state and hidden state are updated horizontally to capture long-term and short-term temporal dependencies. The SwinLSTM network model, constructed using the SwinLSTM Cell as its core, involves dividing wind power data at time t into non-overlapping data blocks, flattening these data blocks, and inputting them into a patch embedding layer. The SwinLSTM layer receives the transformed image patches and the hidden state H. t-1 and unit state C t-1 To generate hidden state H t and unit state C t H t It was copied into two copies, one for rebuilding the layer, and the other for C. t Together they are used in the SwinLSTM layer for the next time step; The SwinLSTM network model is used to process wind power data to obtain predicted wind turbine power generation.

2. The LSTM-Transformer combined model method for medium- and long-term wind power forecasting as described in claim 1, characterized in that, The generation of synthetic historical weather data samples expands upon existing datasets by including: Embedded networks are used to capture features of wind power time series data and map them into a low-dimensional latent space, preserving key information of the time series. Construct a generator G and a discriminator D, where G(z) represents the generator function that maps the latent space vector z sampled from the standard normal distribution to the data space, and D(x) represents the probability that the output x of the discriminator network comes from the training data rather than the generator. The segmented data is input into DCGAN for training. Real wind power data is labeled as 1, and fake data is generated and labeled as 0 for adversarial training. The performance of the discriminator is measured by binary cross-entropy loss. The generated data distribution is compared with the real data distribution, and the Kullback-Leibler divergence is used for evaluation. The model is then fine-tuned using real wind power generation data.

3. The LSTM-Transformer combined model method for medium- and long-term wind power forecasting as described in claim 2, characterized in that, The process of performing empirical mode decomposition on the augmented dataset to obtain intrinsic mode components at different time scales includes: For an original wind power sequence x(i), find all the maximum points in the sequence and determine all the maximum points as the upper envelope, find all the minimum points in the sequence and determine all the minimum points as the lower envelope, m(i) represents the average value of the upper envelope and the lower envelope, and one component is h1(i) = x(i) - m(i); In the next round of screening, h1(i) is regarded as the original data, and m1(i) is the average value of the upper and lower envelopes of h1(i), thus determining the second component h2(i); The screening process is repeated n times until h is reached. n (i) is an intrinsic mode function or residual component r n (i) is a monotonic function, which terminates the decomposition process; Specify q1(i) = h1(i), q2(i) = h2(i), ..., q k (i)=h k x(i) is ultimately decomposed into n eigenmode components q. t (i) and a remaining component r n (i).

4. The LSTM-Transformer combined model method for medium- and long-term wind power forecasting as described in claim 3, characterized in that, The step of performing variational mode decomposition on the augmented dataset to obtain the wind turbine power trend sequence and periodic sequence includes: The wind power data f is decomposed into i components, ensuring that the decomposed sequence is a modal component with a center frequency and a finite bandwidth, while minimizing the sum of the estimated bandwidths of each mode. The constraint is that the sum of all modes is equal to the original data, and the constrained variational problem is transformed into an unconstrained variational problem. The algorithm continuously updates each component and its center frequency using the alternating direction multiplier iterative algorithm to obtain the saddle point of the unconstrained model, which is the optimal solution of the original problem. The algorithm re-estimates the barycenter frequency based on the power spectrum barycenter of each component and initializes the parameters. The execution cycle is n = n + 1. When the frequency ω > 0, each component is updated. The steps are repeated until the stopping condition is met.

5. The LSTM-Transformer combined model method for medium- and long-term wind power forecasting as described in claim 1, characterized in that, Also includes: Based on the wind power signal components obtained from variational mode decomposition, the normalized matrix of the original data is obtained, and the correlation coefficient matrix R = (r) of the normalized matrix is ​​calculated. ij ) m×n ; Calculate the eigenvalue matrix and the eigenvectors corresponding to the eigenvalues, and calculate the contribution rate T. i and cumulative contribution rate η i According to the cumulative contribution rate η i Determine the number of principal components to be selected; The dimensionality-reduced data is obtained by selecting the eigenvalues ​​of the principal components and their corresponding eigenvectors.

6. The LSTM-Transformer combined model method for medium- and long-term wind power forecasting as described in claim 5, characterized in that, The acquisition of the final predicted wind power includes: Define basic parameters, including reward R, action value function Q(s,a), state value function V(s), and advantage function A; construct a SAC network model, and maximize the expected cumulative reward by optimizing the policy network and value function network, while maximizing the entropy of the action distribution; construct an Actor network model and a Critic network model, and formulate a Markov decision process to obtain a combined model prediction method based on deep reinforcement learning flexible action evaluation, and adaptively adjust the prediction weights of the sub-models. Data from the actual physical systems associated with the wind farm are collected, and the theoretical wind turbine power generation is calculated based on the data predicted by the SwinLSTM network model. By adaptively adjusting the weights of the theoretical wind turbine power generation and the predicted wind turbine power generation using the SAC algorithm, the future wind power generation can be predicted.

7. A system applying the LSTM-Transformer combined medium- and long-term wind power forecasting model method as described in any one of claims 1 to 6, characterized in that, include: The dataset processing module is used to acquire and preprocess meteorological data from wind farms, use adversarial generative networks to capture the inherent features of the original dataset, and generate synthetic historical weather data samples to expand the existing dataset. The signal decomposition module is used to perform empirical mode decomposition on the augmented dataset to obtain modal components at different time scales, and to perform variational mode decomposition on the augmented dataset to obtain wind turbine power trend sequences and periodic sequences. The training and prediction module is used to construct the SwinLSTM network model and train the SwinLSTM network model using the data obtained from signal decomposition to predict the comprehensive environmental data of the wind turbine location and the wind power at the wind turbine location, respectively. The results acquisition module is used to construct an adaptive weight adjustment method based on deep reinforcement learning for flexible action evaluation, and to adaptively adjust the theoretical wind power and predicted wind power to obtain the final predicted wind power.

8. An electronic device, comprising: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, they implement the steps of the LSTM-Transformer medium- and long-term wind power prediction combined model method according to any one of claims 1 to 6.

9. A computer-readable storage medium storing computer-executable instructions that, when executed by a processor, implement the steps of the LSTM-Transformer medium-to-long-term wind power forecasting combined model method according to any one of claims 1 to 6.

Images