A dual-branch spatio-temporal wind power prediction method and system based on multi-modal fusion

Abstract

The application belongs to the technical field of wind power prediction, and particularly relates to a double-branch space-time wind power prediction method and system based on multi-modal fusion. The method comprises the following steps: S1, performing maximum minimum normalization on input original wind power sequence data to obtain standardized wind power sequence data; S2, performing multi-mode embedding processing on the standardized wind power sequence to obtain time sequence features integrating space-time information and feature correlation; meanwhile, performing mode conversion processing on the standardized wind power sequence to obtain image features; S3, respectively performing time sequence feature extraction and image feature extraction on the time sequence features and the image features obtained in S2 to obtain enhanced time sequence features and enhanced image features; S4, performing multi-modal feature fusion on the enhanced time sequence features and the enhanced image features obtained in S3 to obtain fusion features comprising time sequence modes and image modes; and S5, generating multi-step wind power prediction by using a prediction architecture based on GRU to provide prediction results for multiple future time periods.

Description

Technical Field

[0001] This invention belongs to the field of wind power prediction technology, specifically a dual-branch spatiotemporal wind power prediction method and system based on multimodal fusion. Background Technology

[0002] As a clean and renewable energy source, wind power has seen its installed capacity grow rapidly worldwide, becoming an important solution for addressing energy demand and environmental challenges. However, the inherent randomness, intermittency, and volatility of wind energy lead to unstable wind power output, posing significant challenges to the safe, stable, and economical operation of the power grid. Therefore, achieving high-precision wind power forecasting is crucial for power system dispatching plans, reserve capacity arrangements, market transactions, and improving wind power absorption capacity.

[0003] Existing wind power forecasting methods can be mainly divided into physical methods, statistical methods, and deep learning methods. Physical methods are based on meteorological principles and wind turbine physical models, requiring high data quality and having limited accuracy in short-term forecasts. Statistical methods (such as ARIMA and SVR) have simple structures but struggle to effectively characterize the complex nonlinear dynamics of wind power sequences. Deep learning methods, represented by recurrent neural networks (RNNs), convolutional neural networks (CNNs), and their variants (such as LSTM and GRUs), have become the mainstream research approach due to their powerful nonlinear fitting and feature learning capabilities. To further improve performance, researchers have developed various hybrid models, such as combining CNNs and RNNs to simultaneously capture spatial and temporal features, or introducing graph neural networks (GCNs) to explicitly model the spatial topological relationships between wind farms.

[0004] Despite significant progress in existing methods, the following key challenges remain when processing real-world wind power data: First, wind power data is inherently a multivariate time series with high dimensionality, strong nonlinearity, and complex intrinsic dependencies. Most existing methods rely primarily on modeling the original one-dimensional time series, making it difficult to fully extract and utilize the deep patterns and multidimensional features inherent in this high-dimensional nonlinear space. Second, wind turbines within a wind farm exhibit not only local spatial correlations determined by geographical proximity but also long-range coupling effects across spatial distances caused by similar meteorological conditions. Existing spatiotemporal modeling methods often struggle to simultaneously and effectively characterize this complex spatial correlation pattern where local and global factors coexist. Third, although signal decomposition techniques (such as variational mode decomposition and seasonal trend decomposition) are widely used for preprocessing to separate different frequency components in the sequence, most methods only treat decomposition as a preliminary step, failing to deeply and adaptively integrate the components with different physical meanings (such as trend terms and periodic terms) obtained from the decomposition with the internal network structure, thus limiting the model's ability to analyze the multi-scale characteristics of the time series. Finally, traditional methods typically only utilize numerical time-series data, neglecting the possibility of converting the data into other modalities to provide complementary information perspectives, thus failing to achieve cross-modal feature enhancement and fusion.

[0005] To address the aforementioned issues, some studies have attempted to introduce strategies such as attention mechanisms, improved Transformer architectures, or contrastive learning. However, there is still room for improvement in terms of the richness of feature representations, the depth of modeling complex spatiotemporal relationships, and the fusion mechanism of multi-scale information. Therefore, developing an end-to-end wind power prediction framework that can comprehensively utilize multimodal information and deeply analyze spatiotemporal dependencies has significant theoretical and practical value for further improving prediction accuracy and enhancing model robustness. Summary of the Invention

[0006] To address the aforementioned problems, this invention provides a dual-branch spatiotemporal wind power prediction method and system based on multimodal fusion.

[0007] This invention adopts the following technical solution: a dual-branch spatiotemporal wind power prediction method based on multimodal fusion, comprising: S1: Perform max-min normalization on the input raw wind power sequence data to obtain standardized wind power sequence data; S2: Perform multi-mode embedding processing on the standardized wind power sequence to obtain temporal features that incorporate spatiotemporal information and feature correlation; at the same time, perform mode transformation processing on the standardized wind power sequence to obtain image features; S3: Based on the temporal features and image features obtained in S2, enhance temporal features and enhance image features are obtained by extracting temporal features and image features respectively. S4: Perform multimodal feature fusion on the enhanced temporal features and enhanced image features obtained in S3 to obtain fused features including temporal modality and image modality; S5: Utilizes a GRU-based forecasting architecture to generate multi-step wind power forecasts, providing forecast results for multiple future time periods.

[0008] In some embodiments, step S2, the multi-mode embedding process includes: Enhanced features are obtained by embedding spatial information into wind power sequence data; Temporal pattern embedding features are obtained by embedding temporal information into wind power sequence data. The enhanced features are embedded with the correlation to obtain the correlation-embedded features; The features embedded with correlation are fused with multi-resolution temporal features to obtain the final representation of the multi-mode embedding module.

[0009] In some embodiments, step S2 includes the following modal transformation process: The normalized sequence is obtained by performing max-min normalization on the one-dimensional wind power history sequence; The coordinate system is transformed to polar coordinates, and the values ​​of the normalized sequence are mapped to angle cosines, and the time points are mapped to radii. Two symmetric matrices for the Gram angle field are constructed using the angle information obtained from normalized time series data through inverse cosine transformation: the Gram angle sum field and the Gram angle difference field. Markov transition matrices are constructed to capture the dynamic behavior and state transition characteristics of time series using Markov transition fields. By fusing the Gram angle sum field, Gram angle difference field, and Markov transition matrix as three channels, a comprehensive high-dimensional image feature is obtained.

[0010] In some embodiments, step S3, the temporal feature extraction includes: Use moving average decomposition to break down the time series into trend and seasonal components; A recurrent neural network is used to model the global features of the trend term, resulting in the trend term features. A TCN-GRU network is used to capture local and global information about seasonal items; The output of the GRU branch is adaptively weighted by the Softmax attention mechanism, and the weighted result is concatenated with the local features extracted by TCN, thereby achieving deep fusion of local details and temporal dynamics and obtaining an enhanced seasonal feature representation. Enhanced time-series features are obtained by splicing and fusing trend features and seasonal features.

[0011] In some embodiments, the TCN-GRU network includes two parallel branches: The TCN branch utilizes dilated convolution to extract local fluctuation features of the seasonal term; GRU branch, which is used to model the long-range time dependency of sequences.

[0012] In some embodiments, in step S3, the image features are extracted using ResNet for deep feature extraction.

[0013] In some embodiments, step S4 includes: S41: Perform linear projection on the enhanced temporal features and enhanced image features, mapping them to the same feature space: S42: Construct graph structures for the projected enhanced temporal features and enhanced image features respectively; S43: Use the enhanced temporal features processed by the graph neural network as queries and the enhanced image features as keys and values, and achieve cross-modal feature fusion through a cross-attention mechanism.

[0014] In some embodiments, step S5 includes: S51: The fused features from step S4 are used as input to the prediction architecture; the GRU encoder processes the input sequence step by step, updates the hidden state, and obtains a summary representation of historical information. S52: The GRU decoder employs a recursive multi-step prediction strategy. It takes the encoder's final state as the initial state, uses the prediction value of the previous step as the input of the current step, and gradually generates prediction results for multiple future time steps. At each time step, the decoder combines the current hidden state, outputs the prediction power through the fully connected layer, and feeds the prediction value back to the next step as input, thus realizing recursive iteration.

[0015] A dual-branch spatiotemporal wind power prediction system based on multimodal fusion includes a data preprocessing module, a dual-branch multimodal embedding module, a dual-branch feature enhancement and extraction module, a cross-modal feature fusion module, and a multi-step prediction module; The data preprocessing module is used to perform max-min normalization on the input raw wind power sequence data and output standardized wind power sequence data. The dual-branch multimodal embedding module is communicatively connected to the data preprocessing module, and the dual-branch multimodal embedding module includes: A time-series multi-mode embedding unit is used to sequentially embed spatial information, temporal information, and feature correlation into a standardized wind power sequence, and to fuse the features after correlation embedding with multi-resolution temporal features to obtain time-series features that incorporate spatiotemporal information and feature correlation. Sequence-image mode conversion unit; The sequence-image mode conversion unit is used to convert the standardized one-dimensional wind power sequence into a polar coordinate system, construct Gram angle sum field, Gram angle difference field and Markov transition matrix based on the converted data, and fuse the three as three channels to obtain high-dimensional image features; The dual-branch feature enhancement and extraction module is communicatively connected to the dual-branch multimodal embedding module, and the dual-branch feature enhancement and extraction module includes: The time series feature extraction unit is used to decompose the time series features into trend items and seasonal items through moving average decomposition, use a recurrent neural network to model the trend items to obtain trend item features, use a TCN-GRU network combined with a Softmax attention mechanism to model the seasonal items to obtain seasonal item features, and concatenate the trend item features and seasonal item features to obtain enhanced time series features. An image feature extraction unit is used to perform deep feature extraction on high-dimensional image features through a ResNet network to obtain enhanced image features. The cross-modal feature fusion module is communicatively connected to the dual-branch feature enhancement extraction module. It is used to linearly project the enhanced temporal features and enhanced image features to map them to the same feature space. It constructs graph structures for the two types of features after projection, and then uses the enhanced temporal features after graph structure processing as queries and the enhanced image features as keys and values. Cross-modal feature fusion is completed through a cross-attention mechanism, and the output is a fused feature containing temporal modality and image modality information. The multi-step prediction module is communicatively connected to the cross-modal feature fusion module. The multi-step prediction module includes a GRU-based encoder-decoder prediction architecture. The encoder is used to process the fused features step by step to summarize historical information. The decoder is used to generate wind power prediction results for multiple future time periods step by step using the encoder output as the initial state and a recursive multi-step prediction strategy.

[0016] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention proposes a wind power prediction method based on a dual-branch spatiotemporal network with multimodal fusion. This method utilizes Gram angle field (GAF) and Markov transition field (MTF) to achieve mode conversion, enabling implicit representation and learning of temporal features in two-dimensional space, effectively improving the expressive power of temporal features.

[0017] 2. This invention designs a multi-modal embedding mechanism that can jointly model spatiotemporal correlations and dependencies between variables. In particular, the spatial pattern embedding simultaneously considers the geographical proximity between wind turbines and wind speed similarity patterns, enabling the model to more comprehensively capture the spatial structure information within the wind farm.

[0018] 3. This invention proposes an LGST network that decouples wind power sequences into seasonal and trend terms. The seasonal term is processed using a TCN-GRU architecture, which can simultaneously capture local features and long-range dependencies, thereby more accurately characterizing the complex temporal dynamics of wind power.

[0019] 4. This invention constructs a graph-based multimodal fusion module to integrate temporal and visual features. This module models cross-modal interactions through a graph attention mechanism, effectively fusing time-series features with spatial structure information contained in images, improving the collaborative expressive power of multimodal features, and thus enhancing the accuracy and robustness of wind power prediction. Attached Figure Description

[0020] Figure 1 This is a technical roadmap of the present invention; Figure 2 This is a schematic diagram illustrating the generation of GASF, GADF, and MTF in this invention; Figure 3 This is a schematic diagram of the temporal feature extraction structure; Figure 4 This is a distribution map of attention weights for multimodal features. Figure 5 The 1-hour wind force prediction results for wind turbine #1 using 8 different methods are plotted on the Brazilian dataset. Figure 6 A graph showing the 12-hour wind forecast results for wind turbine #32 using eight different methods on a Brazilian dataset. Figure 7 A graph showing the 1-hour wind speed prediction results for wind turbine #1 using 8 different methods on a US dataset; Figure 8 A graph showing 12-hour wind forecast results for wind turbine #200 using eight methods on a US dataset; Figure 9 The 1-hour wind power prediction results for wind turbine #1 using 8 methods are plotted on the Chinese dataset. Figure 10 The image shows the 12-hour wind power prediction results for wind turbine #70 using eight different methods on a Chinese dataset. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are some embodiments of the present invention, but 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.

[0022] A dual-branch spatiotemporal wind power prediction method based on multimodal fusion includes: S1: Input raw wind power sequence data Perform min-max normalization to obtain standardized wind power sequence data. ; S2: Standardized wind power sequence Multi-modal embedding processing is performed to obtain temporal features that incorporate spatiotemporal information and feature correlations. Simultaneously, the standardized wind power sequence is subjected to mode transformation processing to obtain image features. ; S3: Based on the temporal features and image features obtained in S2, enhance the temporal features by performing temporal feature extraction and image feature extraction respectively. and enhance image features ; S4: The enhanced temporal features obtained in S3 and enhance image features Multimodal feature fusion is performed to obtain fused features that include temporal modalities and image modalities. ; S5: Utilizes a GRU-based forecasting architecture to generate multi-step wind power forecasts, providing forecast results for multiple future time periods.

[0023] In a specific embodiment, in step S1, the max-min normalization is performed on the input raw wind power sequence data using the following formula. Normalize: .

[0024] In a specific embodiment, the multi-mode embedding process includes: 1) Wind power sequence data Features obtained by embedding spatial information ; Specifically: First, for the target wind turbine i, calculate its Euclidean distance to all other wind turbines j, and select the k nearest wind turbines as geographical neighbors. The index set is as follows: Use the wind speeds of these neighboring turbines as additional characteristics of the target turbine. ,in This represents the wind speed of the kth adjacent wind turbine of the i-th wind turbine.

[0025] Then, by analyzing wind speed differences, neighboring turbines that have a significant impact on the target turbine are identified. The adjacent time step scores for each wind turbine's wind speed sequence V are calculated: Where T represents the timestamp, the cosine similarity matrix between wind turbines is then calculated based on the difference sequence: in This represents the index of the wind turbine. The top s most similar wind turbines are selected as wind speed neighbors, and the index set is: Use the wind from these neighbors as an additional feature. ,in Indicates relative to the first The first typhoon machine Each neighborhood wind speed. Finally, the enhanced input features of each wind turbine are aggregated. The nearest geographical neighborhood and Constructed using neighborhood wind speeds: in It is a time step Spatial embedding. The wind power sequence after spatial embedding is: .

[0026] 2) Temporal pattern embedding features are obtained by embedding time information into wind power sequence data. ; First, define the set of time resolutions as follows: ,in It indicates a specific time resolution (such as seconds, minutes, hours).

[0027] Then, the original time series is sampled at different resolutions to construct a multi-resolution series: in In terms of time resolution Downsampling time series.

[0028] Next, through the embedding operation Map the sequence at each resolution to the latent space: in It's the resolution. Embedded representation.

[0029] Finally, by weighted fusion of features at different resolutions, the multi-resolution temporal pattern can be modeled as follows: in and These are trainable weights and biases; the apostrophe indicates matrix transpose. express The k-th slice matrix.

[0030] 3) Enhance features Embedding correlation yields correlation embedding features ; First, calculate the autocorrelation matrix of the wind power series: Among them, F and F -1 The symbols represent Fast Fourier Transform (FFT) and Inverse Fast Fourier Transform (IFFT), respectively, and the asterisk indicates the complex conjugate operation.

[0031] In addition, the Softmax function is used to normalize the autocorrelation matrix to highlight the correlation between variables and to weight the inputs: .

[0032] 4) Features after embedding relevance With multi-resolution temporal features The fusion process yields the final representation of the multi-mode embedding module. , .

[0033] In a specific embodiment, step S2, the mode conversion process includes the following steps: Input only one-dimensional historical wind power sequence Normalize it: The obtained normalized sequence .

[0034] The normalized values ​​are mapped to angle cosines, and the time points are mapped to radii: in For a point in time, N This is the normalization factor.

[0035] Two symmetric matrices for constructing the Gram Angular Field (GAF) using angular information: GAF captures the changing trends and abrupt changes in time series, while Markov transition fields (MTF) capture the dynamic behavior and state transition characteristics of time series. MTF divides the time series into Q quantile intervals, thereby constructing the Markov transition matrix W: in Indicates from the interval The frequency at which the frequency transitions to interval j.

[0036] Finally, the GASF, GADF, and MTF matrices are fused as three channels (similar to an RGB image) to obtain a comprehensive high-dimensional image feature. .

[0037] In a specific embodiment, step S3, the extraction of temporal features includes: S31: The input is the time feature after multi-modal embedding. The time series is decomposed into trend and seasonal components using moving average decomposition. Among them, trend items Long-term global patterns are captured through average pooling; seasonal items. The trend term is subtracted from the original sequence, preserving local periodic characteristics.

[0038] S32: Use a recurrent neural network (RNN) to model the global features of the trend term and obtain the trend term features; S33: A TCN-GRU network is used to capture local and global information about the seasonal components. This network contains two parallel branches: the TCN branch uses dilated convolution to extract local fluctuation features of the seasonal components; the GRU branch is used to model the long-range temporal dependencies of the sequence. Subsequently, the output of the GRU branch is adaptively weighted using a Softmax attention mechanism, and the weighted result is concatenated with the local features extracted by the TCN, thereby achieving a deep fusion of local details and temporal dynamics to obtain an enhanced representation of seasonal features. S34: Combine and fuse trend features and seasonal features to obtain enhanced time-series features. : , .

[0039] Image features are enhanced by deep feature extraction using ResNet. .

[0040] In a specific embodiment, step S4 includes: S41: To eliminate feature differences between different modalities, linear projection is performed on the enhanced temporal features and enhanced image features, mapping them to the same feature space: .

[0041] S42: Construct graph structures for the projected enhanced temporal features and enhanced image features respectively, use GAT to dynamically model the importance between nodes, and update node features: Among them, attention coefficient The calculation process is as follows: .

[0042] in This represents the importance of different nodes. A It is an adjacency matrix. The Bahdanau attention mechanism was implemented. , and All of these are learnable parameters. This indicates a splicing operation. It is a node in the graph and nodes An infinite set of all its neighbors.

[0043] S43: Time series features processed by graph neural network As a query, image features Cross-modal feature fusion is achieved through a cross-attention mechanism, using both keys and values. .

[0044] Figure 4 This diagram illustrates the distribution of attention weights between queries and keys in multimodal features, where color intensity corresponds to the magnitude of the fusion weights.

[0045] In a specific embodiment, step S5 includes: S51: The fused features from step S4 are used as input to the prediction architecture; the GRU encoder processes the input sequence step by step, updates the hidden state, and obtains a summary representation of historical information. S52: The GRU decoder employs a recursive multi-step prediction strategy. It takes the encoder's final state as the initial state, uses the prediction value of the previous step as the input of the current step, and gradually generates prediction results for multiple future time steps. At each time step, the decoder combines the current hidden state, outputs the prediction power through the fully connected layer, and feeds the prediction value back to the next step as input, thus realizing recursive iteration.

[0046] A dual-branch spatiotemporal wind power prediction system based on multimodal fusion includes a data preprocessing module, a dual-branch multimodal embedding module, a dual-branch feature enhancement and extraction module, a cross-modal feature fusion module, and a multi-step prediction module. The data preprocessing module performs max-min normalization on the input raw wind power sequence data, outputting standardized wind power sequence data. The dual-branch multimodal embedding module is communicatively connected to the data preprocessing module and includes a temporal multimodal embedding unit, which sequentially embeds spatial information, temporal information, and features into the standardized wind power sequence. The system employs a multi-modal embedding module, which embeds features based on relevance and fuses them with multi-resolution temporal features to obtain temporal features incorporating spatiotemporal information and feature correlation. A sequence-image modality conversion unit is used to convert the standardized one-dimensional wind power sequence into a polar coordinate system. Based on the converted data, it constructs Gram angle sum field, Gram angle difference field, and Markov transition matrix, fusing these three as three channels to obtain high-dimensional image features. A dual-branch feature enhancement and extraction module is communicatively connected to the dual-branch multimodal embedding module. The dual-branch feature enhancement and extraction module includes a temporal feature extraction unit, which is used to extract features by... Moving average decomposition decomposes time-series features into trend and seasonal components. A recurrent neural network is used to model the trend component to obtain trend features, and a TCN-GRU network combined with a Softmax attention mechanism is used to model the seasonal component to obtain seasonal features. The trend and seasonal features are then concatenated to obtain enhanced time-series features. An image feature extraction unit is used to perform deep feature extraction on high-dimensional image features using a ResNet network to obtain enhanced image features. A cross-modal feature fusion module is communicatively connected to the dual-branch feature enhancement extraction module and is used to perform linear projection on the enhanced time-series features and enhanced image features to map them to the same feature. The feature space is constructed by building graph structures for the two types of projected features. Then, the enhanced temporal features processed by the graph structure are used as queries and the enhanced image features are used as keys and values. Cross-modal feature fusion is completed through a cross-attention mechanism, and the output is a fused feature containing temporal modality and image modality information. The multi-step prediction module is communicatively connected to the cross-modal feature fusion module. The multi-step prediction module includes an encoder-decoder prediction architecture based on GRU. The encoder is used to process the fused features step by step to summarize historical information. The decoder is used to use the output of the encoder as the initial state and adopt a recursive multi-step prediction strategy to gradually generate wind power prediction results for multiple future time periods.

[0047] To verify the effectiveness of this invention, it was compared with several other algorithms on datasets collected from the Beberibe wind farm on the northeastern coast of Brazil, a wind farm dataset in the United States, and a dataset in southern China, to verify the superiority of this invention in different application scenarios.

[0048] The Brazilian dataset contains SCADA monitoring data for 32 wind turbines collected between August 2013 and July 2014, with a sampling frequency of 10 minutes. Key parameters recorded include wind speed, power generation, and turbine geographic coordinates. To meet forecasting requirements, the raw data was resampled to obtain a 1-hour temporal resolution. The US data contains hourly wind speed, power generation, and geographic coordinate measurements recorded between September 2010 and August 2011. The data provider has performed minimum-maximum normalization on the power values. The Chinese dataset records wind speed, power generation, and geographic location information for 70 wind turbines between January 2023 and January 2024, with a sampling frequency of 15 minutes. To ensure consistency with the baseline dataset, the raw data was resampled to a 1-hour resolution.

[0049] This invention uses the root mean square error (MAE) and root mean square error (RMSE), which are commonly used in regression tasks, to evaluate wind power generation prediction results.

[0050] The specific experimental procedure includes the following steps: (1) Model training: All experiments in this invention were performed 10 times to obtain the average result. During the training process on the dataset, the Adam optimizer was used to optimize the network, with the learning rate set to 0.0001 and the dropout rate set to 0.15. The training process was configured for 300 epochs, and an early stopping strategy was used to prevent overfitting.

[0051] (2) Model testing: The dataset is divided according to empirical proportions: two months for training, one month for validation, and the remainder for testing. Historical observation data for 48 consecutive hours is used as input to predict the power output for the next 12 hours.

[0052] (3) Experimental results: The performance of the two evaluation metrics on different datasets is shown in the table below: Table 1. Comparison of prediction performance on the Brazilian dataset MAE of wind power prediction results generated by different methods on the Brazilian dataset RSME of wind power prediction results generated by different methods on the Brazilian dataset Table 2. Comparison of prediction performance on the US dataset. MAE of wind power forecasts generated by different methods on a US dataset RSME wind power forecasts generated by different methods on a US dataset Table 3 Comparison of prediction performance on the Chinese dataset Wind power prediction results generated by different methods on the Chinese dataset (MAE) RSME of wind power prediction results generated by different methods on Chinese datasets The experimental results above show that, on datasets from three different regions, the method of this invention achieves better prediction accuracy compared to some advanced comparative algorithms. Furthermore, the performance advantage becomes increasingly significant as the prediction timeframe lengthens, validating the superior ability of this method in multi-step prediction. Figure 5-10 The qualitative results show that our method has better predictive performance compared with some advanced methods.

[0053] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A dual-branch spatiotemporal wind power prediction method based on multimodal fusion, characterized in that, include: S1: Perform max-min normalization on the input raw wind power sequence data to obtain standardized wind power sequence data; S2: Perform multi-mode embedding processing on the standardized wind power sequence to obtain temporal features that incorporate spatiotemporal information and feature correlation; at the same time, perform mode transformation processing on the standardized wind power sequence to obtain image features; S3: Based on the temporal features and image features obtained in S2, enhance temporal features and enhance image features are obtained by extracting temporal features and image features respectively. S4: Perform multimodal feature fusion on the enhanced temporal features and enhanced image features obtained in S3 to obtain fused features including temporal modality and image modality; S5: Utilizes a GRU-based forecasting architecture to generate multi-step wind power forecasts, providing forecast results for multiple future time periods.

2. The dual-branch spatiotemporal wind power prediction method based on multimodal fusion according to claim 1, characterized in that, In step S2, the multi-mode embedding process includes: Enhanced features are obtained by embedding spatial information into wind power sequence data; Temporal pattern embedding features are obtained by embedding temporal information into wind power sequence data. The enhanced features are embedded with the correlation to obtain the correlation-embedded features; The features embedded with correlation are fused with multi-resolution temporal features to obtain the final representation of the multi-mode embedding module.

3. The dual-branch spatiotemporal wind power prediction method based on multimodal fusion according to claim 1, characterized in that, In step S2, the mode conversion process includes: The normalized sequence is obtained by performing max-min normalization on the one-dimensional wind power history sequence; The coordinate system is transformed to polar coordinates, and the values ​​of the normalized sequence are mapped to angle cosines, and the time points are mapped to radii. Two symmetric matrices for the Gram angle field are constructed using the angle information obtained from normalized time series data through inverse cosine transformation: the Gram angle sum field and the Gram angle difference field. Markov transition matrices are constructed to capture the dynamic behavior and state transition characteristics of time series using Markov transition fields. By fusing the Gram angle sum field, Gram angle difference field, and Markov transition matrix as three channels, a comprehensive high-dimensional image feature is obtained.

4. The dual-branch spatiotemporal wind power prediction method based on multimodal fusion according to claim 1, characterized in that, In step S3, the temporal feature extraction includes: Use moving average decomposition to break down the time series into trend and seasonal components; A recurrent neural network is used to model the global features of the trend term, resulting in the trend term features. A TCN-GRU network is used to capture local and global information about seasonal items; The output of the GRU branch is adaptively weighted by the Softmax attention mechanism, and the weighted result is concatenated with the local features extracted by TCN, thereby achieving deep fusion of local details and temporal dynamics and obtaining an enhanced seasonal feature representation. Enhanced time-series features are obtained by splicing and fusing trend features and seasonal features.

5. The dual-branch spatiotemporal wind power prediction method based on multimodal fusion according to claim 1, characterized in that, The TCN-GRU network comprises two parallel branches: The TCN branch utilizes dilated convolution to extract local fluctuation features of the seasonal term; GRU branch, which is used to model the long-range time dependency of sequences.

6. The dual-branch spatiotemporal wind power prediction method based on multimodal fusion according to claim 1, characterized in that, In step S3, the image features are enhanced by deep feature extraction using ResNet.

7. The dual-branch spatiotemporal wind power prediction method based on multimodal fusion according to claim 1, characterized in that, Step S4 includes: S41: Perform linear projection on the enhanced temporal features and enhanced image features, mapping them to the same feature space: S42: Construct graph structures for the projected enhanced temporal features and enhanced image features respectively; S43: The enhanced temporal features after the graph neural network are used as the query for processing, and the enhanced image features are used as the key and value. Cross-modal feature fusion is achieved through the cross-attention mechanism.

8. The dual-branch spatiotemporal wind power prediction method based on multimodal fusion according to claim 1, characterized in that, Step S5 includes: S51: The fused features from step S4 are used as input to the prediction architecture; the GRU encoder processes the input sequence step by step, updates the hidden state, and obtains a summary representation of historical information. S52: The GRU decoder employs a recursive multi-step prediction strategy. It takes the encoder's final state as the initial state, uses the prediction value of the previous step as the input of the current step, and gradually generates prediction results for multiple future time steps. At each time step, the decoder combines the current hidden state, outputs the prediction power through the fully connected layer, and feeds the prediction value back to the next step as input, thus realizing recursive iteration.

9. A dual-branch spatiotemporal wind power prediction system based on multimodal fusion, characterized in that, It includes a data preprocessing module, a two-branch multimodal embedding module, a two-branch feature enhancement and extraction module, a cross-modal feature fusion module, and a multi-step prediction module; The data preprocessing module is used to perform max-min normalization on the input raw wind power sequence data and output standardized wind power sequence data. The dual-branch multimodal embedding module is communicatively connected to the data preprocessing module, and the dual-branch multimodal embedding module includes: A time-series multi-mode embedding unit is used to sequentially embed spatial information, temporal information, and feature correlation into a standardized wind power sequence, and to fuse the features after correlation embedding with multi-resolution temporal features to obtain time-series features that incorporate spatiotemporal information and feature correlation. Sequence-image mode conversion unit; The sequence-image mode conversion unit is used to convert the standardized one-dimensional wind power sequence into a polar coordinate system, construct Gram angle sum field, Gram angle difference field and Markov transition matrix based on the converted data, and fuse the three as three channels to obtain high-dimensional image features; The dual-branch feature enhancement and extraction module is communicatively connected to the dual-branch multimodal embedding module, and the dual-branch feature enhancement and extraction module includes: The time series feature extraction unit is used to decompose the time series features into trend items and seasonal items through moving average decomposition, use a recurrent neural network to model the trend items to obtain trend item features, use a TCN-GRU network combined with a Softmax attention mechanism to model the seasonal items to obtain seasonal item features, and concatenate the trend item features and seasonal item features to obtain enhanced time series features. An image feature extraction unit is used to perform deep feature extraction on high-dimensional image features through a ResNet network to obtain enhanced image features. The cross-modal feature fusion module is communicatively connected to the dual-branch feature enhancement extraction module. It is used to linearly project the enhanced temporal features and enhanced image features to map them to the same feature space. It constructs graph structures for the two types of features after projection, and then uses the enhanced temporal features after graph structure processing as queries and the enhanced image features as keys and values. Cross-modal feature fusion is completed through a cross-attention mechanism, and the output is a fused feature containing temporal modality and image modality information. The multi-step prediction module is communicatively connected to the cross-modal feature fusion module. The multi-step prediction module includes a GRU-based encoder-decoder prediction architecture. The encoder is used to process the fused features step by step to summarize historical information. The decoder is used to generate wind power prediction results for multiple future time periods step by step using the encoder output as the initial state and a recursive multi-step prediction strategy.

Images