A wind power prediction method and device based on a double-branch cross-modal alignment

Abstract

The application discloses a wind power prediction method and device based on a double-branch cross-modal alignment, and relates to the technical field of wind power prediction. First, a double-branch feature extraction architecture including a wide-area branch and a local-area branch is built, two branches complete feature extraction in parallel, discrete meteorological features are obtained by discretely representing and encoding regional grid meteorological data covering a target wind farm through the wide-area branch, and time sequence features are obtained by weighting and enhancing wind farm multivariate time sequence data and mining time sequence dependence through the local-area branch; then, cross-modal feature alignment is completed relying on a multi-head cross attention mechanism, meteorological enhanced features are generated, and time sequence features and the meteorological enhanced features are fused through a self-adaptive residual gate module to construct multi-modal fusion features; finally, deep time sequence modeling is carried out on the multi-modal fusion features, and a wind power prediction result of the target wind farm is output. The application can improve the wind power prediction accuracy of a wind farm.

Description

Technical Field

[0001] This application relates to the field of wind power prediction technology, specifically to a wind power prediction method and apparatus based on dual-branch cross-modal alignment. Background Technology

[0002] Wind power forecasting is a core technological support for power system dispatching, grid connection control, reserve capacity configuration, and renewable energy consumption, and is of great significance for ensuring the safe and stable operation of the power grid and improving the utilization rate of wind power. Due to the inherent randomness, intermittency, and volatility of wind resources, the factors influencing the output power of wind farms are complex and diverse. It is not only directly affected by local wind speed, wind direction, temperature, air pressure, humidity, and historical power generation variables, but also heavily dependent on the evolution patterns of surrounding large-scale meteorological systems. Regional grid meteorological data covering target wind farms can capture early warning information of meteorological evolution, effectively compensating for the limitations of local observation data and serving as a key data source for improving wind power forecasting performance.

[0003] Current mainstream AI-based wind power forecasting methods largely focus on modeling local time-series data from wind farms, or simply concatenate numerical weather forecast data with local time-series data before inputting it into the model. This approach has significant technical limitations. Regional grid meteorological data is characterized by a large number of spatial grids, high variable dimensionality, and a mismatch between spatiotemporal scales and locally collected data. Directly inputting this data into modeling easily leads to data dimensionality redundancy, high computational complexity, a significant risk of model overfitting, and low feature representation efficiency. Furthermore, the simple concatenation method assumes natural feature alignment across different data sources, failing to adaptively filter effective meteorological context information based on the real-time operating status of the wind farm, and making it difficult to fully extract effective precursor features from the grid meteorological data. In complex scenarios such as sudden wind speed changes, drastic power fluctuations, and long-term forecasts, existing technologies cannot simultaneously consider local temporal variation patterns and wide-area meteorological evolution trends, generally exhibiting insufficient prediction accuracy and poor prediction stability, making it difficult to meet the actual needs of grid dispatching and operation under high-proportion wind power grid integration. Summary of the Invention

[0004] This application addresses the problems existing in the prior art by providing a wind power prediction method and device based on dual-branch cross-modal alignment, which can solve the problems of insufficient prediction accuracy and poor prediction result stability in the prior art, and meet the actual needs of grid dispatch and operation under high proportion of wind power grid connection.

[0005] To achieve the above objectives, the technical solution adopted in this application is as follows: The first aspect of this application provides a wind power prediction method based on bi-branch cross-modal alignment, comprising: constructing a bi-branch feature extraction architecture, which includes a wide-area branch and a local branch, wherein the wide-area branch and the local branch perform feature extraction in parallel; performing discrete representation encoding on the regional grid meteorological data covering the target wind farm through the wide-area branch to output discrete meteorological features; performing weighted enhancement and extracting temporal dependencies on the multivariate time-series data of the target wind farm through the local branch to output time-series features; constructing a multi-head cross-attention calculation module, using the time-series features as query vectors and the discrete meteorological features as key and value vectors, performing cross-modal feature alignment operations to output meteorological enhancement features; configuring an adaptive residual gating module, which performs weighted fusion of time-series features and meteorological enhancement features through learnable gating parameters to generate multi-modal fusion features; performing deep temporal modeling on the multi-modal fusion features to output the wind power prediction result of the target wind farm.

[0006] In some embodiments of this application, the regional grid meteorological data includes wind speed, wind direction, temperature, air pressure and precipitation rate covering the area surrounding the target wind farm, and the multivariate time series data includes wind speed, wind direction, temperature, air pressure, humidity and historical power of the target wind farm; before performing feature extraction, the regional grid meteorological data is processed to unify the time resolution so that the regional grid meteorological data and the multivariate time series data are matched in time scale.

[0007] In some embodiments of this application, discrete representation encoding is performed on the regional grid meteorological data covering the target wind farm through wide-area branching to output discrete meteorological features. This includes: inputting the regional grid meteorological data into a vector quantization meteorological encoder, mapping the continuous high-dimensional regional grid meteorological data into a discrete codebook index and the corresponding meteorological feature vector through the vector quantization meteorological encoder, and using the discrete codebook index and the corresponding meteorological feature vector as discrete meteorological features.

[0008] In some embodiments of this application, weighted enhancement and extraction of temporal dependencies are performed on the multivariate time series data of the target wind farm through local branching, and time series features are output. This includes: extracting features from each variable in the multivariate time series data, calculating the dynamic weights corresponding to each variable, and performing weighted enhancement on the features of each variable based on the dynamic weights to obtain weighted enhanced features; extracting temporal dependencies from the weighted enhanced features and outputting time series features.

[0009] In some embodiments of this application, a multi-head cross-attention calculation module is constructed, which uses temporal features as query vectors and discrete meteorological features as key vectors and value vectors, performs cross-modal feature alignment operations, and outputs meteorological enhancement features. This includes: mapping temporal features to query vectors, mapping discrete meteorological features to key vectors and value vectors respectively; calculating multi-head cross-attention weights based on query vectors and key vectors, and performing weighted aggregation on value vectors based on multi-head cross-attention weights to output meteorological enhancement features aligned with temporal features.

[0010] In some embodiments of this application, an adaptive residual gating module is configured to perform weighted fusion of temporal features and meteorological enhancement features through learnable gating parameters to generate multimodal fusion features. This includes: generating fusion weights based on learnable gating parameters; determining the contribution ratio of temporal features and meteorological enhancement features in weighted fusion based on the fusion weights; and performing weighted summation of temporal features and meteorological enhancement features to generate multimodal fusion features.

[0011] In some embodiments of this application, deep temporal modeling is performed on the multimodal fusion features to output the wind power prediction results of the target wind farm, including: performing residual temporal processing and self-attention temporal processing on the multimodal fusion features to obtain deep temporal features; and outputting the wind power prediction results of the target wind farm based on the deep temporal features. The wind power prediction results include point prediction results and interval prediction results.

[0012] A second aspect of this application provides a wind power prediction device based on dual-branch cross-modal alignment, comprising: a dual-branch feature extraction architecture, including a wide-area branch and a local branch, wherein the wide-area branch and the local branch perform feature extraction in parallel; the wide-area branch is used to perform discrete representation encoding on the regional grid meteorological data covering the target wind farm, and output discrete meteorological features; the local branch is used to perform weighted enhancement on the multivariate time-series data of the target wind farm and extract time-series dependencies, and output time-series features; a multi-head cross-attention calculation module is used to perform cross-modal feature alignment operation with time-series features as query vectors and discrete meteorological features as key vectors and value vectors, and output meteorological enhancement features; an adaptive residual gating module is used to perform weighted fusion of time-series features and meteorological enhancement features through learnable gating parameters to generate multimodal fusion features; and a time-series modeling and prediction module is used to perform deep time-series modeling on the multimodal fusion features and output the wind power prediction result of the target wind farm.

[0013] In some embodiments of this application, the multi-head cross-attention calculation module is used to map temporal features into query vectors and discrete meteorological features into key vectors and value vectors, respectively; calculate multi-head cross-attention weights based on query vectors and key vectors, and perform weighted aggregation on value vectors based on multi-head cross-attention weights to output meteorological enhancement features aligned with temporal features.

[0014] In some embodiments of this application, the adaptive residual gating module is used to generate fusion weights based on learnable gating parameters, determine the contribution ratio of temporal features and meteorological enhancement features in weighted fusion based on the fusion weights, and perform weighted summation on the temporal features and meteorological enhancement features to generate multimodal fusion features.

[0015] This application also provides an electronic device, comprising: a memory for storing at least one instruction; and a processor for calling the instruction stored in the memory to execute the wind power prediction method based on dual-branch cross-modal alignment in the first aspect and any embodiment of the first aspect.

[0016] This application also provides a computer-readable storage medium storing at least one executable instruction, which is loaded and executed by a processor to implement the wind power prediction method based on dual-branch cross-modal alignment in the first aspect and any embodiment of the first aspect described above.

[0017] This application also provides a computer program product, which includes: computer program code, which, when executed by a computer, causes the computer to perform the wind power prediction method based on bi-branch cross-modal alignment in the first aspect and any embodiment of the first aspect.

[0018] Compared with the prior art, this application has the following advantages: By performing feature extraction in parallel through wide-area and local branches, regional grid meteorological data and multivariate time-series data of the target wind farm are differentiated according to their respective data characteristics, avoiding the high-dimensional redundancy and low representation efficiency problems caused by direct splicing. Cross-modal feature alignment is performed using time-series features as query vectors and discrete meteorological features as key and value vectors, enabling regional grid meteorological data to be dynamically selected and focused according to the current time-series state of the target wind farm, improving the matching ability between meteorological context and local prediction needs. Through an adaptive residual gating module, time-series features and meteorological enhancement features are weighted and fused, and the contribution ratio of the two types of features can be dynamically adjusted based on learnable gating parameters, reducing the risk of information loss caused by fixed-weight fusion or simple splicing. Deep time-series modeling based on multimodal fusion features can retain the local time-series dependence of the target wind farm while introducing regional meteorological evolution information, thereby improving the prediction accuracy and stability of wind power prediction results. Attached Figure Description

[0019] Figure 1 This is a flowchart of a wind power prediction method based on dual-branch cross-modal alignment provided in an embodiment of this application; Figure 2This is a schematic diagram of a multi-head cross-attention mechanism for aligning wide-area features and local features, provided in an embodiment of this application. Figure 3 This is a schematic diagram of an adaptive residual gating structure for dynamic fusion provided in an embodiment of this application; Figure 4 This application provides an overall diagram of wind power prediction results and enlarged partial diagrams of four typical time periods. Figure 5 This is a quantitative index result chart with a 95% confidence level provided in the embodiments of this application; Figure 6 This is a structural block diagram of a wind power prediction device based on dual-branch cross-modal alignment provided in an embodiment of this application. Detailed Implementation

[0020] The present application will now be further described with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solutions of the present application and should not be construed as limiting the scope of protection of the present application. It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of the present application.

[0021] The acquisition, transmission, storage, use, and processing of data in this application comply with relevant national laws and regulations. In the embodiments of this application, certain existing industry solutions such as software, components, and models may be mentioned. These should be considered exemplary, intended only to illustrate the feasibility of implementing the technical solution of this application, and do not imply that the applicant has already used or necessarily used such solutions.

[0022] All terms used in this application have the same meaning as understood by one of ordinary skill in the art to which this application pertains, unless otherwise specifically defined. It should also be understood that terms defined in general dictionaries should be interpreted as having meanings consistent with their meanings in the context of the relevant art, and not as idealized or highly formalized, unless expressly defined herein.

[0023] For ease of understanding, the technical terms appearing in the embodiments of this application are explained in advance. Unless otherwise specified, the following definitions apply throughout the entire specification.

[0024] Regional grid meteorological data refers to gridded meteorological data covering the target wind farm and its surrounding area. Regional grid meteorological data can be derived from numerical weather prediction data and can include one or more of wind speed, wind direction, temperature, air pressure, and precipitation rate.

[0025] Multivariate time series data refers to the data of multiple variables collected locally at the target wind farm over time. Multivariate time series data can include one or more of the following: wind speed, wind direction, temperature, air pressure, humidity, and historical power.

[0026] Discrete representation coding refers to the process of mapping continuous high-dimensional regional grid meteorological data into discrete codebook indices and corresponding meteorological feature vectors.

[0027] Discrete meteorological features refer to the features obtained after discrete characterization and encoding, which are used to characterize regional meteorological evolution patterns.

[0028] Temporal features refer to the features obtained by performing weighted enhancement on the multivariate temporal data of the target wind farm and extracting temporal dependencies in the local branch.

[0029] Meteorological enhancement features refer to the features obtained by the multi-head cross-attention calculation module after performing cross-modal feature alignment operations on discrete meteorological features based on temporal features.

[0030] Multimodal fusion features refer to features generated by the adaptive residual gating module through weighted fusion of time-series features and meteorological enhancement features using learnable gating parameters.

[0031] Example 1; This embodiment provides a wind power prediction method based on dual-branch cross-modal alignment, such as... Figure 1 As shown, the process includes steps S11 to S14.

[0032] S11. Build a dual-branch feature extraction architecture, which includes a wide-area branch and a local branch. The wide-area branch and the local branch perform feature extraction in parallel, and output discrete meteorological features and time-series features respectively.

[0033] In this embodiment, the wide-area branch receives regional grid meteorological data covering the target wind farm, while the local branch receives multivariate time-series data of the target wind farm. To enable the two types of data to be used interchangeably in subsequent cross-modal feature alignment operations, the regional grid meteorological data can be processed to unify the temporal resolution before feature extraction. For example, conformal piecewise cubic Hermite interpolation is used for wind speed, cyclic interpolation is used for wind direction, cubic spline interpolation is used for temperature and air pressure, and precipitation rate is distributed uniformly to match the regional grid meteorological data with the multivariate time-series data on the time scale.

[0034] The wide-area branch performs discrete representation encoding on regional grid meteorological data, outputting discrete meteorological features. Specifically, regional grid meteorological data can be input into a vector quantization meteorological encoder, which maps continuous high-dimensional regional grid meteorological data into discrete codebook indices and corresponding meteorological feature vectors. The vector quantization meteorological encoder can include an encoder, a discrete codebook, and a decoder. After pre-training, the encoder and discrete codebook can be frozen, allowing them to be used as a deterministic feature extractor, thereby reducing the dimensionality of the regional grid meteorological data while preserving meteorological evolution patterns.

[0035] The local branch performs weighted augmentation on the multivariate time-series data of the target wind farm and extracts temporal dependencies, outputting temporal features. Specifically, features can be extracted from each variable in the multivariate time-series data first, and the dynamic weights corresponding to each variable can be calculated. Then, the features of each variable can be weighted and augmented based on the dynamic weights to obtain weighted augmented features. Temporal dependencies can then be extracted from the weighted augmented features to output temporal features. For example, an independent feature processing unit can be set for each variable through a multivariate feature augmentation network, and dynamic weights can be calculated using Softmax. Finally, a minimum long short-term memory network can be used to extract long-term dependencies.

[0036] S12, build a multi-head cross-attention calculation module, use time-series features as query vectors and discrete meteorological features as key vectors and value vectors, perform cross-modal feature alignment operations, and output meteorological enhancement features.

[0037] In this embodiment, temporal features are used to characterize the local state of the target wind farm, and discrete meteorological features are used to characterize the regional meteorological evolution pattern covering the target wind farm. The multi-head cross-attention calculation module maps the temporal features to query vectors and the discrete meteorological features to key vectors and value vectors, respectively. It calculates the multi-head cross-attention weights based on the query vectors and key vectors, and performs weighted aggregation on the value vectors based on the multi-head cross-attention weights to output meteorological enhancement features aligned with the temporal features.

[0038] For example, the query vector represents the current predicted demand of the target wind farm, the key vector represents the searchable and matching meteorological context in the discrete meteorological features, and the value vector provides the meteorological information aggregated by attention weights. Through the above processing, the regional grid meteorological data is no longer a fixed input in a simple splicing form, but is dynamically focused according to the temporal features, thereby achieving cross-modal feature alignment.

[0039] S13 is equipped with an adaptive residual gating module, which uses learnable gating parameters to weight and fuse temporal features and meteorological enhancement features to generate multimodal fusion features.

[0040] In this embodiment, the adaptive residual gating module generates fusion weights based on learnable gating parameters, determines the contribution ratio of temporal features and meteorological enhancement features in weighted fusion based on the fusion weights, and performs a weighted summation of the temporal features and meteorological enhancement features to generate multimodal fusion features. By retaining temporal features and introducing meteorological enhancement features, the loss of local temporal autocorrelation information during feature projection or cross-modal alignment can be reduced.

[0041] S14 performs deep time-series modeling on the multimodal fusion features and outputs the wind power prediction results of the target wind farm.

[0042] In this embodiment, residual temporal processing and self-attention temporal processing can be performed on the multimodal fusion features to obtain deep temporal features; based on the deep temporal features, the wind power prediction results of the target wind farm are output. The wind power prediction results can include point prediction results and interval prediction results. For interval prediction results, one or more upper and lower bounds of the prediction interval at confidence levels can be output, such as upper and lower bounds of the prediction interval at 80%, 90%, 95%, or 99% confidence levels.

[0043] In one optional training method, the model can be trained using a joint optimization loss function formed by weighting quantile loss, physical constraint loss, and interval quality loss. Specifically, quantile loss optimizes point prediction results and / or interval prediction results, physical constraint loss ensures the predicted power conforms to the physical range of wind power, and interval quality loss constrains the coverage and width of the prediction interval, thereby improving the reliability of wind power prediction results.

[0044] The method described in the embodiments is described in its entirety below. The method as a whole includes the following three stages.

[0045] Phase 1: Two-branch spatiotemporal feature extraction and representation: (1) Wide-area branching: Discrete representation learning strategy; The wide-area branch employs a Vector Quantized Weather Encoder (VQWE) to process high-dimensional wide-area numerical weather prediction (NWP) data. The VQWE training architecture includes an encoder, a discrete codebook, and a decoder.

[0046] (2) Local branching: Temporal feature extraction; The input to the local branch is local meteorological data (wind speed, wind direction, temperature, air pressure, humidity) and historical power data. First, a Multi-Variable Feature Extraction Network (MVFEN) is used to independently process each variable using a Dynamic Representation Network (DRN), and Softmax is used to calculate dynamic weights. These weighted data are then fused to obtain a comprehensive representation. Subsequently, the data is input into a lightweight Long Short-Term Memory (minLSTM) network to capture long-term dependencies and output local temporal features.

[0047] Phase Two: Multimodal Feature Fusion In order to effectively integrate wide-area meteorological evolution characteristics into local power prediction, this stage constructs a meteorological context perception layer based on a multi-head cross-attention mechanism, aiming to establish dynamic mapping and feature alignment between multi-source data.

[0048] like Figure 2 As shown, this process first involves temporal characteristics. Discrete meteorological characteristics Three independent linear projection layers are input separately to align the feature dimensions, mapping them to a query matrix, a key matrix, and a value matrix, respectively. These three projected features are then fed into a Scaled Dot-Product Attention module to perform cross-attention calculations, uncovering the correlation between local temporal information and wide-area meteorological information. The features output from the multi-head attention are then concatenated. The concatenated result is passed through a linear layer for feature fusion and dimension mapping, ultimately outputting a meteorological enhancement feature that integrates local temporal information and wide-area meteorological information. The calculation process is as follows: ; ; in, The input feature matrix is, i.e. or Q is the query matrix; K is the key matrix; V is the value matrix; Let Q be a learnable weight matrix, K be a key matrix, and V be a value matrix, respectively, used to transform the input features into the desired query, key, and value spaces; This represents the transpose of matrix K; This indicates that the dot product of matrix Q and matrix K is used to calculate the similarity score matrix. d kThis represents the dimension of each attention head, used to scale the inner product result and avoid excessively large inner product values ​​when the feature dimension is high.

[0049] Therefore, the wide-area meteorological representation (i.e., meteorological enhancement features) aligned based on local feature queries is as follows: .

[0050] Although the above process extracts wide-area meteorological features aligned with local features, the output vector loses the autocorrelation of local time series during projection. To address this issue, an adaptive residual gating mechanism is designed as shown below. By introducing learnable parameters, the model can adaptively adjust the weights between wide-area features and local time series features. The gating structure is as follows: Figure 3 As shown, the parameters can be learned first. Feed into the Sigmoid activation function The parameters are mapped to the interval (0,1) to obtain the adaptive fusion weights. ; Utilizing weights and complementary weights Local temporal features With cross-attention output features Perform scalar weighting; add the two weighted features to output the multimodal fusion feature. .

[0051] Finally, the local time series representation after enhancement by the wide-area meteorological model is as follows: ; in, The multimodal fusion features are obtained after adaptive weighted fusion; It is a time-series feature; This is a meteorological enhancement feature; This represents the fusion weight, with a value range of (0,1); These are learnable fusion gating parameters, with values ​​ranging from (0,1), which can be dynamically learned through gradient backpropagation based on the loss function of the prediction task.

[0052] By incorporating wide-area meteorological field information, local features are enhanced spatially. To further explore the long-range temporal dependencies in these fused features, this stage further employs the proposed Deep Residual Network (DRN) and multi-head self-attention mechanism to ensure training stability during deep temporal processing. The calculation process is as follows: ; in, Features representing the output of the self-attention mechanism; and These are the outputs of internal residual connections and cross-layer residual connections, respectively; This indicates the output features of the pre-feature mapping; This represents multimodal fusion features; LN stands for Layer Normalization; DRN stands for Dynamic Representation Network. The residual extraction process includes two residual optimization pathways: the first pathway: [details of the self-attention output features]. Layer normalization is performed, and the normalization result is fed into the Dynamic Representation Network (DRN) to extract the residual offset; the residual offset is then compared with the original features. Adding them together yields the residual optimization features. Second path: mapping features from the preceding path Layer normalization is performed, and feature correction values ​​are learned through DRN; the obtained correction values ​​are then fused with adaptive gating features. Summation is performed to supplement cross-path feature information and output residual optimization features. .

[0053] Phase 3, Final Prediction Output: In the final stage, the deeply fused features are mapped to the power prediction output, simultaneously optimizing the performance of deterministic and probabilistic predictions.

[0054] To fully verify the effectiveness of the method proposed in this application, this section conducts systematic experiments on real wind farm data. All experiments were conducted in a unified computing environment: the hardware platform consisted of an Intel Core i7-13700K CPU and an NVIDIA GeForce RTX 4070 Ti SUPER GPU, and the software environment consisted of Python 3.10, PyTorch 2.7, and CUDA 11.8.

[0055] The experimental data includes wind farm-level operational data and NWP data. Wind farm data was obtained from an offshore wind farm in Province A (306 MW installed capacity, 36 turbines) and an onshore wind farm in Province B (60 MW installed capacity, 40 turbines), including variables such as wind speed, wind direction, temperature, air pressure, humidity, and power output, with a time resolution of 15 minutes. NWP data came from the weather forecast center, covering the area of ​​Province A (27.0°N~31.1°N, 118.0°E~123.1°E, 0.25°×0.25°) and Province B (40.5°N~46.2°N, 121.3°E~131.2°E, 0.1°×0.1°), with a time resolution of 1 hour, including variables such as wind speed, wind direction, temperature, air pressure, and precipitation rate. To unify the temporal resolution of multi-source data, PCHIP interpolation (wind speed), cyclic interpolation (wind direction), cubic spline interpolation (temperature and air pressure), and uniform distribution (precipitation rate) were used for temporal interpolation of NWP data.

[0056] The baseline models compared to the baseline models included: (1) SA-DeepAR (RNN architecture), (2) LSTMformer (Transformer+LSTM hybrid), (3) PatchTST (Transformer baseline), (4) TimeGrad (diffusion autoregressive model), and (5) DSDP (diffusion probability architecture). All baseline hyperparameters were Bayesian optimized using the Optuna framework.

[0057] Considering four confidence levels (80%, 90%, 95%, 99%) and four prediction step sizes (4 steps / 1 hour, 32 steps / 8 hours, 96 steps / 24 hours, 192 steps / 48 hours), four typical time periods were selected for visualization: low power output (period A), rated power output (period B), severe fluctuations (period C), and sudden wind speed changes (period D). The visualization results are as follows: Figure 4 As shown. Figure 4 As can be seen, the solid black line represents the actual power, the dashed red line represents the power predicted by the method of this application, the gray area represents the prediction error, and the multi-layered blue filled area represents the probability prediction interval at different confidence levels. During the low-power period A and the rated power period B, the predicted power output by this application stably converges to the physically feasible interval (0 to the rated power of the wind farm), effectively avoiding abnormal problems such as negative power or power exceeding limits in the prediction results, which do not conform to the physical laws of wind power operation. For the scenarios of high-frequency and violent power fluctuations in period C and rapid power changes caused by sudden wind speed changes in period D, the prediction curve can closely track the trend of actual power changes. The prediction intervals at different confidence levels can reasonably wrap the actual power sequence, and the interval boundaries fit the power fluctuation amplitude, without the defects of excessive interval expansion or the actual value deviating significantly from the interval.

[0058] Quantitative index results as follows Figure 5 The figures are shown (indicators are calculated based on 95% confidence intervals). In the figures, R² represents the coefficient of determination, RMSE represents the root mean square error, MAE represents the mean absolute error, PICP represents the prediction interval coverage, PINAW represents the average bandwidth of the prediction interval, and |ACE| represents the absolute error of the interval coverage. In the 4-step lookahead short-term prediction scenario, the proposed method achieves R²=0.973, RMSE=19.614 MW, and MAE=11.948 MW, all of which are superior to all baselines. In the 192-step prediction, the proposed method achieves RMSE=38.739 MW, which is 9.55% and 3.64% lower than TimeGrad (41.588 MW) and DSDP (40.880 MW), respectively, and 3.64% lower than the best state-of-the-art baseline. In terms of interval prediction, the |ACE| of this application remains within 4% at all confidence levels (for example, the maximum |ACE| is 1.652% at 95% confidence level), while the baseline model generally suffers from insufficient coverage or excessively wide intervals.

[0059] Example 2; This embodiment provides a wind power prediction device based on dual-branch cross-modal alignment, such as... Figure 6 As shown, the wind power prediction device 100 includes a dual-branch feature extraction architecture 110, a multi-head cross-attention calculation module 120, an adaptive residual gating module 130, and a time series modeling prediction module 140.

[0060] The dual-branch feature extraction architecture 110 includes a wide-area branch and a local branch, which perform feature extraction in parallel. The wide-area branch is used to perform discrete representation encoding on the regional grid meteorological data covering the target wind farm and output discrete meteorological features. The local branch is used to perform weighted enhancement on the multivariate time series data of the target wind farm and extract time series dependencies, and output time series features. The multi-head cross-attention calculation module 120 is used to perform cross-modal feature alignment operations using temporal features as query vectors and discrete meteorological features as key vectors and value vectors, and output meteorological enhancement features. The adaptive residual gating module 130 is used to perform weighted fusion of temporal features and meteorological enhancement features through learnable gating parameters to generate multimodal fusion features; The time series modeling and prediction module 140 is used to perform deep time series modeling on multimodal fusion features and output the wind power prediction results of the target wind farm.

[0061] In some embodiments of this application, the regional grid meteorological data includes wind speed, wind direction, temperature, air pressure and precipitation rate covering the area surrounding the target wind farm, and the multivariate time series data includes wind speed, wind direction, temperature, air pressure, humidity and historical power of the target wind farm; before performing feature extraction, the regional grid meteorological data is processed to unify the time resolution so that the regional grid meteorological data and the multivariate time series data are matched in time scale.

[0062] In some embodiments of this application, the wide-area branch includes a vector quantization meteorological encoder, which is used to map continuous high-dimensional regional grid meteorological data into discrete codebook indices and corresponding meteorological feature vectors, and to use the discrete codebook indices and corresponding meteorological feature vectors as discrete meteorological features.

[0063] In some embodiments of this application, the local branch is used to extract features from each variable in the multivariate time series data, calculate the dynamic weights corresponding to each variable, and perform weighted enhancement on the features of each variable based on the dynamic weights to obtain weighted enhanced features; and extract time-series dependencies from the weighted enhanced features to output time-series features.

[0064] In some embodiments of this application, the multi-head cross-attention calculation module 120 is used to map temporal features into query vectors and discrete meteorological features into key vectors and value vectors respectively; calculate multi-head cross-attention weights based on query vectors and key vectors, and perform weighted aggregation on value vectors based on multi-head cross-attention weights to output meteorological enhancement features aligned with temporal features.

[0065] In some embodiments of this application, the adaptive residual gating module 130 is used to generate fusion weights based on learnable gating parameters, determine the contribution ratio of temporal features and meteorological enhancement features in weighted fusion based on the fusion weights, and perform weighted summation on the temporal features and meteorological enhancement features to generate multimodal fusion features.

[0066] In some embodiments of this application, the time series modeling prediction module 140 is used to perform residual time series processing and self-attention time series processing on the multimodal fusion features to obtain deep time series features; and output the wind power prediction results of the target wind farm based on the deep time series features. The wind power prediction results include point prediction results and interval prediction results.

[0067] For details and benefits of the wind power prediction device based on dual-branch cross-mode alignment provided in the embodiments of this application, please refer to the above description of the wind power prediction method based on dual-branch cross-mode alignment, which will not be repeated here.

[0068] This application also provides an electronic device, comprising: a memory for storing at least one instruction; and a processor for calling the instruction stored in the memory to execute the wind power prediction method based on dual-branch cross-modal alignment in any of the above embodiments.

[0069] This application also provides a computer-readable storage medium storing at least one executable instruction, which is loaded and executed by a processor to implement the wind power prediction method based on dual-branch cross-modal alignment in any of the above embodiments.

[0070] This application also provides a computer program product, which includes: computer program code, which, when executed by a computer, causes the computer to perform the wind power prediction method based on bi-branch cross-modal alignment in any of the above embodiments.

[0071] 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 embodied 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.

[0072] 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.

[0073] 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.

[0074] 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.

[0075] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0076] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0077] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0078] It should be noted that the terms "first," "second," and similar terms used in this application do not indicate any order, quantity, or importance, but are merely used to distinguish different parts. Terms such as "including" or "comprising" mean that the element preceding the word covers the element listed after the word, and do not exclude the possibility of covering other elements as well.

[0079] Although the operations are described in a specific order in the accompanying drawings in this application, it should not be construed as requiring these operations to be performed in the specific order or serial order shown, or requiring all of the shown operations to obtain the desired result. In certain environments, multitasking and parallel processing may be advantageous.

[0080] Finally, it should be noted that the above content is only used to illustrate the technical solution of this application, and is not intended to limit the scope of protection of this application. Simple modifications or equivalent substitutions made by those skilled in the art to the technical solution of this application shall not depart from the substance and scope of the technical solution of this application.

Claims

1. A wind power prediction method based on dual-branch cross-modal alignment, characterized in that, include: A dual-branch feature extraction architecture is constructed, which includes a wide-area branch and a local branch, and the wide-area branch and the local branch perform feature extraction in parallel; The wide-area branch performs discrete representation encoding on the regional grid meteorological data covering the target wind farm and outputs discrete meteorological features; the local branch performs weighted enhancement on the multivariate time series data of the target wind farm and extracts time series dependencies, outputting time series features. A multi-head cross-attention calculation module is built, which uses the temporal features as query vectors and the discrete meteorological features as key vectors and value vectors to perform cross-modal feature alignment operations and output meteorological enhancement features. An adaptive residual gating module is configured to perform weighted fusion of the temporal features and the meteorological enhancement features through learnable gating parameters to generate multimodal fusion features; Deep time-series modeling is performed on the multimodal fusion features to output the wind power prediction results of the target wind farm.

2. The wind power prediction method based on dual-branch cross-modal alignment according to claim 1, characterized in that, The regional grid meteorological data includes wind speed, wind direction, temperature, air pressure, and precipitation rate covering the area surrounding the target wind farm. The multivariate time series data includes wind speed, wind direction, temperature, air pressure, humidity, and historical power of the target wind farm. Before performing feature extraction, the regional grid meteorological data is processed to unify the time resolution so that the regional grid meteorological data matches the multivariate time series data on the time scale.

3. The wind power prediction method based on dual-branch cross-modal alignment according to claim 1, characterized in that, The step of performing discrete representation encoding on the regional grid meteorological data covering the target wind farm through the wide-area branch, and outputting discrete meteorological features, includes: The regional grid meteorological data is input into a vector quantization meteorological encoder. The vector quantization meteorological encoder maps the continuous high-dimensional regional grid meteorological data into discrete codebook indices and corresponding meteorological feature vectors. The discrete codebook indices and corresponding meteorological feature vectors are then used as the discrete meteorological features.

4. The wind power prediction method based on dual-branch cross-modal alignment according to claim 1, characterized in that, The step of performing weighted enhancement and extracting temporal dependencies on the multivariate time-series data of the target wind farm through the local branch, and outputting time-series features, includes: Feature extraction is performed on each variable in the multivariate time series data, dynamic weights are calculated for each variable, and the features of each variable are weighted and enhanced based on the dynamic weights to obtain weighted enhanced features. Extract the temporal dependency from the weighted enhanced features and output the temporal features.

5. The wind power prediction method based on dual-branch cross-modal alignment according to claim 1, characterized in that, The multi-head cross-attention calculation module, using the temporal features as the query vector and the discrete meteorological features as the key and value vectors, performs cross-modal feature alignment operations and outputs meteorological enhancement features, including: The time-series features are mapped to query vectors, and the discrete meteorological features are mapped to key vectors and value vectors, respectively. Calculate the multi-head cross-attention weights based on the query vector and the key vector, and perform weighted aggregation on the value vector based on the multi-head cross-attention weights to output the meteorological enhancement features aligned with the temporal features.

6. The wind power prediction method based on dual-branch cross-modal alignment according to claim 1, characterized in that, The configured adaptive residual gating module performs weighted fusion of the temporal features and the meteorological enhancement features through learnable gating parameters to generate multimodal fusion features, including: Generate fusion weights based on the learnable gating parameters; Based on the fusion weights, the contribution ratios of the temporal features and the meteorological enhancement features in the weighted fusion are determined, and the temporal features and the meteorological enhancement features are weighted and summed to generate the multimodal fusion features.

7. The wind power prediction method based on dual-branch cross-modal alignment according to claim 1, characterized in that, The step of performing deep temporal modeling on the multimodal fusion features and outputting the wind power prediction results of the target wind farm includes: The multimodal fusion features are subjected to residual temporal processing and self-attention temporal processing to obtain deep temporal features; Based on the deep time-series features, the wind power prediction results of the target wind farm are output, including point prediction results and interval prediction results.

8. A wind power prediction device based on dual-branch cross-modal alignment, characterized in that, include: A dual-branch feature extraction architecture includes a wide-area branch and a local branch, wherein the wide-area branch and the local branch perform feature extraction in parallel; The wide-area branch is used to perform discrete representation encoding on the regional grid meteorological data covering the target wind farm and output discrete meteorological features; the local branch is used to perform weighted enhancement and extract time-series dependencies on the multivariate time-series data of the target wind farm and output time-series features. The multi-head cross-attention calculation module is used to perform cross-modal feature alignment operation with the time-series features as query vectors and the discrete meteorological features as key vectors and value vectors, and output meteorological enhancement features. An adaptive residual gating module is used to perform weighted fusion of the temporal features and the meteorological enhancement features through learnable gating parameters to generate multimodal fusion features; The time-series modeling and prediction module is used to perform deep time-series modeling on the multimodal fusion features and output the wind power prediction results of the target wind farm.

9. The wind power prediction device based on dual-branch cross-modal alignment according to claim 8, characterized in that, The multi-head cross-attention calculation module is used to map the temporal features into query vectors and the discrete meteorological features into key vectors and value vectors, respectively; calculate the multi-head cross-attention weights based on the query vectors and the key vectors, and perform weighted aggregation on the value vectors based on the multi-head cross-attention weights to output the meteorological enhancement features aligned with the temporal features.

10. The wind power prediction device based on dual-branch cross-modal alignment according to claim 8, characterized in that, The adaptive residual gating module is used to generate fusion weights based on the learnable gating parameters, determine the contribution ratio of the temporal features and the meteorological enhancement features in the weighted fusion based on the fusion weights, and perform weighted summation on the temporal features and the meteorological enhancement features to generate the multimodal fusion features.

Images