Wind power prediction method based on spatial-variable dual perception and adaptive fusion

Abstract

The application provides a wind power prediction method based on space-variable dual perception and adaptive fusion, and relates to the field of wind power prediction. The application proposes a wind power prediction scheme which fuses historical power sequence, point NWP data and regional two-dimensional NWP field. The scheme is based on a three-module collaborative structure including a time encoder, a space encoder and an adaptive fusion module, and realizes high-precision prediction of wind power through multi-level feature extraction and dynamic fusion. Especially at the adaptive fusion module, the scheme skillfully projects the combined time sequence feature sequence and the space feature sequence to a unified latent space, and learns the dynamic weight of different fusion strategies, realizing on-demand adaptive fusion of multi-source information, so that the finally generated wind power prediction result is reliable under various weather conditions. In addition, the space encoder adopts a training method of "independent pre-training" plus "joint stage freezing", which ensures the stability of the space feature expression.

Description

Technical Field

[0001] This invention relates to the field of wind power prediction, and specifically to a wind power prediction method based on spatial-variable dual perception and adaptive fusion. Background Technology

[0002] Wind power, as a key renewable energy source for achieving carbon neutrality and sustainable development, plays a crucial role in ensuring grid stability, reducing operational risks, and promoting the economic integration of new energy sources. However, the inherent randomness and volatility of wind energy pose significant challenges to the safe and stable operation of the power system. Therefore, high-precision wind power forecasting has become a core technology for solving this problem, as its accuracy directly determines the efficiency of grid dispatching and the level of new energy integration.

[0003] To improve forecast accuracy, related technologies generally adopt data-driven methods, integrating historical operational data with meteorological forecast information for modeling. A typical approach involves multi-source information fusion but employing a static fusion strategy. Specifically, to comprehensively utilize heterogeneous multi-source information, some advanced solutions attempt feature-level or decision-level fusion. The most representative example is the framework proposed by Jonkers et al. in 2024 (Jef Jonkers, Diego Nieves Avendano, Glenn VanWallendael, and Sofie Van Hoecke, “A novel day-ahead regional and probabilistic wind power forecasting framework using deep CNNs and conformalized regression forests,” Applied Energy, vol. 361, pp. 122900, 2024.). This work first uses a deep convolutional neural network (CNN) to extract spatial features from the future regional two-dimensional NWP field, then concatenates these features with the statistical features of historical power sequences, and finally inputs them into a regression forest model for final prediction.

[0004] However, the above methods have a static and fixed mechanism for fusing regional spatial information, which means that the same stitching method is used in all weather scenarios. The rigid fusion strategy makes it difficult to meet the actual needs in terms of prediction accuracy and robustness in complex terrain and variable weather scenarios. Summary of the Invention

[0005] (a) Technical problems to be solved To address the shortcomings of existing technologies, this invention provides a wind power prediction method based on spatial-variable dual sensing and adaptive fusion, which solves the technical problem that fixed fusion methods lack adaptability.

[0006] (II) Technical Solution To achieve the above objectives, the present invention provides the following technical solution: A wind power prediction method based on spatial-variable dual sensing and adaptive fusion includes: The system acquires historical power sequences of the wind farm turbines to be predicted, point-based NWP data for multiple future time steps, and regional two-dimensional NWP fields. The point-based NWP data refers to hourly weather forecast data at the wind farm turbines to be predicted, and the regional two-dimensional NWP fields refer to two-dimensional meteorological field data intercepted within the target spatial range at the center of the wind farm. The historical power sequence and the point NWP data are used as inputs to the time encoder, and after being encoded separately, they are fused by cross-attention to generate a fused time-series feature sequence. The two-dimensional NWP field of the region is used as the input of the spatial encoder. After feature extraction, it is weighted by the variable perception mechanism and the spatial perception mechanism in sequence to generate a spatial feature sequence. The variable perception mechanism is used to dynamically adjust the importance of different meteorological variables to the prediction, and the spatial perception mechanism is used to capture key spatial areas in the regional wind field. The temporal feature sequence and the spatial feature sequence are used as inputs to the adaptive fusion module. After being projected onto a unified latent space, an adaptive weighting mechanism based on multiple strategies is used to select the optimal fusion method to generate wind power prediction results for multiple future time steps.

[0007] Preferably, the step of using the two-dimensional NWP field of the region as input to the spatial encoder, performing feature extraction, and then sequentially weighting it through a variable-aware mechanism and a spatial-aware mechanism to generate spatial features includes: The original spatial feature map of the two-dimensional NWP field in the region is extracted using a feature extractor; Based on the original spatial feature map, global mean pooling and max pooling are used as variable features, and variable-aware weights are calculated. Based on the original spatial feature map and the variable-perceived weights, a first spatial feature map after variable weighting is generated. Based on the first spatial feature map, global mean pooling and max pooling are used as variable features, and spatial awareness weights are calculated. Based on the first spatial feature map and the spatial perception weights, a spatially weighted second spatial feature map is generated. Based on the second spatial feature map, the spatial feature sequence is generated through dimensionality reduction mapping.

[0008] Preferably, the feature extractor is any one of CNN, UNet, ViT, or 3D-CNN.

[0009] Preferably, the adoption of a multi-strategy-based adaptive weighting mechanism to select the optimal fusion method means: Multiple complementary fusion strategies are predefined. Based on the projection results of the temporal feature sequence and the spatial feature sequence, the weights of each fusion strategy are automatically determined by introducing learnable weight coefficients.

[0010] Preferably, the fusion strategy is one or a combination of several of the following: identity mapping, attention fusion, GRU fusion, or MLP fusion.

[0011] Preferably, during the training phase of the wind power prediction method, an overall optimization objective is constructed, which includes overall point error loss, trend consistency loss, and peak segment weighted loss.

[0012] Preferably, during the training phase of the wind power prediction method, the spatial encoder is pre-trained separately and its network parameters are frozen before joint training is performed based on the overall optimization objective.

[0013] A wind power prediction system based on spatial-variable dual sensing and adaptive fusion includes: The data acquisition unit is used to acquire the historical power sequence of the wind farm's turbines to be predicted, point-based NWP data for multiple future time steps, and regional two-dimensional NWP field; wherein, the point-based NWP data refers to the hourly meteorological forecast data at the wind farm's turbines to be predicted, and the regional two-dimensional NWP field refers to the two-dimensional meteorological field data intercepted by the wind farm center within the target spatial range. The temporal coding unit is used to take the historical power sequence and the point NWP data as input to the time encoder, encode them separately, and then use cross-attention fusion to generate a fused temporal feature sequence. The spatial coding unit is used to take the two-dimensional NWP field of the region as the input of the spatial encoder, perform feature extraction, and then perform weighting through the variable perception mechanism and the spatial perception mechanism in sequence to generate a spatial feature sequence; wherein, the variable perception mechanism is used to dynamically adjust the importance of different meteorological variables to the prediction, and the spatial perception mechanism is used to capture key spatial areas in the regional wind field. An adaptive fusion unit is used to take the temporal feature sequence and the spatial feature sequence as inputs to the adaptive fusion module, project them onto a unified latent space, and then use an adaptive weighting mechanism based on multiple strategies to select the optimal fusion method to generate wind power prediction results for multiple future time steps.

[0014] A storage medium storing a computer program, wherein the computer program causes a computer to execute the wind power prediction method based on spatial-variable dual sensing and adaptive fusion as described above.

[0015] An electronic device, comprising: One or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the programs including methods for performing the wind power prediction method based on spatial-variable dual sensing and adaptive fusion as described above.

[0016] (III) Beneficial Effects This invention provides a wind power prediction method based on spatial-variable dual sensing and adaptive fusion. Compared with existing technologies, it has the following advantages: This invention proposes a wind power prediction scheme that integrates historical power sequences, point-based NWP data, and regional two-dimensional NWP fields. Based on a three-module collaborative structure comprising a time encoder, a spatial encoder, and an adaptive fusion module, this scheme achieves high-precision wind power prediction through multi-level feature extraction and dynamic fusion. Particularly noteworthy is the adaptive fusion module, which cleverly projects the combined temporal and spatial feature sequences onto a unified latent space and learns the dynamic weights of different fusion strategies. This enables on-demand adaptive fusion of multi-source information, ensuring reliable wind power prediction results under various weather conditions. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A block diagram of a wind power prediction method based on spatial-variable dual perception and adaptive fusion provided in an embodiment of the present invention; Figure 2 The flowchart illustrates a wind power prediction method based on spatial-variable dual perception and adaptive fusion, provided for an embodiment of the present invention. Detailed Implementation

[0019] 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 are described clearly and completely. Obviously, the described embodiments are only some embodiments of the present invention, 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.

[0020] This application provides a wind power prediction method based on spatial-variable dual perception and adaptive fusion, which solves the technical problem that fixed fusion methods lack adaptability, achieves high-precision prediction of wind power, and ensures the reliability of prediction results.

[0021] The technical solution in this application is to solve the above-mentioned technical problems, and the general idea is as follows: This invention employs a "three-module collaborative structure," designed specifically for the characteristics of historical power sequences, future point-based NWP data, and regional two-dimensional NWP fields; wherein: 1) Temporal Encoder (TE) is responsible for extracting temporal features from historical power sequences and point NWP data, and effectively associating the two types of temporal information through a cross-attention mechanism.

[0022] 2) Spatial Encoder (SE) is specifically designed to process two-dimensional NWP fields in (future) regions. It sequentially performs spatial convolution feature extraction, variable-aware weighting, and spatial region-aware weighting, and generates spatial feature sequences through dimensionality reduction mapping.

[0023] In particular, the spatial encoder has the structural advantage of independent pre-training before participating in joint prediction, which can effectively improve the stability of regional spatial features. Specifically, the spatial encoder employs a training method of "independent pre-training" plus "joint stage freezing": First, the spatial encoder is pre-trained independently to fully learn the spatial structural features (such as wind field distribution and local gradient) in the (future) two-dimensional NWP field of the region; in the subsequent multi-source joint prediction stage, the parameters of the spatial encoder remain frozen, participating only as a stable spatial feature extractor in the fusion process. This design avoids interference from high-dimensional noise during end-to-end training, thus ensuring the stability of spatial feature representation and improving prediction robustness under complex wind field conditions.

[0024] 3) Adaptive Fusion (AF) module By projecting temporal and spatial encoded features into a unified latent space and learning the dynamic weights of different fusion strategies, on-demand adaptive fusion of multi-source information can be achieved. Ultimately, wind power prediction for the next F time steps can be completed using a Multi-Layer Perceptron (MLP).

[0025] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0026] Example 1: like Figure 1 As shown, this embodiment of the invention provides a wind power prediction method based on spatial-variable dual sensing and adaptive fusion, including: S1. Obtain the historical power sequence of the wind farm's turbines to be predicted, point-based NWP data for multiple future time steps, and regional two-dimensional NWP field; wherein, the point-based NWP data refers to the hourly weather forecast data at the wind farm's turbines to be predicted, and the regional two-dimensional NWP field refers to the two-dimensional meteorological field data intercepted by the wind farm center within the target spatial range. S2. The historical power sequence and the point NWP data are used as inputs to the time encoder, and after encoding, cross-attention fusion is used to generate the fused time-series feature sequence. S3. The two-dimensional NWP field of the region is used as the input of the spatial encoder. After feature extraction, it is weighted by the variable perception mechanism and the spatial perception mechanism in sequence to generate a spatial feature sequence. The variable perception mechanism is used to dynamically adjust the importance of different meteorological variables to the prediction, and the spatial perception mechanism is used to capture key spatial areas in the regional wind field. S4. The temporal feature sequence and the spatial feature sequence are used as inputs to the adaptive fusion module. After being projected onto a unified latent space, the optimal fusion method is selected by an adaptive weighting mechanism based on multiple strategies to generate wind power prediction results for multiple future time steps.

[0027] This invention, based on a three-module collaborative structure comprising a time encoder, a spatial encoder, and an adaptive fusion module, achieves high-precision wind power prediction through multi-level feature extraction and dynamic fusion. Particularly in the adaptive fusion module, this invention cleverly projects the combined temporal and spatial feature sequences onto a unified latent space and learns the dynamic weights of different fusion strategies, realizing on-demand adaptive fusion of multi-source information. This ensures that the final wind power prediction results are reliable under various weather conditions.

[0028] like Figure 2 As shown, Figure 2A flowchart of a wind power prediction method based on spatial-variable dual sensing and adaptive fusion is disclosed. The following will combine... Figure 2 The steps of the above scheme are described in detail: In step S1, the historical power sequence of the wind farm to be predicted, point NWP data for multiple future time steps, and regional two-dimensional NWP field are acquired; wherein, the point NWP data refers to the hourly meteorological forecast data of the wind farm to be predicted unit, and the regional two-dimensional NWP field refers to the two-dimensional meteorological field data intercepted by the center of the wind farm within the target space.

[0029] The data obtained in this step includes the following three categories: 1) Historical power series Historical power series refers to the actual operating power data of the wind farm's generating units to be predicted over the past H time steps. This input is used to reflect the short-term operating status of the unit and the dynamic inertial characteristics of the wind farm.

[0030] Where H is the length of the historical time window. For historical time steps Actual operating power data, It is the set of real numbers.

[0031] 2) Point-based NWP data (Future) Point-based NWP data refers to hourly weather forecast data for wind farm turbines to be predicted. This input, rather than data from randomly selected weather points or other stations, includes various meteorological variables directly related to turbine output, such as wind speed, wind direction, temperature, humidity, and air pressure. This input is used to characterize local meteorological changes at the wind farm's location over the next few hours, serving as a crucial basis for predicting future power trends.

[0032] Where F is the length of the future time window. For moment-by-moment weather forecast data at future time step t, superscript Indicates each moment. This represents the number of meteorological variables at each location.

[0033] 3) Regional two-dimensional NWP field (Future) Regional two-dimensional NWP field refers to two-dimensional meteorological field data intercepted by the center of the wind farm within the target spatial range (such as a 40×40 grid). This input is used to describe the neighborhood wind field structure and large-scale meteorological system that affect the operation of wind farms. This input provides the model with spatial awareness, enabling it to capture the local wind field evolution and spatial propagation characteristics under complex terrain.

[0034] in, For two-dimensional meteorological field data at future time step t, superscript Representing two-dimensional space, H represents the number of NWP variables in the region, and H×W represents the grid size.

[0035] In step S2, the historical power sequence and the point NWP data are used as inputs to the time encoder, encoded separately, and then fused using cross-attention to generate a fused time-series feature sequence.

[0036] It should be noted that the embodiments of the present invention do not limit the specific selection of the time encoder. For example, any of the existing models such as the Discrete Linear (DLinear) model, the Temporal Convolutional Network (TCN), and the Information Transformer (Informer) model can be used. Here, only the DLinear model is used as an example for introduction: This step uses the DLinear model to perform decompositional linear modeling of historical power series and point NWP data. The specific steps are as follows: S21, the encoded historical power sequence, is represented as: in, The historical time-series features output by DLinear have a dimension of , This is the hidden space dimension.

[0037] S22, coded dotted NWP data, is represented as: in, The meteorological time-series features output by DLinear, with dimensions... .

[0038] S23, perform cross-attention fusion, represented as: in, The fused temporal feature sequence; For For Query, For cross-attention of the Key / Value ratio, Query represents the query vector used to initiate the matching, Key represents the key vector used to calculate relevance, and Value represents the information vector that is finally weighted and aggregated.

[0039] In step S3, the two-dimensional NWP field of the region is used as the input of the spatial encoder. After feature extraction, it is weighted by the variable perception mechanism and the spatial perception mechanism in sequence to generate a spatial feature sequence. The variable perception mechanism is used to dynamically adjust the importance of different meteorological variables to the prediction, and the spatial perception mechanism is used to capture key spatial areas in the regional wind field.

[0040] In an optional implementation, the step of using the two-dimensional NWP field of the region as input to the spatial encoder, performing feature extraction, and then sequentially weighting it through a variable-aware mechanism and a spatial-aware mechanism to generate spatial features includes: The original spatial feature map of the two-dimensional NWP field in the region is extracted using a feature extractor; Based on the original spatial feature map, global mean pooling and max pooling are used as variable features, and variable-aware weights are calculated. Based on the original spatial feature map and the variable-perceived weights, a first spatial feature map after variable weighting is generated. Based on the first spatial feature map, global mean pooling and max pooling are used as variable features, and spatial awareness weights are calculated. Based on the first spatial feature map and the spatial perception weights, a spatially weighted second spatial feature map is generated. Based on the second spatial feature map, the spatial feature sequence is generated through dimensionality reduction mapping.

[0041] It should be noted that the embodiments of the present invention do not limit the specific selection of the spatial encoder, especially the feature extractor. For example, the feature extractor can be any of the existing models such as Convolutional Neural Network (CNN), U-shaped Network (UNet), Vision Transformer (ViT), and 3D Convolutional Neural Network (3D-CNN). Figure 2 As shown, this example uses a CNN model as the feature extractor and employs a Convolutional Block Attention Module (CBAM) for both the variable awareness and spatial awareness mechanisms. This step extracts spatial structure information from the regional two-dimensional NWP field data, emphasizing key variables and key regions. It includes convolution extraction, variable perception, spatial region perception, and dimensionality reduction projection steps, as detailed below: S31, extract spatial convolution features, represented as: in, Original spatial feature map; two-dimensional meteorological field data It is a three-dimensional tensor with shape as ×H×W, where H represents the number of channels for meteorological variables (e.g., wind speed u, wind speed v, temperature, etc.), and H and W represent the height and width of the spatial grid (e.g., a 40x40 grid), respectively. Convolutional Neural Networks (CNNs) are used to automatically extract local and global spatial patterns (such as wind field structure, pressure gradient, etc.) from two-dimensional gridded meteorological data.

[0042] S32, Variable-aware weighted mechanism.

[0043] The variable-aware mechanism is used to dynamically adjust the importance of different meteorological variables in forecasting. The importance of variables in the two-dimensional NWP field varies dynamically with weather conditions. Here, a variable weight learning mechanism is introduced, specifically using global mean pooling and max pooling as variable features, and calculating the variable-aware weights, expressed as: This weight applies to the original spatial features: in, The variable-aware weight vector is a vector of length . (i.e., original spatial feature map) A vector containing the number of channels (corresponding to different meteorological variables), with each element... [c] indicates the relative importance of the c-th meteorological variable to the final forecast at the current time step t; The Sigmoid activation function compresses the input to the (0,1) interval, making... Each element can be interpreted as a weight or confidence level; All of these are learnable weight matrices, which together form a small shared fully connected network used to learn how to calculate variable weights from the pooled features. ReLU is the activation function of the Rectified Linear Unit, defined as ReLU(x) = max(0,x), used to introduce nonlinearity; This represents global average pooling, which modulates the original spatial feature map along the spatial dimensions (height H and width W). Averaging is performed on each channel to generate a length of ; This represents global max pooling, which operates along the spatial dimension. Take the maximum value from each channel to generate a string of length . ; express A vector consisting of the eigenvalues ​​of all channels (i.e., all meteorological variables) at spatial location (i,j); This represents the first spatial feature map after variable weighting, and its relationship with the original feature map. Same shape; This indicates element-wise multiplication. Here, the variable-aware weight vector... It will broadcast throughout the entire spatial dimension, and... The feature vectors at each spatial location are multiplied channel by channel.

[0044] S33, weighted spatial (regional) perception mechanism.

[0045] The spatial sensing mechanism is used to capture key spatial regions in the regional wind field. It also employs average pooling and max pooling, and calculates the spatial sensing weights through convolutional layers, as follows: This weight applies to the first spatial feature map after the variables are weighted: in, This is a spatially perceived weight map, which is a two-dimensional matrix with the same shape as the spatial grid H×W, and each element... [i,j] represents the relative importance of spatial location (i,j) to the final prediction at the current time step t; Conv7 is a convolutional layer with a kernel size of 7x7. Using a larger kernel helps to capture a wider range of spatial context information. This represents the weighted first spatial feature map along the channel dimension (i.e., the meteorological variable dimension). Average pooling was performed to generate a two-dimensional graph of size H×W; Indicates the pair along the channel dimension Max pooling was performed to generate a two-dimensional graph of size H×W; [ ; The symbol ] indicates channel splicing, which splices the two H×W two-dimensional graphs mentioned above along the channel dimension; This is a refined second spatial feature map, which is actually the final feature after being weighted by both "variable perception" and "spatial perception" attention mechanisms simultaneously; C here refers to The number of channels, and They have the same meaning.

[0046] S34, dimensionality reduction and sequence construction, is represented as: in, It consists of two linear projection matrices; Let be the spatial feature vector of the future time step t, and correspondingly define the spatial feature vectors of the future F time steps as follows: ; Output the dimension for spatial features.

[0047] In step S4, the temporal feature sequence and the spatial feature sequence are used as inputs to the adaptive fusion module. After being projected onto a unified latent space, an adaptive weighting mechanism based on multiple strategies is used to select the optimal fusion method to generate wind power prediction results for multiple future time steps.

[0048] This step utilizes an adaptive fusion module to dynamically fuse temporal and spatial feature sequences. The specific steps are as follows: S41, projected onto the unified latent space, is represented as: in, There are two independent feedforward networks, each containing a linear layer, a ReLU activation function, and layer normalization, used to map features from different sources to a unified latent semantic space; Features mapped to a unified space.

[0049] S42, the splicing feature, is represented as follows: in, This is the result of splicing.

[0050] S42, the generated fusion weights are represented as follows: in, For a lightweight weight generation network, such as a simple MLP; The weights (soft selection) corresponding to the above splicing operation fusion strategy.

[0051] S43, generate the final fusion feature.

[0052] Understandably, in order to dynamically adjust the weights and fusion logic of different information sources, the adoption of a multi-strategy-based adaptive weighting mechanism to select the optimal fusion method in this embodiment of the invention specifically refers to: Multiple complementary fusion strategies are predefined. Based on the projection results of the temporal feature sequence and the spatial feature sequence, the weights of each fusion strategy are automatically determined by introducing learnable weight coefficients.

[0053] Therefore, the embodiments of the present invention do not need to limit the specific selection of the above fusion strategy. For example, it can be one or any combination of identity mapping (simple splicing), attention fusion, gated recurrent unit (GRU) fusion or multi-layer perceptron (MLP) fusion. The above simple splicing is a feasible example.

[0054] Specifically, in the multi-strategy weighted fusion of the present invention, the historical power characteristics of the projection results are considered. Spatial features Different combination methods, predefine K complementary fusion strategies. These strategies characterize the correlation between the two types of features from different perspectives, such as dynamic association modeling based on attention mechanisms, temporal interaction modeling based on GRU, direct concatenation fusion based on identity mapping, and nonlinear fusion based on MLP. Next, the model introduces learnable weight coefficients. The system automatically evaluates the effectiveness of different strategies at each time step and outputs the final fused features in a weighted manner. Weights This reflects the model's level of confidence in the k-th strategy under current meteorological and power conditions: if attention fusion better characterizes local wind field changes, the model will automatically assign it a higher weight; if historical power trends are more decisive over a certain period, the weight of GRU fusion will be increased accordingly. For example, in situations of rapid wind speed changes or wind direction shifts, the attention fusion strategy can more effectively capture the short-term correlation between point NWP and regional wind field, thus its corresponding weight... The power output will increase significantly; however, in situations where the wind field is relatively stable and power changes have strong temporal inertia, the GRU fusion strategy will contribute more significantly to future power, and its weight will also increase accordingly.

[0055] Through the aforementioned multi-strategy-based adaptive weighting mechanism, the model can automatically select the optimal fusion method based on real-time input, thereby improving the overall robustness and generalization ability of the prediction. The final fused features are represented as follows: in, This represents the wind power forecast result at a future time step t; This is the k-th predefined fusion strategy.

[0056] S44, predicted wind power, is represented as: in, This represents the wind power prediction results for the next F time steps; MLP stands for Multilayer Perceptron.

[0057] This concludes the complete description of the application phase of the wind power prediction method provided in this invention. In addition, this invention also includes improvements to the training phase of the method, as detailed below: On the one hand, in the training phase of the wind power prediction method, the embodiments of the present invention also construct an overall optimization objective including overall point error loss, trend consistency loss and peak segment weighted loss.

[0058] Specifically, the three losses are described as follows: 1) Overall point error.

[0059] Mean Squared Error (MSE), as a fundamental loss term, is primarily used to constrain the point-by-point numerical deviation of the model across all future prediction time steps. This loss amplifies the error between the predicted and actual values ​​by squaring them, making the model more sensitive to large deviations during training. This ensures that the overall prediction results do not undergo systematic shifts and maintains high global numerical accuracy. MSE loss is a core foundation for achieving high-precision wind power prediction, helping the model converge quickly and maintain stable and reliable prediction performance under most operating conditions.

[0060] 2) Trend consistency loss.

[0061] Trend consistency loss aims to enhance the model's ability to capture dynamic changes in wind power time series. Relying solely on numerical errors may lead to the model failing to accurately reflect the actual power trend when numerical values ​​are close, such as a rise in actual power while the predicted result falls, or excessive smoothing of fluctuations. To avoid such trend bias, this invention introduces a constraint on the first-order difference between adjacent time steps, ensuring that the rise, fall, or steady change of the predicted sequence remains consistent with the actual sequence. This loss effectively improves the model's temporal response capability, making it more reliable in dynamic scenarios such as rapid changes in wind power (e.g., gusts, wind shear).

[0062] 3) Peak segment weighted loss (peak segment error weighting is increased).

[0063] Peak-weighted loss primarily targets high-power periods in actual wind power operation. Although these periods constitute a small percentage of the total power output, they are crucial for grid dispatch and system security. Traditional MSE (Mean Sequence Estimation) is easily diluted by numerous stable intervals, leading to insufficient learning of peak value variations. To address this, this invention constructs a peak mask based on the statistical characteristics of real-time power, automatically increasing the loss weight for time steps where the actual power exceeds the mean plus one standard deviation. This mechanism ensures the model focuses on predicting high-power intervals, resulting in higher sensitivity and accuracy under conditions of high winds, high loads, or extreme weather, significantly improving the problems of peak value underestimation or response lag.

[0064] Specifically, the optimization objective is expressed as: The three terms on the right side of the equal sign correspond to the overall point error loss, trend consistency loss, and peak segment weighted loss, respectively.

[0065] This represents the total loss function, used to optimize the final target value of the entire model; B represents the batch size, which is the number of samples processed in one forward / backward propagation. F represents the future prediction length; B represents the index of the b-th sample in the batch; t represents the index of the t-th future time step in the sample sequence. This represents the wind power value predicted by the model for the b-th sample at time t. This represents the wind power value of the b-th sample at time t in the actual observation; Let represent the squared prediction error of the b-th sample at time t; The first difference represents the model's predicted value; The first-order difference represents the true power; α represents the weighting coefficient of trend loss (e.g., 0.1); β represents the weighting coefficient of peak point loss (e.g., 0.1). This represents the peak mask, used to increase error weighting during high-power periods; This indicates that if the actual power exceeds the mean plus one standard deviation, then that moment is considered the peak point and the mask is 1; otherwise, it is 0. This represents the mean of the true power sequence of the b-th sample in the batch; This represents the standard deviation of the true power sequence of the b-th sample in the batch; This represents the average over all samples and all time steps; Indicates trend loss at T2 Calculate the average at one difference time point.

[0066] On the other hand, the spatial encoder introduced in this embodiment of the invention has the structural advantage of being pre-trained independently before participating in joint prediction, so as to effectively improve the stability of regional spatial features.

[0067] Specifically, the spatial encoder employs a training approach of "independent pre-training" plus "joint stage freezing." First, the spatial encoder is pre-trained independently to fully learn the spatial structural features (such as wind field distribution and local gradient) in the (future) two-dimensional NWP field of the region. Then, based on the above optimization objectives, it enters the multi-source joint prediction stage, during which the parameters of the spatial encoder remain frozen and participate in the fusion only as a stable spatial feature extractor.

[0068] Understandably, this design avoids interference from high-dimensional noise during end-to-end training, thereby ensuring the stability of spatial feature representation and improving prediction robustness under complex wind field conditions.

[0069] Example 2: This invention provides a wind power prediction system based on spatial-variable dual sensing and adaptive fusion, comprising: The data acquisition unit is used to acquire the historical power sequence of the wind farm's turbines to be predicted, point-based NWP data for multiple future time steps, and regional two-dimensional NWP field; wherein, the point-based NWP data refers to the hourly meteorological forecast data at the wind farm's turbines to be predicted, and the regional two-dimensional NWP field refers to the two-dimensional meteorological field data intercepted by the wind farm center within the target spatial range. The temporal coding unit is used to take the historical power sequence and the point NWP data as input to the time encoder, encode them separately, and then use cross-attention fusion to generate a fused temporal feature sequence. The spatial coding unit is used to take the two-dimensional NWP field of the region as the input of the spatial encoder, perform feature extraction, and then perform weighting through the variable perception mechanism and the spatial perception mechanism in sequence to generate a spatial feature sequence; wherein, the variable perception mechanism is used to dynamically adjust the importance of different meteorological variables to the prediction, and the spatial perception mechanism is used to capture key spatial areas in the regional wind field. An adaptive fusion unit is used to take the temporal feature sequence and the spatial feature sequence as inputs to the adaptive fusion module, project them onto a unified latent space, and then use an adaptive weighting mechanism based on multiple strategies to select the optimal fusion method to generate wind power prediction results for multiple future time steps.

[0070] Example 3: This invention provides a storage medium storing a computer program, wherein the computer program causes a computer to execute the wind power prediction method based on spatial-variable dual perception and adaptive fusion as described in Embodiment 1.

[0071] Example 4: This invention provides an electronic device, comprising: One or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the programs including methods for performing the wind power prediction method based on spatial-variable dual sensing and adaptive fusion as described in Example 1.

[0072] It is understood that the wind power prediction system, storage medium and electronic device based on spatial-variable dual sensing and adaptive fusion provided in the embodiments of the present invention correspond to the wind power prediction method based on spatial-variable dual sensing and adaptive fusion provided in the embodiments of the present invention. The explanation, examples and beneficial effects of the relevant contents can be referred to the corresponding parts of the method, and will not be repeated here.

[0073] In summary, compared with existing technologies, it has the following beneficial effects: 1. This invention, based on a three-module collaborative structure including a time encoder, a spatial encoder, and an adaptive fusion module, achieves high-precision wind power prediction through multi-level feature extraction and dynamic fusion. Particularly in the adaptive fusion module, this invention cleverly projects the combined temporal and spatial feature sequences onto a unified latent space and learns the dynamic weights of different fusion strategies, realizing on-demand adaptive fusion of multi-source information. This ensures that the final wind power prediction results are reliable under various weather conditions.

[0074] 2. The variable perception mechanism proposed in this embodiment of the invention can dynamically adjust the importance of each meteorological variable to the forecast, and the proposed spatial perception mechanism can identify key spatial areas in the future regional wind field, realize the automatic selection of key variables and key areas, and improve the interpretability of the model.

[0075] 3. In this embodiment of the invention, the spatial encoder adopts a training method of "independent pre-training" plus "joint stage freezing", which can avoid high-dimensional noise interference in end-to-end training, thereby ensuring the stability of spatial feature representation, improving prediction robustness under complex wind field conditions, and greatly improving the generalization ability in complex terrain.

[0076] 4. Compared with traditional models, the embodiments of the present invention reduce the reliance on manual feature engineering, have strong generalization ability, and can achieve good prediction results on large-scale and complex data.

[0077] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0078] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to 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 of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A wind power prediction method based on spatial-variable dual sensing and adaptive fusion, characterized in that, include: The system acquires historical power sequences of the wind farm turbines to be predicted, point-based NWP data for multiple future time steps, and regional two-dimensional NWP fields. The point-based NWP data refers to hourly weather forecast data at the wind farm turbines to be predicted, and the regional two-dimensional NWP fields refer to two-dimensional meteorological field data intercepted within the target spatial range at the center of the wind farm. The historical power sequence and the point NWP data are used as inputs to the time encoder, and after being encoded separately, they are fused by cross-attention to generate a fused time-series feature sequence. The two-dimensional NWP field of the region is used as the input of the spatial encoder. After feature extraction, it is weighted by the variable perception mechanism and the spatial perception mechanism in sequence to generate a spatial feature sequence. The variable perception mechanism is used to dynamically adjust the importance of different meteorological variables to the prediction, and the spatial perception mechanism is used to capture key spatial areas in the regional wind field. The temporal feature sequence and the spatial feature sequence are used as inputs to the adaptive fusion module. After being projected onto a unified latent space, an adaptive weighting mechanism based on multiple strategies is used to select the optimal fusion method to generate wind power prediction results for multiple future time steps.

2. The wind power prediction method as described in claim 1, characterized in that, The step of using the two-dimensional NWP field of the region as input to the spatial encoder, performing feature extraction, and then sequentially weighting it through a variable-aware mechanism and a spatial-aware mechanism to generate spatial features includes: The original spatial feature map of the two-dimensional NWP field in the region is extracted using a feature extractor; Based on the original spatial feature map, global mean pooling and max pooling are used as variable features, and variable-aware weights are calculated. Based on the original spatial feature map and the variable-perceived weights, a first spatial feature map after variable weighting is generated. Based on the first spatial feature map, global mean pooling and max pooling are used as variable features, and spatial awareness weights are calculated. Based on the first spatial feature map and the spatial perception weights, a spatially weighted second spatial feature map is generated. Based on the second spatial feature map, the spatial feature sequence is generated through dimensionality reduction mapping.

3. The wind power prediction method as described in claim 2, characterized in that, The feature extractor is any one of CNN, UNet, ViT, or 3D-CNN.

4. The wind power prediction method as described in claim 1, characterized in that, The adoption of a multi-strategy-based adaptive weighting mechanism to select the optimal fusion method refers to: Multiple complementary fusion strategies are predefined. Based on the projection results of the temporal feature sequence and the spatial feature sequence, the weights of each fusion strategy are automatically determined by introducing learnable weight coefficients.

5. The wind power prediction method as described in claim 1, characterized in that, The fusion strategy is one or a combination of several of the following: identity mapping, attention fusion, GRU fusion, or MLP fusion.

6. The wind power prediction method as described in claim 1, characterized in that, During the training phase of the wind power prediction method, an overall optimization objective is constructed, which includes overall point error loss, trend consistency loss, and peak segment weighted loss.

7. The wind power prediction method as described in claim 1, characterized in that, During the training phase of the wind power prediction method, the spatial encoder is pre-trained separately and its network parameters are frozen before joint training is performed based on the overall optimization objective.

8. A wind power prediction system based on spatial-variable dual sensing and adaptive fusion, characterized in that, include: The data acquisition unit is used to acquire the historical power sequence of the wind farm's turbines to be predicted, point-based NWP data for multiple future time steps, and regional two-dimensional NWP field; wherein, the point-based NWP data refers to the hourly meteorological forecast data at the wind farm's turbines to be predicted, and the regional two-dimensional NWP field refers to the two-dimensional meteorological field data intercepted by the wind farm center within the target spatial range. The temporal coding unit is used to take the historical power sequence and the point NWP data as input to the time encoder, encode them separately, and then use cross-attention fusion to generate a fused temporal feature sequence. The spatial coding unit is used to take the two-dimensional NWP field of the region as the input of the spatial encoder, perform feature extraction, and then perform weighting through the variable perception mechanism and the spatial perception mechanism in sequence to generate a spatial feature sequence; wherein, the variable perception mechanism is used to dynamically adjust the importance of different meteorological variables to the prediction, and the spatial perception mechanism is used to capture key spatial areas in the regional wind field. An adaptive fusion unit is used to take the temporal feature sequence and the spatial feature sequence as inputs to the adaptive fusion module, project them onto a unified latent space, and then use an adaptive weighting mechanism based on multiple strategies to select the optimal fusion method to generate wind power prediction results for multiple future time steps.

9. A storage medium, characterized in that, It stores a computer program, wherein the computer program causes the computer to execute the wind power prediction method based on spatial-variable dual perception and adaptive fusion as described in any one of claims 1 to 7.

10. An electronic device, characterized in that, include: One or more processors; Memory; And one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the programs including methods for performing the wind power prediction method based on spatial-variable dual sensing and adaptive fusion as described in any one of claims 1 to 7.

Images