Wind power prediction method and device and electronic equipment

Abstract

The embodiment of the invention provides a wind power prediction method and device and electronic equipment. The method comprises the following steps: acquiring meteorological parameters of a wind power plant in different time and space, and constructing a time and space tensor; performing feature decoupling on the space-time tensor, and separating to obtain a space feature map and a time sequence flow; performing feature extraction on the spatial feature map through a spatial channel to obtain a global meteorological field feature vector; performing feature extraction on the time sequence flow through a time channel to obtain long and short time sequence features; dynamically fusing the global meteorological field feature vector and the long and short term time sequence features through a meta-learning weight distributor to obtain fused features; evaluating each prediction model based on the fusion features, and determining an optimal prediction model; and performing power prediction on the wind power plant according to the optimal prediction model. According to the method, the precision and robustness of wind power prediction can be improved.

Description

Technical Field

[0001] This application relates to the field of wind power generation technology, and in particular to a wind power prediction method, apparatus and electronic equipment. Background Technology

[0002] As the global energy structure transitions towards cleaner and lower-carbon energy, wind power, as a crucial component of renewable energy, is experiencing continuous expansion in installed capacity and grid connection. However, the inherent intermittency and volatility of wind energy pose significant challenges to the safe and stable operation of the power system. Statistics show that a 1% reduction in wind power forecasting error can decrease grid reserve capacity requirements by approximately 2%-3%.

[0003] Currently, existing wind power forecasting methods mainly rely on collecting historical data and then using that data to predict the future power generation of wind farms.

[0004] However, the accuracy of wind power forecasting is currently low. Summary of the Invention

[0005] This application provides wind power prediction methods, apparatus, and electronic devices to improve the accuracy of wind power prediction.

[0006] In a first aspect, embodiments of this application provide a wind power prediction method, including:

[0007] Meteorological parameters of wind farms at different times and spaces are obtained, and spatiotemporal tensors are constructed.

[0008] The spatiotemporal tensor is decoupled by features to separate the spatial feature map and the time series stream;

[0009] The spatial feature map is used to extract features through a spatial channel to obtain a global meteorological field feature vector.

[0010] Feature extraction is performed on the time series stream through the time channel to obtain long- and short-term time series features;

[0011] The global meteorological field feature vector and the long- and short-term time-series features are dynamically fused using a meta-learning weight allocator to obtain the fused features;

[0012] The various prediction models are evaluated based on the fusion features to determine the optimal prediction model;

[0013] Based on the optimal prediction model, the power of the wind farm is predicted.

[0014] Secondly, embodiments of this application provide a wind power prediction device, comprising:

[0015] A module is built to obtain meteorological parameters of the wind farm at different times and spaces, and to construct a spatiotemporal tensor;

[0016] The decoupling module is used to decouple the spatiotemporal tensor by separating the spatial feature map and the time series stream.

[0017] The spatial extraction module is used to extract features from the spatial feature map through a spatial channel to obtain a global meteorological field feature vector;

[0018] The time extraction module is used to extract features from the time series stream through the time channel to obtain long and short time series features;

[0019] The feature fusion module is used to dynamically fuse the global meteorological field feature vector and the long- and short-term time-series features through a meta-learning weight allocator to obtain fused features;

[0020] The model determination module is used to evaluate each prediction model based on the fusion features and determine the optimal prediction model;

[0021] The power prediction module is used to predict the power of the wind farm based on the optimal prediction model.

[0022] Thirdly, embodiments of this application provide an electronic device, including: a memory and a processor;

[0023] The memory stores computer-executed instructions;

[0024] The processor executes computer execution instructions stored in the memory, causing the processor to perform the first aspect and / or various possible implementations of the first aspect as described above.

[0025] Fourthly, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the first aspect and / or various possible implementations of the first aspect.

[0026] Fifthly, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the first aspect and / or various possible implementations of the first aspect.

[0027] The wind power prediction method, apparatus, and electronic equipment provided in this application extract global meteorological field features through a spatial channel, which can effectively capture the multi-scale nonlinear features of the meteorological field and reduce the spatial feature extraction error under extreme weather conditions. By obtaining long- and short-term time series features through a time channel, the root mean square error of traditional time series models can be significantly reduced. Finally, the dynamic optimization and fusion of meteorological and time series features is achieved through a meta-learning weight allocator, which can improve the prediction stability in sudden scenarios such as strong convection and achieve the goal of improving the accuracy of wind power prediction. Attached Figure Description

[0028] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0029] Figure 1 A flowchart illustrating the wind power prediction method provided in this application embodiment;

[0030] Figure 2 This is a flowchart illustrating the overall process structure for wind power prediction provided in an embodiment of this application.

[0031] Figure 3 Spatial feature maps and time series stream separation flowcharts provided for embodiments of this application;

[0032] Figure 4 A flowchart for spatial feature extraction using a spatial channel is provided for embodiments of this application;

[0033] Figure 5 A flowchart for extracting time features using a time channel is provided in an embodiment of this application;

[0034] Figure 6 A schematic diagram of the dynamic fusion dual-channel output provided in the embodiments of this application;

[0035] Figure 7 This is a schematic diagram of the optimal model selection process provided in an embodiment of this application;

[0036] Figure 8 This is a schematic diagram of the structure of the wind power prediction device provided in the embodiments of this application;

[0037] Figure 9 A schematic diagram of the structure of the electronic device provided in this application.

[0038] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0039] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0040] Current wind power prediction technologies are mainly divided into three categories: physical methods, statistical methods, and model learning methods. However, the following technical bottlenecks still exist: (1) Traditional methods (such as Auto-Regressive Moving Average Model (ARMA model), Support Vector Machine (SVM) etc.) are difficult to effectively handle the multi-scale spatial features in numerical weather prediction (NWP), such as the nonlinear distribution of wind speed fields under complex terrain and the meteorological mutation characteristics of extreme weather events (such as typhoons and strong convection), resulting in the prediction error of existing models generally exceeding 20% ​​in the spatial dimension; (2) Although time series models based on recurrent neural networks (RNN) or long short-term memory networks (LSTM) can capture short-term dependencies, they are not capable of modeling the long-term dynamic characteristics of non-stationary power sequences (such as power ramp events caused by sudden changes in wind speed). Studies have shown that within the 4-6 hour prediction interval, the root mean square error (RMS) of traditional time series models is significantly higher than that of other models. Error (RMSE) will increase by 30%-40%; (3) Existing combination methods (such as physical-statistical hybrid models) mostly use static weight allocation, which cannot adaptively adjust the contribution of meteorological data and power sequence. For example, in low wind speed and high turbulence scenarios, when the reliability of meteorological data decreases, the model still mechanically integrates low-quality features, which leads to the amplification of prediction bias.

[0041] To address the aforementioned issues, a forecasting method capable of coordinating the handling of meteorological spatial heterogeneity and power temporal nonstationarity is urgently needed. Based on this, this application presents an NSTransformers wind power forecasting method based on spatiotemporal feature decoupling. Through dynamic feature fusion and multi-scale modeling, it can achieve wind power forecasting in complex scenarios, thereby improving forecast accuracy.

[0042] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0043] Figure 1 This is a flowchart illustrating the wind power prediction method provided in the embodiments of this application, as shown below. Figure 1 As shown, the method includes the following steps:

[0044] S110. Obtain meteorological parameters of the wind farm at different times and spaces, and construct a spatiotemporal tensor.

[0045] S120. Decouple the spatiotemporal tensor by features to separate the spatial feature map and the time series stream.

[0046] S130. Extract features from the spatial feature map through the spatial channel to obtain the global meteorological field feature vector.

[0047] S140. Extract features from the time series stream through the time channel to obtain long and short time series features.

[0048] S150. The global meteorological field feature vector and long- and short-term time series features are dynamically fused through a meta-learning weight allocator to obtain the fused features.

[0049] S160. Evaluate each prediction model based on the fusion features and determine the optimal prediction model.

[0050] S170. Based on the optimal prediction model, predict the power of the wind farm.

[0051] The wind power prediction method provided in this application extracts global meteorological field features through a spatial channel, which can effectively capture the multi-scale nonlinear features of the meteorological field and reduce the spatial feature extraction error under extreme weather conditions. By obtaining long and short-term time series features through a time channel, the root mean square error of traditional time series models can be significantly reduced. Finally, the dynamic optimization and fusion of meteorological and time series features is achieved through a meta-learning weight allocator, which can improve the prediction stability in sudden scenarios such as strong convection and achieve the goal of improving the accuracy of wind power prediction.

[0052] For the above S110, multi-dimensional NWP data can be collected from the wind farm to construct the spatiotemporal tensor.

[0053] NWP integrates observational data from multiple sources, including ground stations, upper-air sounding, radar, and satellites, and generates a three-dimensional gridded initial field through data assimilation technology. Specifically, it divides the Earth's surface into a fine three-dimensional grid, with each grid point independently calculating physical processes (such as radiative transfer and turbulent mixing), thus achieving refined spatial simulation.

[0054] In this embodiment, meteorological parameters may include wind speed, temperature, humidity, air pressure, turbulence intensity, etc. Specifically, when collecting meteorological data in the spatial direction, longitude and latitude data with a resolution of 0.05 degrees, as well as three-layer grid data with heights of 10 meters, 50 meters, and 100 meters, can be collected. When collecting meteorological data in the temporal direction, time dimension data with a time resolution of 15 minutes and containing at least 72 hours can be collected.

[0055] In some embodiments, first meteorological data at different latitude and longitude locations in vertical layers at different altitudes, and second meteorological data within each preset time step in a preset time period are acquired.

[0056] The first meteorological data can be the longitude and latitude data with a collection resolution of 0.05 degrees mentioned above, as well as the three-layer grid data with heights of 10 meters, 50 meters and 100 meters. The second meteorological data can be the time dimension data with a time resolution of 15 minutes and containing at least 72 hours mentioned above.

[0057] After collecting the above data, a five-dimensional spacetime tensor can be constructed:

[0058] X∈R T×H×W×D×C

[0059] In the above formula, X is the five-dimensional spacetime tensor, T represents the preset time step, HxW represents the spatial grid, D is the vertical layer, and C is the meteorological channel.

[0060] In addition, in some embodiments, after constructing the above-mentioned five-dimensional spatiotemporal tensor, spatial data augmentation of the multidimensional spatiotemporal tensor can be performed using elastic deformation; and temporal data augmentation of the multidimensional spatiotemporal tensor can be performed using a non-uniform sampling strategy.

[0061] In this implementation, since real-world data is often affected by natural disturbances (such as wind deflection caused by terrain undulations), elastic deformation enables subsequent models to robustly cope with local distortions. Elastic deformation refers to the nonlinear transformation of the data's geometric structure by simulating physical "stretching," "compression," or "bending" operations, thereby expanding the diversity of training samples and helping the model learn invariant features under different deformation states.

[0062] In this embodiment, the non-uniform sampling strategy employs dynamic step-size or probability distribution resampling to capture the differences in importance among different frequency components in the time series. Data enhancement using the non-uniform sampling strategy can cover a wider spectral range.

[0063] For example, time enhancement employs a non-uniform sampling strategy, using twice the sampling frequency for periods of drastic wind speed changes (acceleration > 3 m / s²), and half the sampling frequency for periods of stability.

[0064] Regarding S120 above, the spatial dimension is mainly used to describe the distribution pattern in physical location, while the temporal dimension reflects the evolution of the state over time. Specifically, a feature decoupling module can be used to separate the spatial feature map and the time series stream. Specifically, the feature decoupling module can employ a dual-path parallel encoder structure, where one path captures spatial context information based on 3D convolution, and the other path uses a gated recurrent unit (GRU) to model temporal dependencies. After independent processing, the two feature paths are dynamically fused through a cross-attention mechanism.

[0065] For example, the spatial feature map may include a wind speed distribution heatmap, and the time series stream may include historical time curves of wind speed.

[0066] For S130 and S140 above, dual-channel NSTransformers (Non-stationary Transformers) can be used for processing. Specifically, the spatial channel uses ConvNeXt-Transformer to extract global meteorological field features, while the temporal channel uses NSTransformers to model dynamic time series and extract long-term and short-term time series features.

[0067] Specifically, the two feature streams (spatial feature map and time series stream) are fed into independently designed sub-networks. The spatial feature processing channel adopts an improved ConvNeXt-Transformer architecture, which scans meteorological data layer by layer through a 7×7×3 3D convolutional window to extract local wind field vortex and temperature and pressure correlation features. It uses a cross-parameter attention mechanism to analyze the interaction between different meteorological elements to comprehensively grasp the characteristics of the weather system. A spatial pyramid structure is constructed to adaptively fuse convolutional features at different scales and output a global meteorological field feature vector with spatial location encoding.

[0068] The time feature processing channel adopts a Non-stationary Transformers architecture, which dynamically segments the time series and automatically adjusts the analysis range according to the degree of data fluctuation. A wide window of 12 hours is used to capture long-term trends during stable periods, while a short window of 3 hours is used to capture abrupt changes during periods of violent fluctuations. The time series data is decomposed into two main components: weather system evolution (low frequency) and turbulent fluctuations (high frequency) for separate processing. The feature weights of key turning points such as sudden increases and decreases in wind speed are enhanced. The processing adopts a phase-aware attention mechanism to ensure that features at different time scales can be naturally connected. The final output is a time series feature that includes both short-term fluctuation details and reflects long-term evolution trends.

[0069] For the aforementioned S150, the dual-channel outputs can be dynamically fused using a meta-learning weight allocator. The meta-learning weight allocator is a dynamic parameter adjustment mechanism based on a meta-learning paradigm. Its core objective is to quickly adapt to changes in new tasks or data distribution using a small number of samples, automatically optimizing the weight configuration of different modules / layers. Specifically, a dedicated meta-feature extraction module can be constructed, employing a two-layer fully connected network to extract high-order statistical features from the original features. The spatial channel extracts features such as the meteorological field uniformity index and gradient intensity, while the temporal channel extracts dynamic features such as Ljung-Box test statistics and mutation point density. The extracted 12-dimensional meta-feature vector is then concatenated with the current environmental context (which may include seasonal flags and weather type codes) to form an 18-dimensional complete meta-feature representation. Next, a weight prediction network based on gated recurrent units is designed. The input layer receives the meta-feature vector, which undergoes temporal modeling using 32-dimensional GRU units. The output layer uses a softmax activation function to generate normalized weights (w_spatial, w_temporal). During network training, a dual loss function is used: the main loss is the negative log-likelihood of the prediction error, and the regularization term is the first-order difference L2 norm of the weight sequence. An error threshold trigger mechanism is set; when the sliding window RMSE of any channel exceeds 1.5 times the historical baseline, the weight recalculation process is initiated. Finally, multi-level feature fusion is implemented.

[0070] For S160 and S170 mentioned above, a composite evaluation system including deterministic and probabilistic indicators can be established to conduct stratified evaluations for different time scales from 15 minutes to 4 hours. At the same time, spatial correlation analysis is introduced to comprehensively cover the spatiotemporal performance characteristics of the prediction model. Meanwhile, an entropy weight method-TOPSIS hybrid decision-making mechanism is adopted. Through adaptive weight allocation and a two-level screening process, combined with stability verification by 7-day rolling tests, the model can be ensured to maintain optimal performance under various meteorological conditions.

[0071] After determining the optimal prediction module, the preferred model can be containerized and deployed, a three-dimensional monitoring system can be established, an automatic alarm mechanism based on dynamic thresholds can be configured, and then continuous performance optimization can be achieved through periodic evaluation reports.

[0072] In some embodiments, the feature decoupling of the spatiotemporal tensor can be achieved through the following steps:

[0073] (1) Normalize the meteorological parameters of each grid point in the spatiotemporal tensor and interpolate the topographic elevation data of the wind farm to each vertical layer to form a three-dimensional topographic feature map.

[0074] (2) Use a lightweight network to predict the three-dimensional position offset of each grid point, and perform non-uniform sampling based on the offset to obtain a spatial feature map;

[0075] (3) Multi-scale convolutional windows are used to extract features from the spatiotemporal tensor after non-uniform sampling to obtain system-level features, mesoscale features and micro-meteorological features, forming a spatial feature map;

[0076] (4) Use dynamic attention mechanism to separate the temporal features of the spatiotemporal tensor to obtain the separated temporal features;

[0077] (5) Wavelet frequency domain decomposition of time features is used to obtain time series streams of different frequency bands.

[0078] For example, Figure 2 This is a flowchart illustrating the overall process structure for wind power prediction provided in the embodiments of this application, as follows: Figure 2 As shown, the wind power forecasting process is divided into five stages:

[0079] Phase 1: Multidimensional NWP data acquisition and spatiotemporal tensor construction;

[0080] Phase Two: Separation of Spatial Feature Map and Time Series Stream;

[0081] Phase 3: Dual-channel NSTransformers processing;

[0082] Phase Four: Dynamic Fusion of Dual-Channel Output;

[0083] Phase 5: Optimal model selection based on multi-scale evaluation system.

[0084] In this implementation, the feature decoupling of the spatiotemporal tensor corresponds to the second stage mentioned above. Specifically, in the process of obtaining the spatial feature map through decoupling, the spatiotemporal tensor data collected in the first stage needs to be independently normalized using meteorological parameters, followed by terrain elevation data interpolation to obtain a three-dimensional terrain feature map. Non-uniform sampling is then performed using three-dimensional position offset prediction, followed by feature extraction using a multi-scale convolutional window to obtain system-level features, mesoscale features, and micro-meteorological features, forming the spatial feature map.

[0085] In this embodiment, reference is continued to be made to the above. Figure 2 The time series stream can introduce a dynamic attention mechanism into the time feature separation process, and then use wavelet frequency domain decomposition of time features to obtain high-frequency components and low-frequency components, thus forming a time series stream.

[0086] Specifically, Figure 3 The spatial feature map and time series stream separation flowchart provided for the embodiments of this application are as follows: Figure 3 As shown, it includes the following steps:

[0087] S2.1: Receives spatiotemporal tensor data and independently normalizes the meteorological parameters of each grid point, while interpolating the terrain elevation data to each vertical layer to form a three-dimensional terrain feature map.

[0088] The normalization operation is used to eliminate dimensional differences.

[0089] S2.2: Predict the 3D position offset of each grid point using a lightweight network, limiting the offset range to no more than 20% of the grid spacing, and then perform non-uniform sampling on the original data based on the offset.

[0090] Non-uniform sampling can enhance adaptability to complex terrain.

[0091] S2.3: Multi-scale convolutional windows are used for feature extraction. Large window convolution (5×5×3) is used to capture weather system-level features, medium window convolution (3×3×3) is used to extract mesoscale features, and point convolution (1×1×1) is used to capture local micro-meteorological features.

[0092] S2.4: Introducing dynamic attention mechanism into the time feature separation process, automatically analyzing the importance of different time points, assigning higher weight to key time periods, and using wavelet transform to decompose time series data into different frequency bands during signal decomposition.

[0093] Among them, the critical period can refer to the period of sudden wind speed change. The high-frequency component is used to capture rapid fluctuations such as turbulence, while the low-frequency component is used to extract the evolution trend of the weather system.

[0094] S2.5: Real-time monitoring of the contribution of spatial and temporal features, automatic calculation of the optimal fusion ratio through neural networks, dynamic adjustment of weight allocation, detection of spatial purity and temporal independence to verify the spatiotemporal decoupling effect.

[0095] Furthermore, in some embodiments, after decoupling the spatial feature map and the time series stream, meteorological field features can be extracted through the following steps:

[0096] The spatial feature maps are rearranged based on vertical layers and meteorological parameters to form a three-dimensional data structure of meteorological parameters;

[0097] Based on a three-dimensional data structure, cross-layer correlation analysis is performed on the multi-layer meteorological parameters of each horizontal grid point to establish a three-dimensional feature representation.

[0098] By using a preset 3D convolution window, the 3D feature representation is scanned layer by layer to extract meteorological parameters and convolution features at different scales.

[0099] By adaptively fusing convolutional features at different scales, a global meteorological field feature vector carrying spatial location encoding is obtained.

[0100] In this embodiment, the purpose of rearranging the data is to integrate the originally scattered two-dimensional planar data into a unified three-dimensional tensor (i.e., a three-dimensional data structure), where each coordinate axis corresponds to longitude, latitude, and altitude, respectively. This three-dimensional data structure can preserve the complete physical spatial relationships.

[0101] The preset 3D convolution window can be a 7×7×3 3D convolution window.

[0102] For example, Figure 4 A flowchart for spatial feature extraction using spatial channels provided in this application embodiment is shown below. Figure 4 As shown, it includes the following steps:

[0103] S3.1: The decoupled spatial feature map is rearranged according to vertical layers and meteorological parameters to form a "meteorological cube" data structure. Cross-layer correlation analysis is performed on the multi-layer meteorological parameters of each horizontal grid point to establish a three-dimensional feature representation.

[0104] S3.2: The spatial feature processing channel adopts an improved ConvNeXt-Transformer architecture. It scans meteorological data layer by layer through a 7×7×3 3D convolutional window, extracts wind field vortex and temperature and pressure correlation features in local areas, and uses a cross-parameter attention mechanism to analyze the mutual influence between different meteorological elements. It constructs a spatial pyramid structure and adaptively fuses convolutional features of different scales to output a global meteorological field feature vector with spatial location encoding.

[0105] Among them, using the cross-parameter attention mechanism to analyze the interaction between different meteorological elements can provide a comprehensive understanding of the characteristics of weather systems.

[0106] In addition, in some embodiments, long-term and short-term time-series features can be extracted through the following steps:

[0107] Based on the degree of data fluctuation in the time series stream, dynamic windows of different sizes are used to capture features of the time series stream;

[0108] A phase-aware attention mechanism is used to weight the temporal features after feature capture to obtain temporal features at different time scales.

[0109] Among them, time series features at different time scales include short-term time series features that characterize short-term fluctuation details and long-term time series features that characterize long-term evolution trends.

[0110] Compared to traditional methods that typically use a fixed-length sliding window to analyze data, this embodiment can employ an adaptive window size strategy. For example, when the data fluctuates drastically, the window range can be reduced to accurately locate the point of change, while during stable periods, the window span can be expanded to cover a longer time frame to capture long-term trends.

[0111] In addition, the attention mechanism of phase perception is mainly used to determine which frequencies are more important. High-frequency components are used to enhance the ability to describe short-term details, while low-frequency components are used to highlight the dominant role of long-term trends.

[0112] For example, Figure 5 A flowchart for extracting time features using a time channel is provided in the embodiments of this application, such as... Figure 5 As shown, it includes the following steps:

[0113] S3.3: The time feature processing channel adopts a non-stationary Transformers architecture to dynamically segment the time series.

[0114] Among them, the analysis range can be automatically adjusted according to the degree of data fluctuation. A 12-hour wide window is used to capture long-term trends during stable periods, and a 3-hour short window is used to capture sudden change features during periods of violent fluctuations. The time series data is decomposed into two main components, weather system evolution (low frequency) and turbulent fluctuations (high frequency), and processed separately to enhance the feature weight of key turning points such as sudden increases and decreases in wind speed.

[0115] In addition, the processing employs a phase-aware attention mechanism to ensure that features at different time scales can be seamlessly connected, and the final output contains both short-term fluctuation details and reflects long-term evolution trends in time-series features.

[0116] S3.4: A multi-level interaction mechanism is used to achieve collaborative work between the two channels. After each processing stage, the spatial channel transmits the meteorological field distribution features it extracts to the temporal channel and receives the evolution trend information fed back by the temporal channel. A "spatial distribution-temporal evolution" correlation model is established through cross-attention. In the final fusion stage, temporal continuity constraints are imposed on the spatial features, and spatial smoothness conditions are added to the temporal features.

[0117] Among them, adding spatial smoothness conditions to the temporal characteristics can ensure that the fused characteristics not only conform to the laws of fluid mechanics but also accurately reflect the actual evolution process of wind power.

[0118] In the above embodiments, the spatial feature channel adopts the improved ConvNeXt-Transformer, which can effectively capture the multi-scale nonlinear features of the meteorological field and reduce the spatial feature extraction error under extreme weather conditions; the temporal feature channel adopts Non-stationary Transformers to model the dynamic evolution of the power sequence, and the RMSE of the 6-hour prediction interval is significantly lower than that of the traditional LSTM.

[0119] In some embodiments, the dynamic fusion of dual-channel outputs via a meta-learning weight allocator can be achieved through the following steps:

[0120] (1) Perform three-dimensional adaptive max pooling on the global meteorological field feature vector to obtain the pooling result; (2) Use thread interpolation to align the long and short-term time series features to the standard length to obtain the alignment result; (3) Based on the meta-learning weight allocator, obtain the first weight of the pooling result and the second weight of the alignment result; (4) Based on the first weight and the second weight, perform weighted fusion on the pooling result and the alignment result to obtain the fused feature.

[0121] In this embodiment, reference is continued to be made to the above. Figure 2 The dynamic fusion of dual-channel outputs through a meta-learning weight allocator can specifically correspond to the above. Figure 2 The fourth stage of the process can include the following steps: (1) standardizing the spatiotemporal features; (2) performing three-dimensional adaptive pooling for the spatial features; (3) performing linear interpolation alignment for the temporal features; (4) predicting weights through a weight prediction network; and (5) performing dynamic fusion.

[0122] in, Figure 6 This is a schematic diagram of the dynamic fusion dual-channel output provided in an embodiment of this application, as shown below. Figure 6 As shown, it specifically includes the following steps:

[0123] S4.1: Standardize the spatial feature tensor and temporal feature sequence of the dual-channel output.

[0124] Specifically, channel-independent Z-score normalization can be used to eliminate dimensional differences, preserving the three-dimensional grid structure of spatial features and maintaining the temporal continuity of temporal features, while also adding channel identifier encoding. The spatial feature tensor has dimensions of H×W×D×C, and the temporal feature sequence has dimensions of T×C'.

[0125] S4.2: Construct a dedicated meta-feature extraction module.

[0126] Among them, a two-layer fully connected network is used to extract high-order statistical features from the original features. The spatial channel extracts features such as the meteorological field uniformity index and gradient intensity, while the temporal channel extracts dynamic features such as Ljung-Box test statistics and mutation point density. The extracted 12-dimensional meta-feature vector is concatenated with the current environmental context to form an 18-dimensional complete meta-feature representation.

[0127] S4.3: Design a weight prediction network based on gated recurrent units.

[0128] The input layer receives meta-feature vectors and performs temporal modeling using 32-dimensional GRU units. The output layer uses a softmax activation function to generate normalized weights (w_spatial represents the first weight, and w_temporal represents the second weight). During network training, a dual loss function is used: the main loss is the negative log-likelihood of the prediction error, and the regularization term is the first-order difference L2 norm of the weight sequence. An error threshold trigger mechanism is set, and when the sliding window RMSE of any channel exceeds 1.5 times the historical benchmark, the weight recalculation process is initiated.

[0129] S4.4: Implement multi-level feature fusion, perform three-dimensional adaptive max pooling on the spatial feature tensor, align the temporal feature sequence to a uniform length using linear interpolation, and perform weighted concatenation operations along the feature dimension. Details are as follows:

[0130] F_fused=w_spatial·Pool(F_spatial)⊕w_temporal·Interp(F_temporal)

[0131] In the above formula, F_fused represents the fused feature, w_spatial represents the first weight, Pool represents the pooling operation, F_spatial represents the global meteorological field feature vector, ⊕ represents channel concatenation, w_temporal represents the second weight, Interp represents the thread interpolation alignment operation, and F_temporal represents the long-term and short-term time series features.

[0132] The original feature information is preserved through residual connections after fusion, as shown below:

[0133] F_out=LayerNorm(F_fused+F_spatial[:,0]+F_temporal[0])

[0134] In the above formula, F_out represents the output after residual connection processing, LayerNorm represents layer normalization processing, F_fused represents fused features, F_spatial[:,0] represents the data of the first channel or slice in the global meteorological field feature vector, that is, the information of a specific spatial location, and F_temporal[0] represents the data of the first time step in the long-term and short-term time series features.

[0135] In this step, an online learning strategy can also be used to update the weight prediction network parameters based on backpropagation of real-time prediction errors.

[0136] S4.5: Establish a fusion effect evaluation system, and calculate the spatial-temporal feature contribution ratio (Signal-to-Response Ratio, STR) and the feature concordance index (FCI). The specific formulas are as follows:

[0137]

[0138] FCI=cos_sim(F_spatial,F_temporal)

[0139] In the above formula, This represents the partial derivative with respect to space. F_spatial represents the spatial feature, F_temporal represents the temporal feature, and cos_sim is the cosine similarity function.

[0140] The meta-features, weight decisions, and evaluation indicators from each fusion are stored in a circular buffer knowledge base (with a capacity of 10,000 entries). Knowledge distillation is performed periodically (e.g., every 24 hours) to compress and store key decision patterns. A dual feedback loop is set up: short-term feedback (e.g., every 15 minutes) adjusts the weight prediction bias, and long-term feedback (e.g., weekly) optimizes the meta-feature extraction strategy.

[0141] Furthermore, in some embodiments, Figure 7 This is a schematic diagram of the optimal model selection process provided in the embodiments of this application, such as... Figure 7 As shown, it includes the following steps:

[0142] S5.1: Establish a comprehensive evaluation system, conduct tiered evaluations of fused features at different time scales ranging from 15 minutes to 4 hours, and introduce spatial correlation analysis to obtain evaluation results. This comprehensively covers the spatiotemporal performance characteristics of the prediction model.

[0143] The composite evaluation system includes deterministic and probabilistic indicators. The deterministic indicators include Normalized Mean Absolute Error (NMAE) and Normalized Root Mean Square Error (NRMSE). The probabilistic indicators include Continuous Ranked Probability Score (CRPS) and quantile loss.

[0144] S5.2: An entropy weight method-TOPSIS hybrid decision-making mechanism is adopted, which uses adaptive weight allocation and a two-level screening process, combined with stability verification through a 7-day rolling test.

[0145] Among them, stability verification can ensure that the selected optimal prediction model maintains its best performance under various meteorological conditions.

[0146] The Entropy Weight Method-TOPSIS Hybrid Decision-Making Mechanism is a comprehensive evaluation method combining the entropy weight method and the TOPSIS method. The entropy weight method measures the dispersion of data based on the concept of "entropy" in information theory. The logic of the TOPSIS method is as follows: by calculating the distance between each solution and the "positive ideal solution" (a virtual solution composed of the optimal values ​​of all attributes) and the "negative ideal solution" (a virtual solution composed of the worst values ​​of all attributes), the solution that is closest to the positive ideal solution and furthest from the negative ideal solution is selected as the optimal solution.

[0147] S5.3: Containerize and deploy the preferred model and establish a 3D monitoring system, configure an automatic alarm mechanism based on dynamic thresholds, and achieve continuous performance optimization through periodic evaluation reports.

[0148] The three-dimensional monitoring system includes error timing, distribution characteristics, and extreme responses. The automatic alarm mechanism can be triggered when NMAE > 2.5% or CRPS deteriorates by 20%, in which case retraining is triggered.

[0149] Through the above embodiments, this solution improves the accuracy and robustness of wind power prediction by employing a dual-channel NSTransformers architecture and a dynamic weight fusion mechanism. Specifically, the spatial feature channel uses an improved ConvNeXt-Transformer, which can effectively capture the multi-scale nonlinear characteristics of the meteorological field and reduce spatial feature extraction errors under extreme weather conditions. The temporal feature channel uses Non-stationary Transformers to model the dynamic evolution of the power sequence, resulting in a significant decrease in the RMSE of the 6-hour prediction interval compared to the traditional LSTM. A meta-learning weight allocator is used to achieve dynamic optimization and fusion of meteorological and temporal features, improving prediction stability in abrupt scenarios such as strong convection.

[0150] Figure 8 This is a schematic diagram of the structure of the wind power prediction device provided in the embodiments of this application, as shown below. Figure 8 As shown, the wind power prediction device 800 provided in this embodiment includes:

[0151] Module 810 is used to obtain meteorological parameters of the wind farm in different times and spaces, and to construct the spatiotemporal tensor;

[0152] The decoupling module 820 is used to decouple the spatiotemporal tensor by separating the spatial feature map and the time series stream.

[0153] The spatial extraction module 830 is used to extract features from the spatial feature map through the spatial channel to obtain the global meteorological field feature vector;

[0154] The time extraction module 840 is used to extract features from the time series stream through the time channel to obtain long and short time series features;

[0155] The feature fusion module 850 is used to dynamically fuse global meteorological field feature vectors and long- and short-term time-series features through a meta-learning weight allocator to obtain fused features;

[0156] The model determination module 860 is used to evaluate each prediction model based on the fusion features and determine the optimal prediction model;

[0157] The power prediction module 870 is used to predict the power of a wind farm based on the optimal prediction model.

[0158] In one possible implementation, the building module is specifically used for:

[0159] Acquire the first meteorological data at different latitude and longitude locations in vertical layers at different altitudes;

[0160] Acquire second meteorological data within each preset time step of a preset time period;

[0161] Based on the first and second meteorological data, the meteorological parameters of the wind farm in different times and spaces are obtained by fusion.

[0162] Based on the meteorological parameters of the wind farm in different times and spaces, a five-dimensional spatiotemporal tensor is constructed:

[0163] X∈R T×H×W×D×C

[0164] In the above formula, X is the five-dimensional spacetime tensor, T represents the preset time step, HxW represents the spatial grid, D is the vertical layer, and C is the meteorological channel.

[0165] In one possible implementation, a data augmentation module is also included, which is used to: perform spatial data augmentation on the multidimensional spatiotemporal tensor using elastic deformation; and perform temporal data augmentation on the multidimensional spatiotemporal tensor using a non-uniform sampling strategy.

[0166] In one possible implementation, the decoupling module is specifically used for:

[0167] The meteorological parameters of each grid point in the spatiotemporal tensor are normalized, and the topographic elevation data of the wind farm are interpolated to each vertical layer to form a three-dimensional topographic feature map.

[0168] A lightweight network is used to predict the 3D position offset of each grid point, and non-uniform sampling is performed based on the offset.

[0169] Multi-scale convolutional windows are used to extract features from the spatiotemporal tensor after non-uniform sampling, resulting in system-level features, mesoscale features, and micro-meteorological features, forming a spatial feature map.

[0170] The spatiotemporal tensor is separated into temporal features using a dynamic attention mechanism to obtain the separated temporal features.

[0171] Wavelet frequency domain decomposition of time features is used to obtain time series streams in different frequency bands.

[0172] In one possible implementation, the space extraction module is specifically used for:

[0173] The spatial feature maps are rearranged based on vertical layers and meteorological parameters to form a three-dimensional data structure of meteorological parameters;

[0174] Based on a three-dimensional data structure, cross-layer correlation analysis is performed on the multi-layer meteorological parameters of each horizontal grid point to establish a three-dimensional feature representation.

[0175] By using a preset 3D convolution window, the 3D feature representation is scanned layer by layer to extract meteorological parameters and convolution features at different scales.

[0176] By adaptively fusing convolutional features at different scales, a global meteorological field feature vector carrying spatial location encoding is obtained.

[0177] In one possible implementation, the time extraction module is specifically used for:

[0178] Based on the degree of data fluctuation in the time series stream, dynamic windows of different sizes are used to capture features of the time series stream;

[0179] The phase-aware attention mechanism is used to weight the temporal features captured by the feature acquisition to obtain temporal features at different time scales. The temporal features at different time scales include short-term temporal features that characterize short-term fluctuation details and long-term temporal features that characterize long-term evolution trends.

[0180] In one possible implementation, the feature fusion module is specifically used for:

[0181] Three-dimensional adaptive max pooling is performed on the feature vector of the global meteorological field to obtain the pooling result;

[0182] Long and short time series features are aligned to standard length using thread interpolation to obtain the alignment result;

[0183] Based on the meta-learning weight allocator, the first weight of the pooling result and the second weight of the alignment result are obtained;

[0184] Based on the first and second weights, the pooling results and alignment results are weighted and fused to obtain the fused features.

[0185] In one possible implementation, the model determination module is specifically used for:

[0186] Using a composite evaluation system, the fusion characteristics are evaluated hierarchically at different time scales, and spatial correlation analysis is conducted to obtain the evaluation results. The composite evaluation system includes deterministic indicators and probabilistic indicators.

[0187] Based on the evaluation and analysis results, the optimal prediction model is determined among the various prediction models using the entropy weight method-TOPSIS hybrid decision-making mechanism.

[0188] The wind power prediction device provided in this embodiment can execute the method provided in the above method embodiment. Its implementation principle and technical effect are similar, and will not be described in detail here.

[0189] Figure 9 A schematic diagram of the structure of the electronic device provided in this application. Figure 9 As shown, the electronic device 90 provided in this embodiment includes at least one processor 901 and a memory 902. Optionally, the device 90 further includes a communication component 903. The processor 901, memory 902, and communication component 903 are connected via a bus.

[0190] In a specific implementation, at least one processor 901 executes computer execution instructions stored in memory 902, causing at least one processor 901 to perform the above-described method.

[0191] The specific implementation process of processor 901 can be found in the above method embodiments, and its implementation principle and technical effect are similar. It will not be repeated here.

[0192] In the above embodiments, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor.

[0193] The memory may include random access memory (RAM) and may also include non-volatile memory (NVM), such as at least one disk storage device.

[0194] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, the buses shown in the accompanying drawings are not limited to a single bus or a single type of bus.

[0195] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0196] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the above-described method.

[0197] The aforementioned readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.

[0198] An exemplary readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and the readable storage medium can exist as discrete components in the device.

[0199] The division of units is merely a logical functional division; in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0200] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0201] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0202] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0203] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0204] Finally, it should be noted that other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This invention is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein, and is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.

Claims

1. A method for predicting wind power output, characterized in that, include: Meteorological parameters of wind farms at different times and spaces are obtained, and spatiotemporal tensors are constructed. The spatiotemporal tensor is decoupled by features to separate the spatial feature map and the time series stream; The spatial feature map is used to extract features through a spatial channel to obtain a global meteorological field feature vector. Feature extraction is performed on the time series stream through the time channel to obtain long- and short-term time series features; The global meteorological field feature vector and the long- and short-term time-series features are dynamically fused using a meta-learning weight allocator to obtain the fused features; The various prediction models are evaluated based on the fusion features to determine the optimal prediction model; Based on the optimal prediction model, the power of the wind farm is predicted.

2. The method according to claim 1, characterized in that, The acquisition of meteorological parameters of the wind farm at different times and spaces, and the construction of a spatiotemporal tensor, includes: Acquire the first meteorological data at different latitude and longitude locations in vertical layers at different altitudes; Acquire second meteorological data within each preset time step of a preset time period; Based on the first meteorological data and the second meteorological data, the meteorological parameters of the wind farm in different times and spaces are obtained by fusion; Based on the meteorological parameters of the wind farm in different times and spaces, a five-dimensional spatiotemporal tensor is constructed: X∈R T×H×W×D×C In the above formula, X is the five-dimensional spacetime tensor, T represents the preset time step, HxW represents the spatial grid, D is the vertical layer, and C is the meteorological channel.

3. The method according to claim 2, characterized in that, After constructing the spatiotemporal tensor, the method further includes: Spatial data augmentation is performed on the spatiotemporal tensor using elastic deformation; Temporal data augmentation is performed on the spatiotemporal tensor using a non-uniform sampling strategy.

4. The method according to claim 2, characterized in that, The step of decoupling the spatiotemporal tensor to separate the spatial feature map and the time series stream includes: The meteorological parameters of each grid point in the spatiotemporal tensor are normalized, and the topographic elevation data of the wind farm are interpolated to each vertical layer to form a three-dimensional topographic feature map. A lightweight network is used to predict the 3D position offset of each grid point, and non-uniform sampling is performed based on the offset. Multi-scale convolutional windows are used to extract features from the spatiotemporal tensor after non-uniform sampling to obtain system-level features, mesoscale features, and micro-meteorological features, forming the spatial feature map. The spatiotemporal tensor is subjected to temporal feature separation using a dynamic attention mechanism to obtain the separated temporal features. The time features are decomposed using wavelet frequency domain to obtain time series streams in different frequency bands.

5. The method according to claim 2, characterized in that, The step of extracting features from the spatial feature map through a spatial channel to obtain a global meteorological field feature vector includes: The spatial feature map is rearranged based on the vertical layer and meteorological parameters to form a three-dimensional data structure of the meteorological parameters; Based on the aforementioned three-dimensional data structure, cross-layer correlation analysis is performed on the multi-layer meteorological parameters of each horizontal grid point to establish a three-dimensional feature representation. By using a preset three-dimensional convolution window, the three-dimensional feature representation is scanned layer by layer to extract meteorological parameters and convolution features at different scales. By adaptively fusing convolutional features at different scales, a global meteorological field feature vector carrying spatial location encoding is obtained.

6. The method according to claim 1, characterized in that, The step of extracting features from the time series stream through the time channel to obtain long- and short-term time series features includes: Based on the data fluctuation level of the time series stream, dynamic windows of different sizes are used to capture features of the time series stream; The phase-aware attention mechanism is used to weight the temporal features captured by the feature acquisition to obtain temporal features at different time scales. The temporal features at different time scales include short-term temporal features that characterize short-term fluctuation details and long-term temporal features that characterize long-term evolution trends.

7. The method according to claim 1, characterized in that, The process of dynamically fusing the global meteorological field feature vector and the long- and short-term time-series features through a meta-learning weight allocator to obtain fused features includes: Three-dimensional adaptive max pooling is performed on the global meteorological field feature vector to obtain the pooling result; The long and short-term time series features are aligned to the standard length using thread interpolation to obtain the alignment result; Based on the meta-learning weight allocator, the first weight of the pooling result and the second weight of the alignment result are obtained; Based on the first weight and the second weight, the pooling result and the alignment result are weighted and fused to obtain the fused feature.

8. The method according to claim 1, characterized in that, The step of evaluating each prediction model based on the fusion features and determining the optimal prediction model includes: Using a composite evaluation system, the fusion features are evaluated hierarchically at different time scales, and spatial correlation analysis is performed to obtain the evaluation results. The composite evaluation system includes deterministic indicators and probabilistic indicators. Based on the evaluation and analysis results, the optimal prediction model is determined among the various prediction models using the entropy weight method-TOPSIS hybrid decision-making mechanism.

9. A wind power prediction device, characterized in that, include: A module is built to obtain meteorological parameters of the wind farm at different times and spaces, and to construct a spatiotemporal tensor; The decoupling module is used to decouple the spatiotemporal tensor by separating the spatial feature map and the time series stream. The spatial extraction module is used to extract features from the spatial feature map through a spatial channel to obtain a global meteorological field feature vector; The time extraction module is used to extract features from the time series stream through the time channel to obtain long and short time series features; The feature fusion module is used to dynamically fuse the global meteorological field feature vector and the long- and short-term time-series features through a meta-learning weight allocator to obtain fused features; The model determination module is used to evaluate each prediction model based on the fusion features and determine the optimal prediction model; The power prediction module is used to predict the power of the wind farm based on the optimal prediction model.

10. An electronic device, characterized in that, include: Memory, processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory, causing the processor to perform the method as described in any one of claims 1-8.

Images