Wind farm image and multi-physical constraint fused wind power prediction method and system

Abstract

The application discloses a wind power prediction method and system fusing wind field images and multiple physical constraints, and belongs to the technical field of new energy power generation, aiming to improve the accuracy and stability of short-term wind power prediction. The method comprises the following steps: obtaining wind field image data and performing pretreatment; extracting features by using a dynamic time sequence encoder, wherein the dynamic time sequence encoder comprises two parallel branches of a Fourier neural operator and a convolution gate recurrent unit, which are respectively used for capturing global frequency spectrum features and local transient features of the wind field; introducing spatial variable gate parameters and learnable channel weights to fuse features of each branch; based on the fused features, a neural network architecture driven by physical information is used to realize multi-step wind field simulation; and multi-step wind power prediction is performed on multiple wind farms by using a time sequence fusion Transformer according to multi-step wind field images obtained through the wind field simulation. The application overcomes the shortcomings of existing methods in capturing the space-time dependence and local rapid disturbance of the wind field.

Description

Technical Field

[0001] This invention belongs to the field of new energy power generation technology, specifically relating to a wind power prediction method and system that integrates wind field images with multiple physical constraints. Background Technology

[0002] With the continuous transformation of the renewable energy structure, wind power has become a key component of the global clean energy supply system. Relying on natural wind energy resources, wind power generation boasts significant advantages such as cleanliness, low carbon emissions, sustainable utilization, and environmental friendliness, and has received widespread attention in the energy strategies of various countries, becoming an important support for promoting the green energy transition. However, due to the influence of multiple environmental factors such as wind speed, wind direction, atmospheric boundary layer evolution, and topographic disturbances, wind power output exhibits strong intermittency, randomness, and spatiotemporal non-stationarity.

[0003] This volatile characteristic not only restricts the stability and controllability of wind farm operation itself, but also puts significant pressure on the real-time balance of the power grid, reserve capacity allocation, and market transactions. Especially with the continuous increase in wind power penetration, high-precision, short-cycle wind power forecasting has become a key technical requirement for ensuring grid security, improving absorption capacity, and assisting dispatch decisions. Traditional forecasting methods mostly rely on historical data from single sites or numerical weather forecast outputs, which often fail to effectively characterize the complex spatiotemporal correlation structure and rapid local disturbances in wind farms, leading to increased forecast errors and insufficient robustness under conditions of sudden weather changes or complex terrain.

[0004] In recent years, with the advancement of spatiotemporal data processing technology and artificial intelligence models, prediction methods that integrate wind field images, multi-scale meteorological information, and physical constraints have gradually attracted attention. In particular, the technical approach based on high-resolution wind field simulation, multi-physical information coupling, and end-to-end deep learning frameworks has provided new possibilities for overcoming existing prediction bottlenecks. How to construct a wind power prediction model that can simultaneously ensure physical consistency, spatiotemporal accuracy, and operational efficiency has become one of the core challenges in promoting large-scale grid connection and intelligent regulation of wind power. Summary of the Invention

[0005] Purpose of the invention: The purpose of this invention is to provide a wind power prediction method and system that integrates wind field images with multiple physical constraints, thereby improving the accuracy of short-term wind power prediction.

[0006] Technical Solution: To achieve the above-mentioned objectives, the wind power prediction method merging wind field images and multiple physical constraints proposed in this invention includes the following steps:

[0007] Wind field image data is acquired and preprocessed using a time window to obtain a four-channel wind field image composed of pixel coordinates and horizontal wind components;

[0008] Based on the four-channel wind field image, a dynamic temporal encoder is used to extract features. The dynamic temporal encoder includes two parallel branches: a Fourier neural operator and a convolutional gated recursive unit, which are used to capture the global spectral features and local transient features of the wind field, respectively. The features of each branch are fused by spatially variable gating parameters and learnable channel weights to obtain the fused spatiotemporal features of the wind field.

[0009] Based on the fused spatiotemporal characteristics of the wind field, an encoder is used to simulate and generate multi-step future wind field images through a multi-objective physical constraint-driven approach. Multi-objective physical constraint-driven approach refers to the simultaneous introduction of multiple soft constraint terms based on atmospheric physical laws during the simulation process. These include at least wind field divergence constraints to ensure mass conservation, vorticity retention constraints to maintain circulation structure stability, and spectral consistency constraints to ensure that energy distribution conforms to turbulence spectral laws. Each constraint is embedded into the loss function through a weighted combination to guide the model to achieve a balance between data fitting and physical rationality.

[0010] Using multi-step future wind field images obtained from wind field simulation, combined with historical power data, wind farm geographical location information and meteorological data, a time-series fusion Transformer architecture is used to perform multi-step power prediction for multiple wind farms, and output multi-step future power prediction results.

[0011] Furthermore, the preprocessing of wind field image data using time windows specifically includes:

[0012] Extract a length of T using a sliding window method. w The continuous wind field image frames are processed by normalizing, imputing missing values, aligning spatial coordinates, and stacking channels for each frame to form a four-channel input tensor with a unified format. Where X, Y are the normalized pixel coordinates. Let H and W represent the components of the horizontal wind vector at time τ in the meridional and zonal directions, respectively, and H and W be the height and width of the wind field grid.

[0013] Furthermore, the process of feature extraction by the dynamic temporal encoder includes:

[0014] In the window of history Internally, each wind field image frame is mapped to channel features through a shared backbone network. : ,in, This represents the feature mapping function, where H and W are the height and width of the wind field grid, C is the number of channels after feature mapping, and T is the feature mapping function. w The length to be truncated for the sliding window, where t represents the t-th time step;

[0015] The Fourier neural operator branch uses two Fourier function mapping modules to act on channel features. The formula is as follows: ,in, This represents the output of global spectral characteristics. This represents the Fourier function mapping module;

[0016] Convolutional gated recursive units recursively traverse the features of each channel. Output local transient features The formula is as follows:

[0017] ;

[0018] in, It is the Sigmoid activation function. For Hadamard products, This is the hidden state from the previous moment. To update the door, control the degree to which its hidden state from the previous moment is retained. To reset the gate, control the degree of influence of the previous hidden state on the current candidate state. , , These represent the convolution operations of the corresponding gates, and each of them has independent learnable parameters.

[0019] Furthermore, features from each branch are fused using spatially variable gating parameters and learnable channel weights, including:

[0020] ,

[0021] ,

[0022] Where G represents the gated unit function, Indicates adaptive weights, These represent global spectral features and local transient features, respectively. It is the Sigmoid activation function. This represents the learnable scaling factor. This represents the output fusion vector, i.e., the fused spatiotemporal features of the wind field.

[0023] Furthermore, the encoder simulates and generates multi-step future wind field images through a multi-objective physical constraint-driven approach, including:

[0024] The fused spatiotemporal features of the wind field F are input into the shared decoder. The decoded feature vector is obtained. T w The length of the sliding window is truncated, and τ represents the time step τ; a single-step predicted wind field is generated based on the mapping function. ,in For those with learnable parameters The mapping function is used to convert the decoded feature vector into a four-channel wind field image; then the generated predicted wind field is slidably stitched with the historical wind field sequence as the input for the next iteration step, and multi-step autoregressive inference is performed.

[0025] During the training phase, the encoder calculates the predicted wind field at step h. Calculate divergence loss vorticity loss Spectrum loss The summation constitutes the physical loss. And together with reconstruction loss and power loss, the model parameters are optimized; among them, divergence loss constitutes the wind field divergence constraint to ensure mass conservation, vorticity loss constitutes the vorticity retention constraint to maintain the stability of the circulation structure, and spectrum loss constitutes the spectrum consistency constraint to ensure that the energy distribution conforms to the turbulent spectrum law.

[0026] in, This represents the horizontal component of the wind speed predicted by the model. This represents the vertical component of the wind speed predicted by the model. This represents the vorticity value predicted by the model. This represents the actual vorticity value; This represents the spectrum predicted by the model. Represents the actual spectrum, This indicates the difference between predicted vorticity and actual vorticity. This indicates the spectral difference between the predicted wind field and the actual wind field. , , The learnable or preset weight coefficients for each loss term, and h represents the time step index.

[0027] Furthermore, physical loss Together with reconstruction loss and power loss, the model parameters are optimized, and the total loss function for model optimization is:

[0028] ,

[0029] in, These are learnable variance weights used to balance heterogeneous losses; the first term is the reconstruction loss, which is the mean square error between the predicted wind speed and the actual wind speed at step t+1. Indicates predicted wind speed, Indicates the actual wind speed; Indicates physical loss; The power loss is represented by the average of the squared differences between the predicted and actual power at all wind farm sites over all predicted time steps.

[0030] Furthermore, the temporal fusion Transformer architecture performs multi-step power prediction for multiple wind farms, including:

[0031] For station i, the grid features are transformed into a sequence through differentiable sampling: ,in, This represents the grid wind field feature vector of station i. Represents the mesh feature mapping function;

[0032] At each time step, the projected view of the sampled features, historical power, auxiliary measurements, and static metadata are concatenated: ,in, Indicates splicing characteristics, Indicates historical power. Indicates measurement data; This refers to static metadata, specifically the geographical location information of the wind farm, including latitude and longitude coordinates (x, y, y). i , y i and terrain conditions c i ; , , , Indicates learnable weights;

[0033] Stack them above the window to form a sequence ordered by site: , where d represents the feature dimension of the hidden layer after feature concatenation and mapping;

[0034] A recursive encoder is used to include historical information: , Indicates a recursive encoder;

[0035] The timing decoder uses learnable time-domain markings Perform a query The process generates embedding vectors in the specified time domain, and the post-processing head maps these embedding vectors to scalars. , The decoding function combines the encoded context information with a learnable temporal label to generate an embedded representation for a specified time step. This represents the post-processing function used to map the high-dimensional embedding vector output by the decoder to the final scalar power prediction value;

[0036] After summarizing all sites, a multi-site, multi-step prediction matrix is ​​generated: N is the total number of stations. The length of time in the future.

[0037] The wind power prediction system fused with wind field images and multiple physical constraints provided by this invention includes:

[0038] The wind field image data preprocessing module is used to acquire wind field image data and preprocess it through a time window to obtain a four-channel wind field image composed of pixel coordinates and horizontal wind components.

[0039] The wind field spatiotemporal feature extraction module is used to extract features from four-channel wind field images using a dynamic temporal encoder. The dynamic temporal encoder includes two parallel branches: a Fourier neural operator and a convolutional gated recursive unit, which are used to capture the global spectral features and local transient features of the wind field, respectively. The features of each branch are fused by spatially variable gating parameters and learnable channel weights to obtain the fused spatiotemporal features of the wind field.

[0040] The wind field simulation and extrapolation module is used to simulate and generate multi-step future wind field images based on the fused spatiotemporal characteristics of the wind field. The encoder uses a multi-objective physical constraint-driven approach to simulate and generate the images. The multi-objective physical constraint-driven approach refers to the simultaneous introduction of multiple soft constraint terms based on atmospheric physical laws during the simulation process. These include at least the wind field divergence constraint to ensure mass conservation, the vorticity retention constraint to maintain the stability of the circulation structure, and the spectral consistency constraint to ensure that the energy distribution conforms to the turbulence spectrum law. Each constraint is embedded into the loss function through a weighted combination to guide the model to achieve a balance between data fitting and physical rationality.

[0041] The power time series prediction module is used to perform multi-step power prediction on multiple wind farms by combining multi-step future wind farm images obtained from wind farm simulation with historical power data, wind farm geographical location information and meteorological data, and output multi-step future power prediction results through the time series fusion Transformer architecture.

[0042] The present invention also 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, wherein when the programs are executed by the processors, they implement the steps of the wind power prediction method for fusing wind field images and multiple physical constraints as described above.

[0043] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the wind power prediction method that fuses wind field images and multiple physical constraints as described above.

[0044] The present invention also provides a computer program product, including a computer program, characterized in that, when the computer program is executed by a processor, it implements the steps of the wind power prediction method that fuses wind field images and multiple physical constraints as described above.

[0045] Beneficial effects:

[0046] (1) This invention employs a dual-branch dynamic temporal coding method, which extracts the global spectral features and local spatiotemporal features of wind field images by fusing Fourier neural operators and convolutional gated recurrent units. This method can simultaneously model the long-range meteorological evolution law and short-term rapid disturbances in the wind field, thereby improving the ability to characterize the dynamics of the wind field under complex terrain and sudden weather conditions.

[0047] (2) This invention employs a multi-objective learning method for physical information. By introducing multiple physical constraints such as mass conservation, vorticity preservation, and spectral consistency, and combining them with an adaptive loss weight adjustment mechanism, the method collaboratively optimizes data fitting and physical consistency during the training process. This method ensures that the wind field simulation results conform to the laws of fluid mechanics while maintaining high numerical accuracy, providing reliable input for subsequent power prediction.

[0048] (3) This invention constructs an end-to-end wind power prediction framework, which realizes adaptive fusion of wind field regional features through spatially variable gating units, and performs unified modeling of wind field embedding vectors and historical power data in a time-series fusion transformer. This method effectively avoids feature distribution shift caused by task decoupling in traditional two-stage prediction, and improves the robustness and consistency of multi-step prediction. Attached Figure Description

[0049] Figure 1 This is a flowchart of the wind power prediction method of the present invention;

[0050] Figure 2 This is a schematic diagram of the Fourier neural operator branches used in this invention;

[0051] Figure 3 This is a schematic diagram of the branching of the convolution gated loop unit used in this invention;

[0052] Figure 4 This is a schematic diagram of the physical information neural network module used in this invention;

[0053] Figure 5 This is a schematic diagram of wind field simulation in an embodiment of the present invention;

[0054] Figure 6 This is a schematic diagram of the prediction result fitting using the proposed method in an embodiment of the present invention;

[0055] Figure 7 This chart compares the performance metrics of different prediction methods across different prediction time domains. Detailed Implementation

[0056] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0057] Reference Figure 1This invention provides a wind power prediction method that integrates wind field images with multiple physical constraints, comprising the following steps:

[0058] Step 1: Acquire wind field image data and preprocess the wind field image through a time window to obtain a four-channel wind field image composed of pixel coordinates and horizontal wind components.

[0059] Short-term wind power prediction is defined as a spatiotemporal sequence prediction task with multimodal inputs and multiple objective goals. This invention addresses the problem of short-term prediction on a wind field grid. It uses pixel values ​​from wind field images to quantitatively represent wind speed intensity, forming a continuous spatial distribution on an 81×81 (or other sizes) grid, and uses a length of... The time window. Wind field image data refers to image-type data organized in the form of a two-dimensional grid that characterizes the spatial wind field state. Each pixel corresponds to a grid point in geographic space, and the pixel value quantitatively reflects the horizontal wind component at that location. Wind speed, intensity, and wind direction angle; a time window refers to continuously collected data of a fixed length. Historical wind field image sequences are used to construct temporal input samples.

[0060] Preprocessing includes normalizing the original wind field image, imputing missing values, aligning spatial coordinates, and stacking channels to form a four-channel input tensor with a uniform format. Where X, Y are the normalized pixel coordinates. These represent the meridional and zonal components of the horizontal wind vector at time τ, respectively. This preprocessing ensures that the input data meets the processing requirements of the subsequent dynamic time encoder in terms of spatial scale, numerical range, and channel semantics.

[0061] In this embodiment, the time window preprocessing method is as follows: first, the original wind field image is interpolated for missing values ​​and aligned to spatial coordinates; then, a sliding window is used to truncate the image to a length of T. w =12 consecutive wind field image frames, and finally perform Z-score normalization on the pixel values ​​in each frame, and then compare the X,Y coordinate map with The wind component map is stitched together along the channel dimension to form a single four-channel tensor.

[0062] Furthermore, the preprocessing method also employs the following steps: applying Gaussian filtering to the wind component image to suppress high-frequency noise while preserving key dynamic features such as vortex structures, and then scaling it to the [0,1] interval using Min-Max to adapt to the dynamic range of the neural network input.

[0063] After the processing in step 1, the subsequent model will receive a stack of four-channel wind field images. The formula is as follows:

[0064]

[0065] Indicates from a historical moment up to the current moment A collection of continuous wind field image sequences.

[0066] Step 2: The wind field image is processed using a dynamic temporal encoder, which includes two parallel branches: a Fourier Neural Operator (FNO) and a Convolutional Gated Recurrent Unit (ConvGRU), used to capture the global spectral features and local transient features of the wind field, respectively. An adaptive weighting mechanism is used to optimize the fusion effect of the wind field features based on the learned spatially varying gating units and channel weights.

[0067] A dynamic temporal encoder is a module specifically designed for processing spatiotemporal feature extraction from wind field image sequences. Its input is a preprocessed four-channel wind field image sequence within a time window, and its output is a fused high-dimensional spatiotemporal feature tensor. Fourier neural operators (FNO) are used in this context. Figure 2 As shown, this is a deep learning operator based on frequency domain modeling, used to model the large-scale, low-frequency spatial structure evolution of wind fields. Its application is to the spatial spectral response of the entire wind field image. The processing method of the FNO branch is: [The text abruptly ends here, likely due to an incomplete translation or source material.] Each frame of feature map is subjected to a two-dimensional fast Fourier transform. In the frequency domain, a learnable complex weight matrix is ​​applied for spectral modulation, and then the inverse FFT is used to restore it to the spatial domain features. This process is repeated twice to complete multi-scale spectral modeling.

[0068] Convolutional gated recurrent unit (ConvGRU) such as Figure 3 As shown, this is a loop structure combining convolution operations and gating mechanisms, used to model nonlinear, short-term dynamic changes in local wind field regions, such as sudden gusts and boundary layer disturbances. The ConvGRU branch processing method includes: using the initial hidden state... Start, sequentially ,…, Input ConvGRU cells, update hidden state at each step Output the local transient representation at the current time. Ultimately This will be the output of that branch.

[0069] The specific processing procedure of the dynamic timing encoder is as follows:

[0070] In the window of history Internally, each wind field image frame is first mapped to a shared backbone network. Channel characteristics : ,in, Represents the feature mapping function; to capture complementary temporal dependencies, two parallel branches act on { };

[0071] The FNO branch uses two FNO modules to act. To model large-scale and low-frequency structures, the formula is as follows:

[0072]

[0073] in, Represents the global spectral characteristics output. This represents the Fourier function mapping module.

[0074] ConvGRU branches use ConvGRU units to recursively traverse... This integrates local and rapidly changing patterns, outputting a local transient representation. The formula is as follows:

[0075] ;

[0076] in, It is the Sigmoid activation function. For Hadamard products, This is the hidden state from the previous moment. To update the door, control the degree to which its hidden state from the previous moment is retained. To reset the gate, control the degree of influence of the previous hidden state on the current candidate state. , , These represent the convolution operations of the corresponding gates, and each of them has independent learnable parameters.

[0077] Temporal dynamic features are decomposed into global spectral patterns and local transient features through parallel processing. This invention introduces spatially variable gating parameters and learnable channel weights to fuse the branches. First, the output features of FNO and ConvGRU are concatenated along the channel dimension, and then a spatially gated map G∈[0,1] is generated by 1×1 convolution and Sigmoid activation. Then, the outputs of the two branches are multiplied by the learnable channel vectors α and (1−α) respectively. Finally, the fused features are obtained by pixel-level weighted summation, as shown in the following formula:

[0078]

[0079]

[0080] Where G represents the gated unit function, Indicates adaptive weights, This represents the output fusion vector.

[0081] The adaptive weighting mechanism refers to a spatial-channel joint weighting strategy, which consists of two core parts: spatially variable gating units refer to soft masks with values ​​ranging from [0,1] that are generated independently at each spatial location of the feature map, used to control the contribution ratio of FNO and ConvGRU branches at that location; channel weights refer to learnable scaling coefficients α applied to the output features of FNO and ConvGRU in the channel dimension, used to adjust the information gain of different semantic channels.

[0082] Simultaneously, this dynamic temporal encoder returns the channel feature vector for each time step. This is used for downstream sampling at the site level. Fusion vector While maintaining sensitivity to local mutations, it incorporates large-scale terrain information.

[0083] Step 3: Based on the fused spatiotemporal features of the wind field, wind field simulation and extrapolation are performed. The wind field simulation and extrapolation generates multi-step future wind field images through a multi-objective physical constraint-driven approach.

[0084] The fused spatiotemporal features of the wind field refer to the feature tensor F output in step 2, which integrates the global spectral pattern and local transient disturbances as the initial hidden state of the wind field evolution model. Wind field simulation and extrapolation refer to the process of generating a sequence of wind field images for multiple future time steps through the decoder, starting from this feature. Multi-objective physical constraint driving refers to the simultaneous introduction of multiple soft constraint terms based on atmospheric physical laws during the simulation process, including at least wind field divergence constraints (ensuring mass conservation), vorticity retention constraints (maintaining circulation structure stability), and spectral consistency constraints (ensuring energy distribution conforms to turbulent spectral laws). Each constraint is embedded into the loss function through weighted combination to guide the model to achieve a balance between data fitting and physical rationality.

[0085] The wind field simulation method is as follows: input F into the shared decoder to generate a single-step predicted wind field. , For those with learnable parameters The mapping function is used to convert the decoded feature vector into a four-channel wind field image; then it is slidably stitched with the historical wind field sequence and used as the input for the next iteration step, and the process is repeated. This is the first time to complete. Step-by-step regression deduction;

[0086] Multi-objective physics-constrained methods include: during the training phase, for each step of the predicted wind field... Calculate divergence loss vorticity loss With spectral loss The weighted sum of all losses constitutes the physical constraint term. That is, physical loss, which, together with reconstruction loss, optimizes model parameters. Its structural diagram is as follows: Figure 4 As shown.

[0087] The wind field simulation employs a physical guided sampling strategy during the inference phase, which applies a lightweight projection operation to each prediction result and redistributes the wind components in the neighborhood of pixels that do not meet the divergence threshold using the least squares method, ensuring that the output wind field is basically reasonable in terms of physical interpretability.

[0088] Specifically, the fused feature vector F obtained in step 2 is input into the decoder, and the output is the predicted wind field embedding and the predicted wind field image:

[0089]

[0090] A physical information-driven neural network architecture is used to simulate multi-step wind fields. The physical information-based module collaboratively optimizes field accuracy, physical plausibility, and power accuracy through mean squared error uncertainty weights. The total loss function is:

[0091]

[0092] in, These are learnable variance weights used to balance heterogeneous losses; the first term is the reconstruction loss (i.e., the mean square error between predicted and actual wind speeds). Indicates predicted wind speed, Indicates the actual wind speed. Indicates physical loss. The power loss is represented by the mean square error. It is calculated as the average of the squares of the difference between the predicted power and the actual power at all prediction time steps for all wind farm sites. This average is used to measure the error between the predicted power and the actual power.

[0093] Physical loss In this invention, the following are obtained through multi-objective physical constraints:

[0094] Divergence (mass conservation) loss:

[0095]

[0096] in, The weighting coefficients represent the divergence loss. This represents the x-direction component of the wind speed predicted by the model. This represents the y-direction component of the wind speed predicted by the model.

[0097] Rotation (circulation structure) loss, also known as vorticity loss:

[0098]

[0099] in, The weighting coefficients representing curl loss, This represents the curl value predicted by the model. This represents the true curl (vorticity) value of the model.

[0100] Spectral (energy distribution) loss, also known as spectral loss:

[0101]

[0102] in, The weighting coefficients representing the spectral loss. Represents the predicted spectrum, Represents the actual spectrum, It represents the set of frequency components in wavenumber space.

[0103] Physical information belongs to aggregated complementary prior information, which functions as soft constraints (inspired by PINN but not for solving partial differential equations). Furthermore, physical constraint terms... In addition to the three basic constraints mentioned above (divergence, vorticity, and spectral consistency), the following extended form is adopted to introduce correlation constraints, gradient constraints, and structural similarity constraints to further enhance the physical rationality of the wind field simulation:

[0104]

[0105] in, The vorticity is represented by s = |(u,v)|², which is the modulus of wind speed. To predict the gradient of the wind speed modulus and preserve the edge structure of the wind field, This represents the spectral difference between the predicted wind field and the actual wind field, i.e., the spectral loss. This represents the correlation between the predicted wind speed at step t+1 and the actual wind speed at step t, i.e., the correlation loss. This represents the difference between the predicted vorticity and the actual vorticity, i.e., vorticity loss; This represents the difference between the gradient of the predicted wind speed model and the gradient of the actual wind speed model, i.e., gradient loss. The difference between the predicted wind speed modulus and the actual wind speed modulus is the structural similarity loss. , , , , These are the learnable or preset weight coefficients for each constraint term. and Performing a model warm-up helps avoid premature over-regularization when data terms are dominant. All spatial derivatives are obtained using centered finite difference with grid-aware spacing; orthogonal FFT is used for spectral terms.

[0106] The framework proposed in this invention enables multi-step wind field simulation, automatically generating a series of future wind field images. This process aims to model the spatiotemporal evolution of atmospheric flow as a dynamic system. The simulation begins with an initial state defined by features from the most recently observed wind field frame. The model first generates authoritative single-step predictions as the initial state for the entire simulation:

[0107]

[0108] in, This represents a spatiotemporal prediction module that combines a dynamic temporal encoder and a physical information decoder to capture potential wind field dynamics and generate predictions for the next moment.

[0109] For subsequent time steps, the model enters a recursive autoregressive expansion phase. It automatically advances the system state by using its previous predictions as new initial conditions. From Initially, the model iteratively generates multi-step wind field images:

[0110]

[0111] Step 4: Using the multi-step wind field images obtained from wind field simulation, perform multi-step power prediction on multiple wind farms through time-series fusion Transformer (TFT) and output multi-step future power prediction results.

[0112] The wind field simulation results are as follows, derived from multi-step wind field images obtained through wind field simulation. Figure 5 As shown, a multi-step power prediction for multiple wind farms is performed using a temporal fusion Transformer (TFT). The TFT employs historical power data, wind farm geographic location information, wind farm simulation data, and meteorological data to perform time-series modeling and output multi-step future power prediction results. The multi-step wind field image obtained from wind farm simulation refers to the wind field prediction vector sequence output in step 3, which covers the future... Wind farm status is displayed hourly or per hourly; the Time-Series Fusion Transformer (TFT) is an encoder-decoder architecture for multi-source heterogeneous time-series data, specifically designed to fuse site-level static attributes, historical observations, and physical simulation outputs; historical power data refers to the actual power generation sequence of each wind farm over the past Tw hours, used to characterize equipment response inertia and power curtailment behavior; wind farm geographic location information includes latitude and longitude coordinates (x, y, y). i , y i and terrain conditions c iThe system includes elements such as wind field sampling mapping for spatial perception; meteorological data refers to auxiliary variables other than wind field images, such as gridded or station observation data of temperature, air pressure, and humidity, which serve as external condition inputs; and multi-step future power prediction results refer to the output of future power for each of the N wind farms. Hourly power prediction sequence The final sum is N× Prediction matrix:

[0113]

[0114] The TFT processing method is as follows: For each wind farm i, first, at its geographical location (x i ,y i For each simulated wind field Perform bilinear differentiable sampling to obtain the wind field feature vector corresponding to the site. Then compare it with historical power Measurement data and static metadata [x i , y i ,c i [Concatenate into multimodal features] After stacking, the timing context representation H is extracted by the input TFT encoder. i Then, the decoder with time-domain query markers generates hourly power predictions.

[0115] Specifically, for station i, the grid features are transformed into a sequence through differentiable sampling:

[0116]

[0117] in, This represents the grid wind field feature vector of station i. Mesh feature mapping function.

[0118] At each time step, the projected view of the sampled features, historical power, auxiliary measurements, and static metadata are concatenated:

[0119]

[0120] in, Indicates splicing characteristics, Indicates historical power. Indicates measurement data, Represents static metadata. , , , This represents the learnable weights.

[0121] Stack them above the window to form a sequence ordered by site:

[0122]

[0123] Where d represents the feature dimension of the hidden layer after feature concatenation and mapping.

[0124] A recursive encoder is used to include historical information:

[0125]

[0126] in, This indicates a recursive encoder.

[0127] The timing decoder uses learnable time-domain markings Perform a query This generates embedding vectors in the specified time domain. The post-processing head then maps these vectors to scalars.

[0128]

[0129] in, The decoding function combines the encoded context information with a learnable temporal label to generate an embedded representation for a specified time step. This represents the post-processing function used to map the high-dimensional embedding vector output by the decoder to the final scalar power prediction value.

[0130] After summarizing all (N) sites, a multi-site, multi-step prediction matrix can be generated:

[0131]

[0132] To verify the effectiveness of the proposed method, the following experiments were conducted in the embodiments. The meteorological dataset used was a reanalysis dataset provided by a regional medium-range weather forecasting center, covering a geographical area from 28° to 38° north latitude and 113° to 123° east longitude, with a spatial resolution of 0.1° × 0.1° (approximately 9.6 kilometers). The wind field image data had a temporal resolution of 1 hour, an image size of 81 × 81 pixels, and included horizontal wind speed component information. After data alignment and normalization preprocessing, data from January 2021 to August 2022 was used as the training set, and data from August 2022 to August 2023 was used as the test set. Simultaneously, an end-to-end wind field-power joint prediction experiment was conducted using power record data from three actual wind farms located in a region of Jiangsu Province. The performance of the model in both wind field simulation and power prediction was comprehensively evaluated through multiple sets of comparative experiments and quantitative indicators. In this embodiment, the following error evaluation metrics are used: Normalized Mean Absolute Error (NMAE), Normalized Root Mean Square Error (NRMSE), Mean Absolute Percentage Error (MAPE), and Coefficient of Determination (R²), which are defined as follows:

[0133]

[0134] Where n represents the total number of test samples, and Let these represent the predicted value and the actual value of the i-th sample, respectively. This represents the mean of the actual values.

[0135] To visually reflect the prediction results, Figure 6 The wind power generation prediction curve obtained by the method of this invention with a one-time-step advance is shown, demonstrating that the method can achieve accurate short-term wind power prediction. To further evaluate the prediction performance of the prediction models, four comparison models were set up at three different wind farms: Persistence, Informer, TCN-BiGRU, and U-Net-BiGRU. These models are all cutting-edge ultra-short-term wind power prediction models. Furthermore, Tables 1-3 show the prediction errors of each model for wind power prediction with a one-time-step advance at the three wind farms. Figure 7 This table compares the error metrics of the model constructed according to this invention with those of other models. (See Tables 1-3 and...) Figure 7 As can be seen, compared with the comparison model, the model constructed based on this invention has smaller prediction errors and better performance.

[0136] Table 1 Comparison of Power Generation Prediction Performance of Wind Farm No. 1

[0137] Model NMAE (%) NRMSE (%) MAPE (%) R2 Persistence 16.11 21.49 26.75 0.823 Informer 14.26 19.12 23.19 0.911 TCN-BiGRU 15.44 19.98 24.54 0.883 U-Net-BiGRU 11.79 16.71 22.36 0.931 This invention 10.95 16.07 20.87 0.965

[0138] Table 2 Comparison of Power Generation Prediction Performance of Wind Farm No. 2

[0139] Model NMAE (%) NRMSE (%) MAPE (%) R2 Persistence 18.69 20.12 27.28 0.811 Informer 15.15 19.15 23.25 0.905 TCN-BiGRU 13.69 18.47 24.44 0.898 U-Net-BiGRU 12.87 16.84 23.87 0.919 This invention 11.13 16.47 21.19 0.963

[0140] Table 3 Comparison of Power Generation Prediction Performance of Wind Farm No. 3

[0141] Model NMAE (%) NRMSE (%) MAPE (%) R2 Persistence 18.11 22.14 27.56 0.817 Informer 14.01 18.87 23.99 0.926 TCN-BiGRU 14.55 20.88 24.73 0.893 U-Net-BiGRU 12.47 15.20 21.60 0.935 This invention 10.42 15.98 20.22 0.969

[0142] In summary, this invention proposes a novel end-to-end wind power prediction framework. By fusing high-resolution wind field simulation with a multi-physics-constrained neural network, it systematically overcomes key problems in existing wind power prediction methods, such as insufficient spatiotemporal modeling, lack of physical consistency, and decoupling of the prediction stages. Its superior performance is verified through experiments. Firstly, this method effectively captures the complex interaction between long-range meteorological evolution and short-term local disturbances in the wind field through a dual-branch dynamic temporal encoder that integrates Fourier neural operators and convolutionally gated recurrent units. Furthermore, a multi-objective learning strategy based on physical information, by introducing physical conservation and spectral constraints, ensures the physical rationality of wind field evolution during model training, providing a high-precision and reliable input foundation for power prediction. Finally, by leveraging a spatially variable gating mechanism and a temporal fusion transformer, end-to-end joint modeling from wind field images to wind power is achieved, significantly alleviating the feature distribution shift problem in traditional staged methods and enhancing the model's stability and generalization ability under abrupt meteorological changes. Experimental results show that this invention outperforms existing benchmark methods in multiple metrics of wind field simulation and power prediction, demonstrating good robustness and forecast consistency under different terrain and weather scenarios. This method can effectively improve the accuracy of ultra-short-term wind power forecasts, providing reliable technical support for grid dispatch, reserve allocation, and market transactions, and possesses certain engineering application value.

[0143] Based on the same technical concept as the method embodiments, the present invention also provides a wind power prediction system that integrates wind field images and multiple physical constraints, including:

[0144] The wind field image data preprocessing module is used to acquire wind field image data and preprocess it through a time window to obtain a four-channel wind field image composed of pixel coordinates and horizontal wind components.

[0145] The wind field spatiotemporal feature extraction module is used to extract features from four-channel wind field images using a dynamic temporal encoder. The dynamic temporal encoder includes two parallel branches: a Fourier neural operator and a convolutional gated recursive unit, which are used to capture the global spectral features and local transient features of the wind field, respectively. The features of each branch are fused by spatially variable gating parameters and learnable channel weights to obtain the fused spatiotemporal features of the wind field.

[0146] The wind field simulation and extrapolation module is used to simulate and generate multi-step future wind field images based on the fused spatiotemporal characteristics of the wind field. The encoder uses a multi-objective physical constraint-driven approach to simulate and generate the images. The multi-objective physical constraint-driven approach refers to the simultaneous introduction of multiple soft constraint terms based on atmospheric physical laws during the simulation process. These include at least the wind field divergence constraint to ensure mass conservation, the vorticity retention constraint to maintain the stability of the circulation structure, and the spectral consistency constraint to ensure that the energy distribution conforms to the turbulence spectrum law. Each constraint is embedded into the loss function through a weighted combination to guide the model to achieve a balance between data fitting and physical rationality.

[0147] The power time series prediction module is used to perform multi-step power prediction on multiple wind farms by combining multi-step future wind farm images obtained from wind farm simulation with historical power data, wind farm geographical location information and meteorological data, and output multi-step future power prediction results through the time series fusion Transformer architecture.

[0148] It should be understood that the wind power prediction system that integrates wind field images and multiple physical constraints in the embodiments of the present invention can realize all the technical solutions in the above method embodiments. The functions of each functional module can be specifically implemented according to the methods in the above method embodiments. The specific implementation process can be referred to the relevant descriptions in the above embodiments, which will not be repeated here.

[0149] The present invention also 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, wherein when the programs are executed by the processors, they implement the steps of the wind power prediction method for fusing wind field images and multiple physical constraints as described above.

[0150] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the wind power prediction method that fuses wind field images and multiple physical constraints as described above.

[0151] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, apparatus (systems), electronic devices, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0152] This invention is described with reference to a flowchart of a method according to embodiments of the invention. It should be understood that each step in the flowchart and combinations thereof can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing device, generate instructions for implementing the process. Figure 1 A device for a function specified in one or more processes.

[0153] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 The function specified in one or more processes.

[0154] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 Steps of a specified function in one or more processes.

Claims

1. A wind power prediction method that integrates wind field images with multiple physical constraints, characterized in that, The method includes the following steps: Wind field image data is acquired and preprocessed using a time window to obtain a four-channel wind field image composed of pixel coordinates and horizontal wind components; Based on the four-channel wind field image, a dynamic temporal encoder is used to extract features. The dynamic temporal encoder includes two parallel branches: a Fourier neural operator and a convolutional gated recursive unit, which are used to capture the global spectral features and local transient features of the wind field, respectively. The features of each branch are fused by spatially variable gating parameters and learnable channel weights to obtain the fused spatiotemporal features of the wind field. Based on the fused spatiotemporal characteristics of the wind field, an encoder is used to simulate and generate multi-step future wind field images through a multi-objective physical constraint-driven approach. Multi-objective physical constraint-driven approach refers to the simultaneous introduction of multiple soft constraint terms based on atmospheric physical laws during the simulation process. These include at least wind field divergence constraints to ensure mass conservation, vorticity retention constraints to maintain circulation structure stability, and spectral consistency constraints to ensure that energy distribution conforms to turbulence spectral laws. Each constraint is embedded into the loss function through a weighted combination to guide the model to achieve a balance between data fitting and physical rationality. Using multi-step future wind field images obtained from wind field simulation, combined with historical power data, wind farm geographical location information and meteorological data, multi-step power prediction is performed on multiple wind farms through a time-series fusion Transformer architecture, and the multi-step future power prediction results are output. The process of feature extraction by the dynamic temporal encoder includes: In the window of history Internally, each wind field image frame is mapped to channel features through a shared backbone network. : ,in, This represents the feature mapping function, where H and W are the height and width of the wind field grid, C is the number of channels after feature mapping, and T is the feature mapping function. w The length to be truncated for the sliding window, where t represents the t-th time step; The Fourier neural operator branch uses two Fourier function mapping modules to act on channel features. The formula is as follows: ,in, This represents the output of global spectral characteristics. This represents the Fourier function mapping module; Convolutional gated recursive units recursively traverse the features of each channel. Output local transient features The formula is as follows: ； in, It is the Sigmoid activation function. For Hadamard products, This is the hidden state from the previous moment. To update the door, control the degree to which its hidden state from the previous moment is retained. To reset the gate, control the degree of influence of the previous hidden state on the current candidate state. , , These represent the convolution operations of the corresponding gates, and each of them has independent learnable parameters; The encoder simulates and generates multi-step future wind field images through a multi-objective physical constraint-driven approach, including: The fused spatiotemporal features of the wind field F are input into the shared decoder. The decoded feature vector is obtained. T w The length of the sliding window is truncated, and τ represents the time step τ; a single-step predicted wind field is generated based on the mapping function. ,in For those with learnable parameters The mapping function is used to convert the decoded feature vector into a four-channel wind field image; then the generated predicted wind field is slidably stitched with the historical wind field sequence as the input for the next iteration step, and multi-step autoregressive inference is performed. During the training phase, the encoder calculates the predicted wind field at step h. Calculate divergence loss vorticity loss Spectrum loss The summation constitutes the physical loss. And together with reconstruction loss and power loss, the model parameters are optimized; among them, divergence loss constitutes the wind field divergence constraint to ensure mass conservation, vorticity loss constitutes the vorticity retention constraint to maintain the stability of the circulation structure, and spectrum loss constitutes the spectrum consistency constraint to ensure that the energy distribution conforms to the turbulent spectrum law. in, This represents the horizontal component of the wind speed predicted by the model. This represents the vertical component of the wind speed predicted by the model. This represents the vorticity value predicted by the model. This represents the actual vorticity value; This represents the spectrum predicted by the model. Represents the actual spectrum, This indicates the difference between predicted vorticity and actual vorticity. This indicates the spectral difference between the predicted wind field and the actual wind field. , , The learnable or preset weight coefficients for each loss term, and h represents the time step index.

2. The method according to claim 1, characterized in that, Preprocessing of wind field image data using time windows specifically includes: Extract a length of T using a sliding window method. w The continuous wind field image frames are processed by normalizing, imputing missing values, aligning spatial coordinates, and stacking channels for each frame to form a four-channel input tensor with a unified format. Where X, Y are the normalized pixel coordinates. Let H and W represent the components of the horizontal wind vector at time τ in the meridional and zonal directions, respectively, and H and W be the height and width of the wind field grid.

3. The method according to claim 1, characterized in that, Features from each branch are fused using spatially variable gating parameters and learnable channel weights, including: ， ， Where G represents the gated unit function, Indicates adaptive weights, These represent global spectral features and local transient features, respectively. It is the Sigmoid activation function. This represents the learnable scaling factor. This represents the output fusion vector, i.e., the fused spatiotemporal features of the wind field.

4. The method according to claim 1, characterized in that, Physical loss Together with reconstruction loss and power loss, the model parameters are optimized, and the total loss function for model optimization is: ， in, These are learnable variance weights used to balance heterogeneous losses; the first term is the reconstruction loss, which is the mean square error between the predicted wind speed and the actual wind speed at step t+1. Indicates predicted wind speed, Indicates the actual wind speed; Indicates physical loss; The power loss is represented by the average of the squared differences between the predicted and actual power at all wind farm sites over all predicted time steps.

5. The method according to claim 2, characterized in that, The temporal fusion Transformer architecture performs multi-step power prediction for multiple wind farms, including: For station i, the grid features are transformed into a sequence through differentiable sampling: ,in, This represents the grid wind field feature vector of station i. Represents the mesh feature mapping function; At each time step, the projected view of the sampled features, historical power, auxiliary measurements, and static metadata are concatenated: ,in, Indicates splicing characteristics, Indicates historical power. Indicates measurement data; This refers to static metadata, specifically the geographical location information of the wind farm, including latitude and longitude coordinates (x, y, y). i , y i and terrain conditions c i ; , , , Indicates learnable weights; Stack them above the window to form a sequence ordered by site: , where d represents the feature dimension of the hidden layer after feature concatenation and mapping; A recursive encoder is used to include historical information: , Indicates a recursive encoder; The timing decoder uses learnable time-domain markings Perform a query The process generates embedding vectors in the specified time domain, and the post-processing head maps these embedding vectors to scalars. , The decoding function combines the encoded context information with a learnable temporal label to generate an embedded representation for a specified time step. This represents the post-processing function used to map the high-dimensional embedding vector output by the decoder to the final scalar power prediction value; After summarizing all sites, a multi-site, multi-step prediction matrix is ​​generated: N is the total number of stations. The length of time in the future.

6. 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, wherein when the programs are executed by the processors, they implement the steps of the wind power prediction method fused with wind field images and multiple physical constraints as described in any one of claims 1-5.

7. A computer-readable storage medium, characterized in that, It stores a computer program, which, when executed by a processor, implements the steps of the wind power prediction method that fuses wind field images and multiple physical constraints as described in any one of claims 1-5.

8. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the wind power prediction method that fuses wind field images and multiple physical constraints as described in any one of claims 1-5.

Images