A multi-stage hybrid physics-data driven wind farm correction method

Abstract

The application discloses a kind of based on multi-stage mixed physics-data driven wind field revision method, it is related to wind field data processing and climate model application technical field. Utilize regional climate simulation tool to execute dynamic downscaling simulation, obtain multidimensional candidate variable field;According to ERA5 wind speed data, target candidate climate variable is screened, and multidimensional key variable field is formed;Hierarchical multi-scale spatio-temporal fusion network for revising wind field data is obtained by constructing the to-be-trained model including multi-scale spatial feature extraction module, time dynamic feature extraction module, multi-scale spatio-temporal feature fusion, output module and training using multidimensional key variable field;With wind field revision field as boundary condition, high-resolution wind field revision field is generated by using complex terrain microscale flow field refining solver.The application not only improves the accuracy of wind speed prediction, but also ensures the physical authenticity and consistency of high-resolution wind field under the influence of complex terrain through physical constraint.

Description

Technical Field

[0001] This invention relates to the field of wind field data processing and climate model application technology, specifically to a wind field correction method based on multi-stage hybrid physics-data driven approach. Background Technology

[0002] High-resolution wind field data can accurately capture the complex structure of local wind fields, significantly improving the accuracy of short-term weather forecasts and extreme event warnings, while also providing key support for wind farm optimization, aviation and maritime safety, and atmospheric pollution diffusion simulation. Currently, the mainstream technologies for obtaining high-resolution wind fields include statistical methods, pure data-driven methods, and pure physical models. However, statistical methods rely on "linear assumptions" or "stationarity assumptions," which cannot capture the nonlinear coupling between atmospheric circulation and the underlying surface under complex terrain. Even relying on climate models that can output future trends, they cannot accurately generate detailed wind fields over long periods due to assumption limitations. Pure data-driven methods can generate high-resolution wind fields by fitting nonlinear relationships through multi-layer networks, but they lack physical constraints and are prone to phenomena that violate atmospheric hydrodynamics, such as unrealistic airflows "penetrating mountains" and sudden changes in wind speed at boundary layer height. Furthermore, they have poor generalization ability to extreme weather, resulting in a significant increase in prediction errors. They also cannot effectively extrapolate the evolution characteristics of wind fields over long periods based on historical data, and cannot provide stable detailed wind field outputs. Pure physical models can accurately simulate the coupling effect between microscale wind fields and terrain, with strong physical consistency, but they have extremely high computational costs and are sensitive to errors in initial boundary conditions. This makes them difficult to apply on a large scale to regional assessments, and they cannot complete high-resolution wind field simulations for long-term future series within an acceptable timeframe, thus limiting their practicality.

[0003] Therefore, there is an urgent need to develop a comprehensive wind field correction scheme that integrates the advantages of multiple technologies to improve the accuracy, reliability, and acquisition efficiency of high-resolution wind field data. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a wind field correction method based on a multi-stage hybrid physics-data driven approach.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A multi-stage hybrid physics-data-driven wind field correction method is proposed, which constructs a hierarchical multi-scale spatiotemporal fusion network for correcting wind field data according to steps S1 to S4, and generates a high-resolution wind field correction field according to step A:

[0007] Step S1: Based on the preset temporal and spatial resolutions, using the global climate dataset as the initial field and lateral boundary conditions, perform dynamic downscaling simulation using regional climate simulation tools to obtain climate data for each target candidate climate variable in the target region within the target time period. Then, the variable fields corresponding to each target candidate climate variable are used to form a multidimensional candidate variable field.

[0008] Step S2: Obtain ERA5 wind speed data for the target area during the target time period, and align the ERA5 wind speed data with the multidimensional candidate variable field in terms of time and spatial resolution; further, based on the ERA5 wind speed data, select a preset number of target candidate climate variables, i.e., each key variable, by spatial display association sorting for the multidimensional candidate variable field, and splice the variable fields of each key variable to form a multidimensional key variable field.

[0009] Step S3: Based on the preset time series length, the multidimensional key variable field is divided into time slices using a sliding time window to form each sub-multidimensional key variable field corresponding to each time period. The sub-multidimensional key variable fields are then standardized to form a multi-channel input dataset. This is further used to form a sample set with the multi-channel input data of the target area for each time period and the ERA5 wind speed data corresponding to the last time step in each time period as samples.

[0010] Step S4: Construct a training model that includes a multi-scale spatial feature extraction module, a temporal dynamic feature extraction module, a multi-scale spatiotemporal feature fusion module, and an output module. Based on the sample set and combined with a preset batch size, use the multi-channel input data of each time period in the sample and the ERA5 wind speed data corresponding to the last time step of each time period as input, and the wind field correction field corresponding to the last time step of each time period as output to train the training model and obtain a hierarchical multi-scale spatiotemporal fusion network.

[0011] Step A: Use steps S1 to S3 to generate input data to be corrected, and input the input data to be corrected into the trained hierarchical multi-scale spatiotemporal fusion network to obtain the corresponding wind field correction field. Then, use the wind field correction field as the boundary condition and use the complex terrain microscale flow field refinement solver to generate a high-resolution wind field correction field.

[0012] Furthermore, the spatial display association sorting described in step S2 specifically includes:

[0013] Step S21: For each spatial grid point in the multidimensional candidate variable field at a preset spatial resolution, acquire the time series data of each target candidate climate variable at that spatial grid point, and simultaneously acquire the corresponding ERA5 wind speed data for that spatial grid point. Calculate the Pearson correlation coefficient and normalized root mean square error for each target candidate climate variable using the following formulas:

[0014] ;

[0015] in, The Pearson correlation coefficient is used. i For time steps, X i For the target candidate climate variable in the first i Time series data at each time step Y i For the first i ERA5 wind speed data at each time step N The total number of time steps. The time series mean of the target candidate climate variable in the spatial grid. This represents the average wind speed data for ERA5 within the spatial grid.

[0016] ;

[0017] in, This is the normalized root mean square error;

[0018] Step S22: Based on the Pearson correlation coefficient and normalized root mean square error of the target candidate climate variables at each spatial grid point, a comprehensive index based on spatial distribution characteristics is generated for each target candidate climate variable; further, for each target candidate climate variable, the corresponding comprehensive score is obtained by normalizing and weighting the comprehensive index; based on the comprehensive scores of each target candidate climate variable, the candidate variables are sorted in descending order to generate a quantitative ranking list.

[0019] Furthermore, the comprehensive indicators of the target candidate climate variables based on spatial distribution characteristics mentioned in step S22 include the area proportion of highly correlated regions, the weighted average correlation index of key areas, the spatial clustering index of highly correlated regions, and the weighted normalized root mean square error index of key areas. Each of these indicators is specifically:

[0020] High correlation area percentage index: Based on the preset high correlation threshold, count the number of spatial grid points whose absolute value of Pearson correlation coefficient is greater than the high correlation threshold, i.e., the number of high correlation grid points, and calculate the percentage of them in the total number of grid points.

[0021] The weighted average correlation index for key areas: The mean of ERA5 wind speed data is used as the spatial weight map, and the absolute value of the Pearson correlation coefficient of each spatial grid point in the spatial grid is weighted and averaged with the spatial weight map.

[0022] Spatial clustering index of highly correlated regions: Using the connected component analysis algorithm, generate the spatially connected clusters formed by each highly correlated grid point, and calculate the proportion of the number of spatial grid points contained in the largest connected cluster to the total number of highly correlated grid points;

[0023] The weighted normalized root mean square error index for the critical region is calculated by weighting the normalized root mean square error of each spatial grid point with the spatial weight map.

[0024] Furthermore, the input end of the multi-scale spatial feature extraction module constitutes the input end of the hierarchical multi-scale spatiotemporal fusion network, and the output end of the multi-scale spatial feature extraction module is connected in series with the time dynamic feature extraction module, the multi-scale spatiotemporal feature fusion module, and the output module. The output end of the output module constitutes the output end of the hierarchical multi-scale spatiotemporal fusion network.

[0025] The multi-scale spatial feature extraction module is used to merge various multi-channel input data along the time dimension based on a preset batch size to form multi-channel reconstructed input data. It then uses residual convolution operations and downsampling operations to gradually generate multi-level spatial features corresponding to different scales. The time dynamic extraction module is used to generate multi-level spatiotemporal features corresponding to different scales. The multi-scale spatiotemporal feature fusion module generates spatiotemporal fusion features at the target scale by fusing multi-level spatial features and multi-level spatiotemporal features and combining upsampling operations. The output module receives the spatiotemporal fusion features and uses convolution operations to generate the wind field correction field corresponding to the last time step of the target time period.

[0026] Furthermore, the multi-scale spatial feature extraction module includes a residual convolution module, a first downsampling module, and a second downsampling module;

[0027] The input of the residual convolution module constitutes the input of the multi-scale spatial feature extraction module, and the output of the residual convolution module is connected in series with the first downsampling module and the second downsampling module. The output of the second downsampling module constitutes the output of the multi-scale spatial feature extraction module.

[0028] The residual convolution module includes two 3x3 convolutional layers and a batch normalization layer, used to map the number of channels of the multi-channel reshaped input data to 32 channels and generate the first-level spatial features; the first downsampling module is used to receive the first-level spatial features and, through a max pooling layer with a stride of 2 and a residual convolution module, increase the number of channels to 64 channels and generate the second-level spatial features; the second-level spatial features are then discarded and input into the second downsampling module, and through a max pooling layer with a stride of 2 and a residual convolution module, increase the number of channels to 128 channels and generate the third-level spatial features.

[0029] Furthermore, the time-dynamic feature extraction module includes a first convolutional long short-term memory network and a second convolutional long short-term memory network, and the input terminals of the first and second convolutional long short-term memory networks together constitute the input terminal of the time-dynamic feature extraction module, and the output terminals of the first and second convolutional long short-term memory networks together constitute the output terminal of the time-dynamic feature extraction module.

[0030] After restoring the temporal series dimensions of the second-level and third-level spatial features using the reshaping operation, the third-level and second-level spatial features are respectively input into the first convolutional long short-term memory network and the second convolutional long short-term memory network, and the third-level spatiotemporal features of the third-level spatial features at the last time step and the second-level spatiotemporal features of the second-level spatial features at the last time step are generated respectively.

[0031] Furthermore, the multi-scale spatiotemporal feature fusion module includes a skip connection module, a first upsampling module, and a second upsampling module.

[0032] The skip connection module extracts the second-level spatial features at the last time step and concatenates them with the second-level spatiotemporal features along the channel dimension to obtain a concatenated tensor with 96 channels. A 1x1 convolutional layer is then used to reduce the number of channels in the concatenated tensor to 64, generating the fused skip connection features. The first upsampling module uses a 2x2 transposed convolution to expand the spatial resolution of the third-level spatiotemporal features and concatenates them with the skip connection features. Further feature fusion is performed using residual convolution to generate the first decoder-level output, which is then concatenated with the first-level spatial features at the last time step to generate the first decoder-level concatenated features. The second upsampling module uses a 2x2 transposed convolution to restore the spatial resolution of the first decoder-level output, restoring the number of channels to 32. This is then concatenated with the first decoder-level concatenated features, and further residual convolution is used to generate the second decoder-level output, i.e., the spatiotemporal fusion features at the target scale.

[0033] Furthermore, the training of the hierarchical multi-scale spatiotemporal fusion network is based on a two-stage training strategy, optimized through a loss function driven by physical perception residuals. This physical perception residual-driven loss includes a mean squared error loss function, a spatial gradient loss function, a correction mean squared error loss function, and a correction spatial gradient loss function.

[0034] ; ;

[0035] ;

[0036] ;

[0037] ;

[0038] in, This is the loss function for the physical perception residual-driven loss, i.e., the total loss function; Let the mean squared error loss function be . Let be the spatial gradient loss function. The mean square error loss function is used to correct the error. The correction space gradient loss function, These are the weighting coefficients of the mean squared error loss function. These are the weighting coefficients of the spatial gradient loss function. The weighting coefficients of the mean square error loss function for the correction are... These are the weighting coefficients of the correction space gradient loss function; i For time steps, N The total number of samples, For the first i Predicted wind speed at each time step For the first i ERA5 wind speed data at each time step; The spatial gradient map of the output of the hierarchical multi-scale spatiotemporal fusion network. This is a spatial gradient map of ERA5 wind speed data. The wind speed at the i-th time step is generated using dynamic downscaling simulation.

[0039] The beneficial effects of adopting the above technical solution are as follows:

[0040] (1) This invention selects key variables through a physical screening mechanism, eliminates dimensional differences by combining Z-score standardization, effectively strengthens physical prior information and eliminates redundant variables, and reduces the numerical error of wind field prediction by learning complex nonlinear error patterns through a deep learning model.

[0041] (2) The present invention is based on a microscale flow field refinement solver for complex terrain. By solving the RANS equations, the wind field is forced to meet the laws of atmospheric hydrodynamics, thus avoiding unrealistic airflow phenomena.

[0042] (3) The present invention adopts a layered strategy of “coarse resolution data-driven correction + fine resolution physical refinement”, which reduces the generation time of regional wind fields while ensuring accuracy. Attached Figure Description

[0043] Figure 1 This is a flowchart of the present invention;

[0044] Figure 2 This invention relates to a hierarchical multi-scale spatiotemporal fusion network;

[0045] Figure 3 This is an example diagram of the microscale flow field refinement data for complex terrain according to the present invention;

[0046] Figure 4 This is a comparison chart of the data before and after correction in this invention with the ERA5 data;

[0047] Figure 5 This is a bar chart comparing the indicators of the model of this invention with those of various comparative models;

[0048] Figure 6 This is a monthly RMSE line graph of the model of this invention and various comparative models. Detailed Implementation

[0049] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings.

[0050] refer to Figure 1 A multi-stage hybrid physics-data-driven wind field correction method is proposed, which constructs a hierarchical multi-scale spatiotemporal fusion network for correcting wind field data according to steps S1 to S4, and generates a high-resolution wind field correction field according to step A:

[0051] Step S1: Set the time resolution to 1 hour and the spatial resolution to 20 km. Using the EC-EARTH global climate model in CMIP6 as the initial field and lateral boundary conditions, perform dynamic downscaling simulation using the RegCM4.7 regional climate model to obtain climate data for each target candidate climate variable in the concentrated wind farm area of ​​Guizhou (23°N-30°N, 103°E-110°E) from 2017 to 2022. Then, construct a multidimensional candidate variable field from the variable fields corresponding to each target candidate climate variable.

[0052] Step S2: Obtain ERA5 wind speed data for the concentrated wind farm area in Guizhou (23°N-30°N, 103°E-110°E) from 2017 to 2022. The obtained ERA5 wind speed data has a spatial resolution of 0.1° (11km) and a temporal resolution of 1 hour. The ERA5 wind speed data and the multidimensional candidate variable field are aligned in terms of temporal and spatial resolution using bilinear interpolation, so that both have a temporal resolution of 1 hour and a spatial grid of 40×40. Further, based on the ERA5 wind speed data, the multidimensional candidate variable field is spatially displayed, correlated, and sorted to select a preset number of target candidate climate variables, i.e., each key variable. The variable fields of each key variable are then spliced ​​together to form a multidimensional key variable field.

[0053] Step S3: Based on the preset time series length, the multidimensional key variable field is divided into time slices using a sliding time window to form each sub-multidimensional key variable field corresponding to each time period. The Z-score standardization method is used to standardize each sub-multidimensional key variable field to form a multi-channel input dataset. The dataset is then divided into a training set, a validation set, and a test set in a ratio of 7:1.5:1.5. This further forms a sample set with the multi-channel input data of the target area for each time period and the ERA5 wind speed data corresponding to the last time step in each time period as samples.

[0054] Step S4: Construct a training model that includes a multi-scale spatial feature extraction module, a temporal dynamic feature extraction module, a multi-scale spatiotemporal feature fusion module, and an output module. Based on the sample set and combined with a preset batch size, use the multi-channel input data of each time period in the sample and the ERA5 wind speed data corresponding to the last time step of each time period as input, and the wind field correction field corresponding to the last time step of each time period as output to train the training model and obtain a hierarchical multi-scale spatiotemporal fusion network.

[0055] Step A: Use steps S1 to S3 to generate input data to be corrected, and input the input data to be corrected into the trained hierarchical multi-scale spatiotemporal fusion network to obtain the corresponding wind field correction field. Then, use the wind field correction field as the boundary condition and use the complex terrain microscale flow field refinement solver to generate a high-resolution wind field correction field.

[0056] Furthermore, the spatial display association sorting described in step S2 specifically includes:

[0057] Step S21: For each spatial grid point in a 40×40 spatial grid of the multidimensional candidate variable field, acquire the time series data of each target candidate climate variable at that spatial grid point, and simultaneously acquire the corresponding ERA5 wind speed data for that spatial grid point. Calculate the Pearson correlation coefficient and normalized root mean square error (RMSE) for each target candidate climate variable using the following formulas: The Pearson correlation coefficient measures the linear correlation between the target candidate climate variable and the true wind speed value, with a value range of [-1, 1]. The closer the absolute value is to 1, the stronger the correlation. The normalized root mean square error is used to eliminate dimensional differences between different target candidate climate variables; the smaller its value, the smaller the numerical deviation between the target candidate climate variable and the true wind speed value.

[0058] ;

[0059] in, The Pearson correlation coefficient is used. i For time steps, X i For the target candidate climate variable in the first iTime series data at each time step Y i For the first i ERA5 wind speed data at each time step N The total number of time steps. The time series mean of the target candidate climate variable in the spatial grid. This represents the average wind speed data for ERA5 within the spatial grid.

[0060] ;

[0061] in, This is the normalized root mean square error;

[0062] Step S22: Based on the Pearson correlation coefficient and normalized root mean square error of the target candidate climate variables at each spatial grid point, a comprehensive index based on spatial distribution characteristics is generated for each target candidate climate variable; further, for each target candidate climate variable, the corresponding comprehensive score is obtained by normalizing and weighting the comprehensive indexes; based on the comprehensive scores of each target candidate climate variable, the candidate variables are sorted in descending order to generate a quantitative ranking list.

[0063] Furthermore, the comprehensive indicators of the target candidate climate variables based on spatial distribution characteristics mentioned in step S22 include the area proportion of highly correlated regions, the weighted average correlation index of key areas, the spatial clustering index of highly correlated regions, and the weighted normalized root mean square error index of key areas. Each of these indicators is specifically:

[0064] Highly correlated area percentage (APHC): Based on a preset high correlation threshold of 0.4, this index counts the number of spatial grid points whose absolute Pearson correlation coefficient is greater than the high correlation threshold, i.e., the number of highly correlated grid points, and calculates their percentage of the total number of grid points. A higher APHC indicates that in most areas of the target region, the target candidate climate variable is more closely correlated with the true wind speed value.

[0065] The Key Area Weighted Average Correlation Index (KWAC) uses the mean of ERA5 wind speed data as the spatial weight map. The absolute values ​​of the Pearson correlation coefficients of each spatial grid point are then weighted and averaged with the spatial weight map. A higher KWAC indicates a stronger predictive ability of the target candidate climate variable for wind speed in high-wind-speed regions.

[0066] Spatial Clustering Index (SCHC): Using a connected component analysis algorithm, this index generates spatially connected clusters formed by each highly correlated grid point, and calculates the proportion of grid points contained in the largest connected cluster to the total number of highly correlated grid points. A higher SCHC indicates that the highly correlated regions are more concentrated on the spatial grid.

[0067] The weighted normalized root mean square error index (KWNRE) for key areas is a weighted average of the normalized root mean square error of each spatial grid point and the spatial weight map. A lower KWNRE index indicates a smaller numerical deviation between the target candidate climate variable and the true wind speed value, resulting in higher prediction accuracy.

[0068] Furthermore, for each candidate climate variable, the comprehensive indicators are normalized according to the following formula:

[0069] In this embodiment, for the three indicators APHC, KWAC, and SCHC, where higher values ​​are preferred, the calculation results of the target candidate climate quantities for each indicator are first obtained, forming a set corresponding to each indicator. Then, global minimum-maximum normalization is applied, as shown in the following formula:

[0070] ;

[0071] in The value in the set, The minimum value in the set. X is the maximum value in the set. norm The normalized index is X. norm ={ }

[0072] For the KWNRE index, which is considered "better with lower values", we first obtain the calculation results of the target candidate climate quantities on this index, form a set, and then use inverse minimum-maximum normalization, as shown in the following formula:

[0073] .

[0074] Furthermore, based on preset weights, the comprehensive score of each target candidate climate variable is calculated by weighted summation according to the following formula:

[0075] .

[0076] Further, refer to Figure 2 ,and Figure 2 Each block represents a feature of a different dimension. The input of the multi-scale spatial feature extraction module constitutes the input of a hierarchical multi-scale spatiotemporal fusion network. The output of the multi-scale spatial feature extraction module is connected in series with the time dynamic feature extraction module, the multi-scale spatiotemporal feature fusion module, and the output module. The output of the output module constitutes the output of the hierarchical multi-scale spatiotemporal fusion network.

[0077] In this embodiment, the preset batch size is 16, the time series length is 25, and the spatial grid is 40*40, i.e., H=40 and W=40. The multi-scale spatial feature extraction module is used to merge various multi-channel input data along the time dimension based on the preset batch size to form multi-channel reconstructed input data. It uses residual convolution operation and downsampling operation to gradually generate multi-level spatial features corresponding to different scales. The time dynamic extraction module is used to generate multi-level spatiotemporal features corresponding to different scales. The multi-scale spatiotemporal feature fusion module generates spatiotemporal fusion features of the target scale by fusing multi-level spatial features and multi-level spatiotemporal features and combining upsampling operation. The output module is used to receive spatiotemporal fusion features and use convolution operation to generate the wind field correction field corresponding to the last time step of the target time period, and the dimension of the wind field correction field is [B,1,40,40].

[0078] Furthermore, the multi-scale spatial feature extraction module includes a residual convolution module, a first downsampling module, and a second downsampling module;

[0079] The input of the residual convolution module constitutes the input of the multi-scale spatial feature extraction module, and the output of the residual convolution module is connected in series with the first downsampling module and the second downsampling module. The output of the second downsampling module constitutes the output of the multi-scale spatial feature extraction module.

[0080] The residual convolution module includes two 3x3 convolutional layers and a batch normalization layer, used to map the number of channels of the multi-channel reshaped input data to 32 channels and generate a first-level spatial feature with a dimension of [400, 32, 40, 40]. The first downsampling module receives the first-level spatial feature and increases the number of channels to 64 channels through a max pooling layer with a stride of 2 and a residual convolution module, generating a second-level spatial feature with a dimension of [400, 64, 20, 20]. The second-level spatial feature is discarded and then input into the second downsampling module, which increases the number of channels to 128 channels through a max pooling layer with a stride of 2 and a residual convolution module, generating a third-level spatial feature with a dimension of [400, 128, 10, 10].

[0081] Furthermore, the time-dynamic feature extraction module includes a first convolutional long short-term memory network and a second convolutional long short-term memory network, and the input terminals of the first and second convolutional long short-term memory networks together constitute the input terminal of the time-dynamic feature extraction module, and the output terminals of the first and second convolutional long short-term memory networks together constitute the output terminal of the time-dynamic feature extraction module.

[0082] After restoring the temporal series dimensions of the second-level and third-level spatial features using the reshaping operation, the dimensions of the second-level spatial features become [16,25,64,20,20] and the dimensions of the third-level spatial features become [16,25,128,10,10]. The third-level and second-level spatial features are then input into the first and second convolutional long short-term memory networks, respectively, to generate the third-level spatiotemporal features of the third-level spatial features at the last time step (dimension [16,64,10,10]) and the second-level spatiotemporal features of the second-level spatial features at the last time step (dimension [16,32,20,20]).

[0083] Furthermore, the multi-scale spatiotemporal feature fusion module includes a skip connection module, a first upsampling module, and a second upsampling module;

[0084] The skip connection module extracts the second-level spatial features at the last time step and concatenates them with the second-level spatiotemporal features along the channel dimension, resulting in a concatenated tensor with 96 channels. A 1x1 convolutional layer is then used to reduce the number of channels in the concatenated tensor to 64, generating the fused skip connection features. The first upsampling module uses a 2x2 transposed convolution to expand the spatial resolution of the third-level spatiotemporal features and concatenates them with the skip connection features. Further feature fusion is performed using residual convolution operations to generate the first decoder layer output, which has dimensions [16, 64, 20, 20]. This output is then combined with the first layer... The spatial features of the first level are concatenated with the first level features (dimension [16,32,40,40]) at the last time step to generate the first decoder level concatenated features, and the dimensions of the first decoder level concatenated features are [16,32,40,40]. The second upsampling module uses 2x2 transposed convolution to restore the spatial resolution of the first decoder level output, that is, the number of channels is restored to 32, and concatenates it with the first decoder level concatenated features. Further residual convolution operation is used to generate the second decoder level output, and the dimensions of the second decoder level output are [16,32,40,40], that is, the spatiotemporal fusion features at the target scale.

[0085] Furthermore, the training of the hierarchical multi-scale spatiotemporal fusion network is based on a two-stage training strategy, optimized through a loss function driven by physical perception residuals. This physical perception residual-driven loss includes a mean squared error loss function, a spatial gradient loss function, a correction mean squared error loss function, and a correction spatial gradient loss function.

[0086] ;

[0087] ;

[0088] ;

[0089] ;

[0090] ;

[0091] in, This is the loss function for the physical perception residual-driven loss, i.e., the total loss function; Let the mean squared error loss function be . Let be the spatial gradient loss function. The mean square error loss function is used to correct the error. The correction space gradient loss function, These are the weighting coefficients of the mean squared error loss function. These are the weighting coefficients of the spatial gradient loss function. The weighting coefficients of the mean square error loss function for the correction are... These are the weighting coefficients of the correction space gradient loss function; i For time steps, N The total number of time steps. For the first i Predicted wind speed at each time step For the first i ERA5 wind speed data at each time step; The spatial gradient map of the output of the hierarchical multi-scale spatiotemporal fusion network. This is a spatial gradient map of ERA5 wind speed data. The wind speed at the i-th time step is generated using dynamic downscaling simulation.

[0092] In this embodiment, each spatial gradient map is generated based on ERA5 wind speed data and wind speed data output from the RegCM4.7 regional climate model. The rate of change of wind speed in geographic space is calculated by convolution using the Sobel operator to quantify the spatial differences in wind speed in different regions. Its core function is to constrain the model prediction results to conform to the laws of atmospheric physics and ensure the physical consistency and accuracy of future high-resolution wind speed data under complex terrain.

[0093] Specifically, the two-stage training strategy adopted in this embodiment is as follows: In the initial training phase (epoch < preset switching period), the first set of weight configurations is used: =1.0, =0, =0, =0. The core objective of this stage is to enable the model to focus on learning the fundamental numerical mapping relationship between low-resolution inputs and high-resolution ground truth, quickly achieve initial convergence of model parameters, and avoid instability in the early stages of training due to complex physical constraints.

[0094] During the refined training phase (epoch ≥ preset switching period), a second set of weight configurations is used: =0.5, =0.5, =0.6, =0.5. At this stage, the model already has basic predictive capabilities. By increasing the weight of the physics-related loss term, the model is forced to pay attention to the spatial details of the predicted wind field (such as wind speed gradient changes caused by terrain), physical texture (such as airflow around characteristics), and accurate fitting of the actual physical deviations, thereby further improving the physical consistency and prediction accuracy of the model.

[0095] Furthermore, step A generates a high-resolution wind field correction field using a complex terrain microscale flow field refinement solver as follows:

[0096] Step A1: Using the blockMesh function in the OpenFOAM fluid dynamics software, an initial hexahedral background mesh covering the entire target area is generated based on the preset vertex coordinates and block division rules of the computational region. The spatial resolution of the initial hexahedral background mesh is set to 100 meters, and the number of meshes is determined according to the computational region and resolution. In this embodiment, it is set to 40×40×30 (x×y×z direction).

[0097] Step A2: Using snappyHexMesh, load the STL file describing the real terrain, and control the adaptive mesh refinement near the STL geometric surface using the refinementSurfaces parameter. The refinement level is configured as level(12) (i.e., the minimum refinement level is 1 and the maximum refinement level is 2), so that the mesh resolution near the terrain surface is gradually increased to about 25 meters. By setting the nSolveIter parameter in snapControls to solve for 30 iterations, the refined mesh nodes are accurately fitted to the STL geometric surface, and finally a body-fitted unstructured mesh that can resolve complex terrain is generated.

[0098] Step A3: Before solving the RANS equations, configure the physical and numerical model parameters using the OpenFOAM dictionary file;

[0099] Specifically, in the turbulenceProperties dictionary, the simulation type is specified as Reynolds-averaged simulation (simulationType RAS), and the standard k-ε turbulence model (suitable for neutral atmospheric boundary layer flow simulation) is selected. Initial values ​​for k (turbulent kinetic energy) and ε (turbulent kinetic energy dissipation rate) and boundary conditions are set. In the transportProperties dictionary, the fluid is defined as a Newtonian fluid, and the kinematic viscosity ν is set to 1.5 × 10⁻⁶. -5m² / s (kinematic viscosity of air under standard atmospheric conditions); The discretization format is specified for different equation terms in the fvSchemes dictionary: gradient terms (gradSchemes) use Gaussian... The linear format ensures the accuracy of spatial gradient calculations. The convection terms (divSchemes) employ a hybrid upwind format (upwind difference format) to balance computational stability and accuracy, with a weight of 0.7 for the convection terms. The Laplacian terms (laplacianSchemes) use a corrected format with non-orthogonal corrections to adapt to the non-orthogonal characteristics of unstructured meshes. The solution algorithm uses the SIMPLE (semi-implicit pressure correlation equation solution) algorithm to handle pressure-velocity coupling in steady-state solutions. In the fvSolution dictionary, the linear solver is configured as follows: GAMG (geometric multigrid) linear solver for pressure p, and smoothSolver for velocity U, turbulent flow k, and ε. The convergence tolerance for pressure p is set to 1e-6, and the convergence tolerance for velocity U, turbulent flow k, and ε is set to 1e-5. Relaxation factors are set to 0.3, 0.7, and 0.7 for pressure field p, velocity equation U, and turbulent equations k and ε, respectively, to ensure the stability of the solution process.

[0100] Step A4: Extract background wind profile information based on the wind field correction field, and set boundary conditions for the boundaries of the computational domain (inlet, outlet, surface, top). The specific boundary conditions are as follows:

[0101] Inlet boundary (upstream): Using the velocity inlet (fixedValue) boundary condition, the processed wind profile data is interpolated and mapped to the grid nodes of the inlet boundary, setting the magnitude and direction of the wind speed;

[0102] Outlet boundary (downstream): Use pressure outlet boundary conditions (e.g., relative pressure with a fixedValue of 0) to ensure smooth outflow.

[0103] Surface boundary: No-slip boundary conditions are used (fixedValue, velocity is set to 0), and it can be used in conjunction with wall functions to reflect the obstruction effect of surface roughness on airflow;

[0104] Top boundary: Use symmetric or slip boundary conditions to reduce the influence of the upper atmospheric boundary on the flow field inside the computational domain.

[0105] Step A5: After setting the boundary conditions, start the steady-state solver of OpenFOAM (such as simpleFoam) to iteratively solve the RANS equations. By monitoring the residual curve and physical quantities (such as wind speed and pressure) at key locations, when the residual drops below the convergence tolerance and the physical quantities no longer change significantly, the solution is considered converged. After convergence, the output is as follows: Figure 3 The high-resolution three-dimensional wind field data shown includes parameters such as wind speed magnitude, wind speed components, pressure, and turbulent kinetic energy.

[0106] The following evaluation and comparison experiment will be conducted using data from 2024 as an example to further explain the technical effects of the present invention in detail.

[0107] Figure 4 This is a comparison chart of the data before and after correction according to the present invention with the ERA5 data. The chart visually demonstrates the spatial prediction capability of the hierarchical multi-scale spatiotemporal fusion network at a time point automatically selected as the optimal correction effect. The chart compares the target true wind speed field (ERA5), the input wind speed field before correction (RegCM), and the output wind speed field after correction side by side. It can be clearly observed from the chart that the wind speed field of RegCM before correction has obvious systematic bias, with its wind speed values ​​generally being too high (showing large areas of bright yellow), and its spatial distribution being too smooth, failing to depict the fine vortex and gradient structure in the real wind field. In contrast, the output wind speed field after correction by the hierarchical multi-scale spatiotemporal fusion network of the present invention is highly consistent with the target true wind speed field in both numerical range and spatial distribution.

[0108] Figure 5 A bar chart comparing the metrics of each model is provided, where the comparison models are attention-series U-shaped network, U-shaped network, convolutional neural network, and naive convolutional network; RMSE is root mean square error, MAE is mean absolute error, and Bias is the bias. The results show that the hierarchical multi-scale spatiotemporal fusion network model proposed in this invention performs best in both root mean square error and mean absolute error metrics, reflecting the absolute error of prediction. Its root mean square error is 0.945 m / s, and its mean absolute error is 0.684 m / s, both the lowest values ​​among all comparison models. This indicates that the hierarchical multi-scale spatiotemporal fusion network has the highest overall prediction accuracy and the smallest deviation between its prediction results and the actual observations. Compared with the classic U-shaped network (root mean square error of 0.971 m / s) and the second-best performing naive convolutional network (root mean square error of 0.963 m / s), the method of this invention has a significant advantage in error control. In terms of the coefficient of determination, which measures the ability of a model to explain the variance of real data, the hierarchical multi-scale spatiotemporal fusion network achieved the highest score of 0.357. This value not only far exceeds all other comparative models, but more importantly, it is significantly higher than the second-best performing Naive Convolutional Network (0.332) and Attention Temporal U-shaped Network (0.337).

[0109] Figure 6The monthly RMSE line graphs for each model in this invention show the monthly fluctuations in the root mean square error performance of each model in 2024. As can be seen from the graph, the errors of all models exhibit seasonal fluctuations, particularly in spring (February-March) and summer (July), when wind field changes are more complex, with peak values ​​appearing for each model's RMSE. However, in most months, especially autumn and winter (August-December), the performance curve (green line) of the hierarchical multi-scale spatiotemporal fusion network of this invention consistently ranks below or at the best position among all curves.

[0110] The above description is merely a preferred embodiment of the present invention and does not constitute any limitation on the present invention. Any simple modifications, alterations, or equivalent structural changes made to the above embodiments based on the technical essence of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A wind field correction method based on multi-stage hybrid physics-data driven approach, characterized in that, Construct a hierarchical multi-scale spatiotemporal fusion network for correcting wind field data according to steps S1 to S4 below, and generate a high-resolution wind field correction field according to step A: Step S1: Based on the preset temporal and spatial resolutions, using the global climate dataset as the initial field and lateral boundary conditions, perform dynamic downscaling simulation using regional climate simulation tools to obtain climate data for each target candidate climate variable in the target region within the target time period. Then, the variable fields corresponding to each target candidate climate variable are used to form a multidimensional candidate variable field. Step S2: Obtain ERA5 wind speed data for the target area during the target time period, and align the ERA5 wind speed data with the multidimensional candidate variable field in terms of time and spatial resolution; further, based on the ERA5 wind speed data, select a preset number of target candidate climate variables, i.e., each key variable, by spatial display association sorting for the multidimensional candidate variable field, and splice the variable fields of each key variable to form a multidimensional key variable field. Step S3: Based on the preset time series length, the multidimensional key variable field is divided into time slices using a sliding time window to form each sub-multidimensional key variable field corresponding to each time period. The sub-multidimensional key variable fields are then standardized to form a multi-channel input dataset. This is further used to form a sample set with the multi-channel input data of the target area for each time period and the ERA5 wind speed data corresponding to the last time step in each time period as samples. Step S4: Construct a training model that includes a multi-scale spatial feature extraction module, a temporal dynamic feature extraction module, a multi-scale spatiotemporal feature fusion module, and an output module. Based on the sample set and combined with a preset batch size, use the multi-channel input data of each time period in the sample and the ERA5 wind speed data corresponding to the last time step of each time period as input, and the wind field correction field corresponding to the last time step of each time period as output to train the training model and obtain a hierarchical multi-scale spatiotemporal fusion network. Step A: Use steps S1 to S3 to generate input data to be corrected, and input the input data to be corrected into the trained hierarchical multi-scale spatiotemporal fusion network to obtain the corresponding wind field correction field. Then, use the wind field correction field as the boundary condition and use the complex terrain microscale flow field refinement solver to generate a high-resolution wind field correction field.

2. The wind field correction method based on multi-stage hybrid physics-data driven according to claim 1, characterized in that, The spatial display association sorting mentioned in step S2 specifically refers to: Step S21: For each spatial grid point in the multidimensional candidate variable field at a preset spatial resolution, acquire the time series data of each target candidate climate variable at that spatial grid point, and simultaneously acquire the corresponding ERA5 wind speed data for that spatial grid point. Calculate the Pearson correlation coefficient and normalized root mean square error for each target candidate climate variable using the following formulas: ； in, The Pearson correlation coefficient is used. i For time steps, X i For the target candidate climate variable in the first i Time series data at each time step Y i For the first i ERA5 wind speed data at each time step N The total number of time steps. The time series mean of the target candidate climate variable in the spatial grid. This represents the average wind speed data for ERA5 within the spatial grid. ； in, This is the normalized root mean square error; Step S22: Based on the Pearson correlation coefficient and normalized root mean square error of the target candidate climate variables at each spatial grid point, a comprehensive index based on spatial distribution characteristics is generated for each target candidate climate variable; further, for each target candidate climate variable, the corresponding comprehensive score is obtained by normalizing and weighting the comprehensive index; based on the comprehensive scores of each target candidate climate variable, the candidate variables are sorted in descending order to generate a quantitative ranking list.

3. The wind field correction method based on multi-stage hybrid physics-data driven according to claim 2, characterized in that, The comprehensive indicators of the target candidate climate variables based on spatial distribution characteristics mentioned in step S22 include the area proportion of highly correlated regions, the weighted average correlation index of key areas, the spatial clustering index of highly correlated regions, and the weighted normalized root mean square error index of key areas. Each of these indicators is specifically: High correlation area percentage index: Based on the preset high correlation threshold, count the number of spatial grid points whose absolute value of Pearson correlation coefficient is greater than the high correlation threshold, i.e., the number of high correlation grid points, and calculate the percentage of them in the total number of grid points. The weighted average correlation index for key areas: The mean of ERA5 wind speed data is used as the spatial weight map, and the absolute value of the Pearson correlation coefficient of each spatial grid point in the spatial grid is weighted and averaged with the spatial weight map. Spatial clustering index of highly correlated regions: Using the connected component analysis algorithm, generate the spatially connected clusters formed by each highly correlated grid point, and calculate the proportion of the number of spatial grid points contained in the largest connected cluster to the total number of highly correlated grid points; The weighted normalized root mean square error index for the critical region is calculated by weighting the normalized root mean square error of each spatial grid point with the spatial weight map.

4. The wind field correction method based on multi-stage hybrid physics-data driven according to claim 1, characterized in that, The input end of the multi-scale spatial feature extraction module constitutes the input end of the hierarchical multi-scale spatiotemporal fusion network. The output end of the multi-scale spatial feature extraction module is connected in series with the time dynamic feature extraction module, the multi-scale spatiotemporal feature fusion module, and the output module. The output end of the output module constitutes the output end of the hierarchical multi-scale spatiotemporal fusion network. The multi-scale spatial feature extraction module is used to merge various multi-channel input data along the time dimension based on a preset batch size to form multi-channel reconstructed input data, and to gradually generate multi-level spatial features corresponding to different scales using residual convolution operations and downsampling operations. The time dynamic extraction module is used to generate multi-level spatiotemporal features corresponding to different scales; the multi-scale spatiotemporal feature fusion module generates spatiotemporal fusion features at the target scale by fusing multi-level spatial features and multi-level spatiotemporal features and combining upsampling operations. The output module is used to receive spatiotemporal fusion features and use convolution operations to generate the wind field correction field corresponding to the last time step of the target time period.

5. The wind field correction method based on multi-stage hybrid physics-data driven according to claim 4, characterized in that, The multi-scale spatial feature extraction module includes a residual convolution module, a first downsampling module, and a second downsampling module; The input of the residual convolution module constitutes the input of the multi-scale spatial feature extraction module, and the output of the residual convolution module is connected in series with the first downsampling module and the second downsampling module. The output of the second downsampling module constitutes the output of the multi-scale spatial feature extraction module. The residual convolution module includes two 3x3 convolutional layers and a batch normalization layer, which are used to map the number of channels of the multi-channel reshaped input data to 32 channels and generate the first-level spatial features. The first downsampling module is used to receive the first-level spatial features, and through a max pooling layer with a stride of 2 and a residual convolution module, the number of channels is increased to 64 channels, and the second-level spatial features are generated. The second-level spatial features are discarded and then input into the second downsampling module. The number of channels is increased to 128 through a max pooling layer with a stride of 2 and a residual convolution module, and the third-level spatial features are generated.

6. The wind field correction method based on multi-stage hybrid physics-data driven according to claim 5, characterized in that, The time-dynamic feature extraction module includes a first convolutional long short-term memory network and a second convolutional long short-term memory network. The input terminals of the first and second convolutional long short-term memory networks together constitute the input terminal of the time-dynamic feature extraction module, and the output terminals of the first and second convolutional long short-term memory networks together constitute the output terminal of the time-dynamic feature extraction module. After restoring the temporal series dimensions of the second-level and third-level spatial features using the reshaping operation, the third-level and second-level spatial features are respectively input into the first convolutional long short-term memory network and the second convolutional long short-term memory network, and the third-level spatiotemporal features of the third-level spatial features at the last time step and the second-level spatiotemporal features of the second-level spatial features at the last time step are generated respectively.

7. The wind field correction method based on multi-stage hybrid physics-data driven according to claim 6, characterized in that, The multi-scale spatiotemporal feature fusion module includes a skip connection module, a first upsampling module, and a second upsampling module. The skip connection module extracts the second-level spatial features at the last time step and concatenates them with the second-level spatiotemporal features in the channel dimension to obtain a concatenated tensor with 96 channels. Then, a 1x1 convolutional layer is used to reduce the number of channels of the concatenated tensor to 64, generating the fused skip connection features. The first upsampling module uses a 2x2 transposed convolution to expand the spatial resolution of the third-level spatiotemporal features and concatenates them with the skip connection features. Further feature fusion is performed using residual convolution to generate the first decoder level output, which is then concatenated with the first-level spatial features at the last time step to generate the first decoder level concatenated features. The second upsampling module uses 2x2 transposed convolution to restore the spatial resolution of the first decoder layer output, that is, the number of channels is restored to 32, and it is concatenated with the first decoder layer splicing features. Then, residual convolution operation is used to generate the second decoder layer output, that is, the spatiotemporal fusion features at the target scale.

8. The wind field correction method based on multi-stage hybrid physics-data driven according to claim 7, characterized in that, The training of the hierarchical multi-scale spatiotemporal fusion network is based on a two-stage training strategy, which optimizes the loss function through the physical perception residual driven loss. The physical perception residual driven loss includes the mean squared error loss function, the spatial gradient loss function, the correction mean squared error loss function, and the correction spatial gradient loss function. ； ； ； ； ； in, This is the loss function for the physical perception residual-driven loss, i.e., the total loss function; Let the mean squared error loss function be . Let be the spatial gradient loss function. The mean square error loss function is used to correct the error. The correction space gradient loss function, These are the weighting coefficients of the mean squared error loss function. These are the weighting coefficients of the spatial gradient loss function. The weighting coefficients of the mean square error loss function for the correction are... These are the weighting coefficients of the correction space gradient loss function; i For time steps, N The total number of samples, For the first i Predicted wind speed at each time step For the first i ERA5 wind speed data at each time step; The spatial gradient map of the output of the hierarchical multi-scale spatiotemporal fusion network. This is a spatial gradient map of ERA5 wind speed data. To generate the first using dynamic downscaling simulation i Wind speed at each time step.

Images