Regional wind power coordinated prediction method and system for center-oriented platform-wind farm station data fusion
By integrating data between the central platform and wind farms, and utilizing wind vector field simulation and data assimilation, combined with differential privacy algorithms and federated learning, the problem of regional wind power prediction error accumulation was solved, thereby improving the safety, stability, and computational efficiency of the power grid.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- HOHAI UNIV
- Filing Date
- 2025-04-22
- Publication Date
- 2026-07-14
AI Technical Summary
Regional wind power forecasting errors are prone to spatial accumulation, affecting regional power balance and grid security and stability control.
A data integration approach oriented towards the central platform and wind farms is adopted, which combines differential privacy algorithms and federated learning mechanisms. Through wind vector field simulation and data assimilation, gradient calibration and weight updates of the wind power prediction model are achieved, thereby improving prediction accuracy and privacy protection.
It effectively mitigated the accumulation of wind power prediction errors, improved the regional power balance, promoted the economical and safe operation of the power grid, and enhanced computational efficiency and prediction accuracy.
Smart Images
Figure CN120410076B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of renewable energy development and utilization technology, specifically relating to a regional wind power coordination and prediction method and system for data integration between a central platform and wind farms. Background Technology
[0002] With the large-scale integration of wind turbines into regional power grids, the continuous nature of numerical meteorological field errors can lead to a "spatial accumulation" problem in regional wind power output, affecting regional power balance and the safe and stable control of the power grid. Therefore, it is necessary to combine advanced meteorological field simulation technology to improve the accuracy of regional wind power forecasting, reduce grid reserve capacity and auxiliary costs, and ultimately enhance the safe and stable operation of the power grid. Summary of the Invention
[0003] Purpose of the invention: This invention proposes a regional wind power coordinated prediction method and system for data integration between a central platform and wind farms, which effectively alleviates the problem of accumulated regional wind power prediction errors and provides effective data support for improving the regional power balance and promoting the economical and safe operation of the power grid.
[0004] Technical Solution: To achieve the above-mentioned objectives, the present invention proposes a regional wind power coordination and prediction method for data integration between a central platform and wind farms, comprising the following steps:
[0005] S1: Each wind farm uses a local wind speed prediction model for forward propagation, and differential privacy noise is applied to the activation features of the final layer of the local wind speed prediction model to obtain A. i,t Activation feature A i,t and predicted targets Both are reported to the central platform, where d is the time window to be predicted;
[0006] S2: The central platform utilizes the wind vector field from the previous time period and the activation features collected from various wind farms {A} i,t} i∈[1,N] Wind vector field simulation and data assimilation are performed, where N is the total number of wind farms, and the loss function is based on the wind vector field prediction at the central platform. Calculate the wind vector field prediction loss to obtain the gradient information of the activated feature set. The gradient information is then distributed sequentially to the corresponding wind farms; the central platform then updates the weights W of the wind vector field prediction model based on the gradient information obtained from backpropagation. t S ;
[0007] S3: Each wind farm station uses the gradient information dA distributed by the central platform. i,t Gradient calculation based on local wind speed prediction model of wind farm site Then, the gradient information is calibrated based on the differential privacy algorithm to obtain the calibrated gradient information. based on Update the weights of the local wind speed prediction model for wind farms.
[0008] S4: Repeat steps S1-S3 until the wind vector field prediction model of the central platform and the local wind speed prediction model of all wind farms converge; finally, each wind farm trains a wind speed-power conversion model based on the local wind speed prediction results to obtain the multi-step wind power prediction results.
[0009] Furthermore, in step S1, the local wind speed prediction model for the wind farm is as follows:
[0010]
[0011] in, and v i f represents the predicted wind speed and the measured wind speed of the i-th wind farm, respectively. Traf,i Let be the local Transformer prediction model for the i-th wind farm.
[0012] Furthermore, in step S2, the prediction model for the crosswind vector field of the central platform is as follows:
[0013]
[0014] The superscript * indicates that the data is encrypted using a differential privacy protection algorithm. and These are the wind vector field simulation and data assimilation results, respectively, where P represents the word embedding operation; This represents the number of layers from the 1st layer to the nth layer. a Execute Fourier neural operators f sequentially afo,n ; This indicates a data assimilation method.
[0015] Furthermore, in step S2, the wind vector field simulation performed by the central platform specifically includes:
[0016] For the spacetime wind vector field Where c = (x, y) is the spatial index of the wind vector field, and t is the temporal position. Based on prior physical knowledge, the spatiotemporal motion law of the wind vector field can be constructed using the Navier-Stokes equations:
[0017]
[0018] Where γ represents the fluid viscosity coefficient, and f(c) represents the time-invariant scalar matrix. Indicates uncertain motion;
[0019] Introducing stacked Fourier neural operators to achieve iterative solution:
[0020]
[0021] in, To iteratively update the wind vector field feature map of the nth layer, Let w represent the Fourier neural operator model of the nth layer. n and b n f represents the weights and biases of the nth layer; afo,n This represents the Fourier neural operator of the nth layer;
[0022] Utilizing the property that time-domain partial differential equations are equivalent to frequency-domain coefficient products, the single-layer Fourier neural operator is solved using Fast Fourier Transform:
[0023]
[0024] Where D represents the spatial extent of the wind vector field, k φ Let x represent a learnable filter function. k Then it is its index. and This refers to the Fast Fourier Transform and its inverse transform; R φ This represents the integrated filtering function.
[0025] Furthermore, in step S2, the data assimilation performed by the central platform specifically includes:
[0026] Collect wind measurement data and their spatial coordinates from wind farm stations within the region, and crop the spatial neighborhood of the wind vector field simulation results to characterize the local micro-meteorological features of the wind farm stations:
[0027]
[0028] in, The results of the wind vector field simulation are as follows. and , where are the longitude and latitude indices of the i-th wind farm, respectively, and r is the size of the spatial neighborhood;
[0029] Constructing key features using the difference method:
[0030]
[0031] Where (m,n) is the local region S of the i-th wind farm. i The location index; |·| represents the local region S. i The number of pixels in;
[0032] The pixel-level key feature aggregation is achieved using a partitioning aggregation algorithm to obtain the data assimilation results:
[0033]
[0034] st T∈{S i} i∈[1,N] ,(m,n)∈{T1∩...∩T |T|}
[0035] in, and These are the wind vector field simulation and data assimilation results, respectively, f Conv (·) represents a convolutional neural network, and T represents the aggregation space of the corresponding wind vector field pixels (m,n).
[0036] Furthermore, in step S2, the central platform updates the weights of the wind vector field prediction model as follows:
[0037]
[0038] Where WtS is the wind vector field prediction model weight at time t, and η is the model weight update parameter.
[0039] Furthermore, in step S3, the wind farm calibration gradient information is as follows:
[0040]
[0041] Where σ represents the noise scale, and C′ represents the noise modulus applied to the gradient information. This indicates that the expected value is 0 and the variance is σ. 2 The normal distribution of C′I, where I represents a matrix of all 1s.
[0042] Furthermore, in step S3, the wind farm updates the weights of its local wind speed prediction model, as follows:
[0043]
[0044] Where η is the model weight update parameter.
[0045] Furthermore, in step S4, each wind farm station trains a wind speed-power conversion model based on the local wind speed prediction results to obtain multi-step wind power prediction results, as detailed below:
[0046]
[0047] in, and P i f represents the predicted power value and the measured input value of the i-th wind farm, respectively. MLP,i Let i be the local wind speed-power conversion model for the i-th wind farm. This is digit-wise multiplication.
[0048] This invention also provides a regional wind power coordinated prediction system, including a central platform and several wind farms. The central platform deploys a central platform sidewind vector field prediction model, and each wind farm deploys a local wind speed prediction model. The central platform and wind farms repeatedly perform the following operations until the central platform sidewind vector field prediction model and all wind farm local wind speed prediction models converge:
[0049] Each wind farm uses a local wind speed prediction model for forward propagation, and differential privacy noise is applied to the activation features of the final layer of the local wind speed prediction model to obtain A. i,t Activation feature A i,t and predicted targets Both are reported to the central platform, where d is the time window to be predicted;
[0050] The central platform utilizes the wind vector field from the previous time period and the activation features collected from various wind farms {A} i,t} i∈[1,N] Wind vector field simulation and data assimilation are performed, where N is the total number of wind farms, and the loss function is based on the wind vector field prediction at the central platform. Calculate the wind vector field prediction loss to obtain the gradient information of the activated feature set. The gradient information is then distributed sequentially to the corresponding wind farms; the central platform then updates the weights W of the wind vector field prediction model based on the gradient information obtained from backpropagation. t S ;
[0051] Each wind farm station uses gradient information dA distributed by the central platform. i,t Gradient calculation based on local wind speed prediction model of wind farm site Then, the gradient information is calibrated based on the differential privacy algorithm to obtain the calibrated gradient information. according to Update the weights of the local wind speed prediction model for wind farms.
[0052] Each wind farm also deploys a power prediction model, which obtains multi-step wind power prediction results based on local wind speed prediction results and a wind speed-power conversion model.
[0053] Beneficial effects: (1) This invention proposes a regional central platform-wind farm coordinated prediction mechanism. The regional central platform and wind farms respectively implement wind vector field prediction and local power prediction. Through federated separation learning, the data of the central platform and the edge wind farm nodes are integrated, which can improve the accuracy of regional wind power prediction, privacy protection and computational efficiency. (2) This invention uses stacked Fourier neural operators to extract the spatiotemporal dynamic features of wind vector field. It can effectively solve the NS equation through iterative method and give full play to the advantages of high generalization of physical model and strong generalization fitting ability of deep learning. (3) This invention combines the simulation results of wind vector field and the wind speed measurement of regional wind farms to carry out data assimilation, which can efficiently integrate dense wind field data and sparse wind speed measurement data, so that the wind field data inference is more in line with the actual situation and effectively avoids the accumulation of wind field multi-step prediction errors. (4) This invention fits the actual needs of grid dispatch automation and can be applied to regional power dispatch control centers. It can effectively alleviate the problem of regional wind power prediction error accumulation and provide data support for improving the regional power balance level and promoting the economic and safe operation of the power grid. Attached Figure Description
[0054] Figure 1 This is a schematic diagram of the regional wind power coordinated prediction method proposed in this invention;
[0055] Figure 2 This is a schematic diagram of the various factors involved in regional wind power prediction in this invention;
[0056] Figure 3 This is a schematic diagram of the spatiotemporal dynamic feature extraction method for wind vector fields based on Fourier operators proposed in this invention;
[0057] Figure 4 This is a schematic diagram of the multi-source data assimilation method proposed in this invention;
[0058] Figure 5 This is the wind vector field image prediction result in an embodiment of the present invention;
[0059] Figure 6 This refers to the regional wind power prediction results in this embodiment of the invention.
[0060] Figure 7 This is the predicted result of the total regional wind power in this embodiment of the invention. Detailed Implementation
[0061] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.
[0062] Understanding the spatiotemporal characteristics of wind fields is crucial for determining the accuracy of regional wind power prediction. Compared to traditional methods that rely solely on historical wind field spatiotemporal dynamic modeling to capture future wind field characteristics, incorporating wind speed measurement data from wind farms into data assimilation is key to improving the performance of spatiotemporal wind field prediction modeling. Furthermore, the accuracy of wind field simulation can be enhanced by introducing wind vector fields. Therefore, achieving efficient simulation and data assimilation of wind vector fields, coordinating information exchange between the regional central dispatch platform and wind farms, and collaboratively improving prediction accuracy, privacy protection, and computational friendliness are current key focuses and challenges in regional wind power prediction research.
[0063] Based on the above analysis, this invention proposes a regional wind power coordinated prediction method oriented towards data integration between a central platform and wind farms. Drawing on the federated learning mechanism, the regional power grid dispatch center platform (also known as the regional central dispatch platform, hereinafter referred to as the central platform, dispatch center, or dispatch platform) serves as the central server for federated learning, while wind farms distributed across various locations act as clients, also known as edge wind farms. The central platform implements large-scale wind vector field simulation and data assimilation for ubiquitous power measurement; edge wind farms utilize numerical meteorological forecasts issued by the central platform to achieve efficient local power prediction. The overall architecture is as follows: Figure 1 As shown, the participating elements are as follows: Figure 2 As shown. Taking time segment t as the reference position, the regional wind power coordinated prediction mechanism consists of the following three communication steps. In all data communication stages, all transmitted data is encrypted using a differential privacy protection algorithm.
[0064] S1: Each edge wind farm uses a local wind speed prediction model for forward propagation, and differential privacy noise is applied to the activation features of the final layer of the local prediction model to obtain A. i,t Activation feature A i,t and predicted targets They are reported together to the regional dispatch platform, where d is the time window to be predicted.
[0065] S2: The central platform utilizes the wind vector field from the previous time period (time window length p) and the activation features {A} collected from various wind farms. i,t} i∈[1,N] Wind vector field simulation and data assimilation were performed, and a loss function based on the wind vector field prediction at the central platform was used. Calculate the wind vector field prediction loss to obtain the gradient information of the activated feature set. The gradient information is then distributed sequentially to the corresponding wind farms. The central platform then updates the weights of the wind vector field prediction model based on the gradient information obtained from backpropagation.
[0066]
[0067] Among them, Wt S Let η be the weight of the wind vector field prediction model at time t, and η be the model weight update parameter.
[0068] S3: The wind farm station uses gradient information dA distributed by the central platform. i,t Gradient calculation based on local prediction model of wind farm site Then, the gradient information is calibrated based on the differential privacy algorithm to obtain the calibrated gradient information.
[0069]
[0070] Where σ represents the noise scale, and C′ represents the noise modulus applied to the gradient information. This indicates that the expected value is 0 and the variance is σ. 2 The normal distribution of C′I is used, where I represents a matrix of all 1s. Finally, the local model of the wind farm is updated:
[0071]
[0072] Repeat steps S1-S3 until the wind vector field prediction model of the regional dispatch center platform and the local prediction models of all wind farms converge. The wind vector field prediction model of the regional dispatch center platform can be directly used for regional wind vector field prediction. Each local wind farm then trains a wind speed-power conversion model based on the local wind speed prediction results to obtain multi-step wind power prediction results:
[0073]
[0074] in, and P i f represents the predicted power value and the measured input value of the i-th wind farm, respectively. MLP,i Let i be the local wind speed-power conversion model for the i-th wind farm. This is digit-wise multiplication.
[0075] According to an embodiment of the present invention, in step S1, the local wind speed prediction model of the wind farm is... The construction method is as follows:
[0076]
[0077] in, and f represents the predicted wind speed and the measured wind speed for the i-th wind farm, respectively. Traf,i This is the local Transformer prediction model for the i-th wind farm site. Local prediction loss for the wind farm site. It can be designed as:
[0078] In step S2, the wind vector field prediction model (numerical meteorological prediction model) on the central platform side is constructed as follows:
[0079]
[0080] The superscript * indicates that the data is encrypted using a differential privacy protection algorithm. This represents the number of layers from the 1st layer to the nth layer. a Execute Fourier neural operators f sequentially afo,n ; This represents the data assimilation method. The loss function for predicting the crosswind vector field at the central platform. The design is as follows:
[0081]
[0082] in, This represents the wind vector field prediction loss. Indicates data assimilation loss. This represents the Fourier approximation loss. λ1 and λ2 are the coefficients of the data assimilation loss and the Fourier loss, respectively. This indicates the wind speed field.
[0083] In step S2, for wind field simulation, refer to Figure 3 Inputting historical wind vector fields, the central platform utilizes the frequency domain mode separation mechanism of stacked Fourier neural operators to extract the temporal partial differential dynamics of the wind vector fields, obtaining numerical simulation results of the wind vector fields. Furthermore, it employs Fourier approximation loss to constrain the spatiotemporal physical laws of the wind vector fields, specifically including:
[0084] For the spacetime wind vector field Where c = (x, y) is the spatial index of the wind vector field, and t is the temporal position. Based on prior physical knowledge, the spatiotemporal motion law of the wind vector field can be constructed using the Navier-Stokes (NS) equations:
[0085]
[0086] Where γ represents the fluid viscosity coefficient, and f(c) represents the time-invariant scalar matrix. This refers to other uncertainties caused by elevation compensation, energy transfer, etc.
[0087] To address the computational dissipation and nonlinear fitting limitations of directly solving the Navier-Stokes equations, a stacked Fourier neural operator f is introduced. afo,n Implement iterative solution:
[0088]
[0089] in, To iteratively update the wind vector field feature map of the nth layer, Let w represent the Fourier neural operator model of the nth layer. n and b n Let P represent the weights and biases of the nth layer, and let P represent the word embedding operation.
[0090] Since the partial differential dynamics of the wind vector field in the time domain are difficult to extract effectively, the property that the partial differential in the time domain is equivalent to the product of coefficients in the frequency domain can be utilized to solve for the single-layer Fourier neural operator through fast Fourier transform:
[0091] Where D represents the spatial extent of the wind vector field, k φ Let x represent a learnable filter function. k Then it is its index. and This is the Fast Fourier Transform and its inverse transform. R φ This represents the integrated filter function.
[0092] In step S2, for data assimilation, efficient data assimilation is achieved by fusing dense wind vector fields and sparse wind speed measurement correction results through a dense-sparse adaptive convolutional network, referring to... Figure 4 The specific method is as follows:
[0093] Collect wind measurement data and their spatial coordinates from wind farm stations within the region, and crop the spatial neighborhood of the wind vector field simulation results to characterize the local micro-meteorological features of the wind farm stations:
[0094]
[0095] in, The results of the wind vector field simulation are as follows. and ...
[0096] To accurately describe the differences in spatiotemporal phase and magnitude between actual wind speed measurements and wind vector fields at wind farms, key features are constructed using the finite difference method:
[0097]
[0098] Where (m,n) is the local region S of the i-th wind farm. i The location index. |·| represents the local region S. i The number of pixels in the image.
[0099] The pixel-level key feature aggregation is achieved using a partitioning aggregation algorithm to obtain the data assimilation results:
[0100]
[0101] st T∈φ{S i} i∈[1,N] ,(m,n)∈{T1∩...∩T |T|}
[0102] in, and These are the wind vector field simulation and data assimilation results, respectively, f Conv (·) represents a convolutional neural network, and T represents the aggregation space of the corresponding wind vector field pixels (m,n).
[0103] This invention takes into account the needs of data privacy encryption and computational efficiency. It performs gradient feature encryption processing on communication data through a horizontal federation mechanism of separate federated learning. The central platform uses a federated averaging algorithm to process the gradient features reported by each wind farm, thereby realizing efficient data integration and iterative solution between the central platform and the edge wind farm models.
[0104] To verify the performance of the method proposed in this invention, the following experiments were conducted in this embodiment. The test area used in this embodiment is located in Jiangsu Province. Jiangsu Province includes 156 wind farms, including 40 offshore wind farms and 126 onshore wind farms, with a total installed capacity of 22.6 GW. The wind vector field used in this embodiment is the ERA-5 meteorological reanalysis dataset developed by the European Centre for Weather Forecasting, which provides high-precision wind vector field data for a geographic area (23°N-43°N, 110°E-130°E) with a geographic resolution of 0.25° and a temporal resolution of 1 hour. The measured and simulated elevation of the wind field is 1000 hPa (approximately 111.8 m).
[0105] The example uses standard absolute reference error to evaluate wind power prediction results:
[0106]
[0107] in, and P t The predicted and actual wind power values at time t are given respectively. t P is the time step of the test set. c This refers to the installed capacity of wind farms.
[0108] For the ERA-5 meteorological reanalysis dataset and on-site measurement data from wind farms across Jiangsu Province, time-series synchronization matching was performed. Data from June 2022 to December 2023 was selected as the training set (13,140 samples), and 8,760 samples from January 2024 to December 2024 were selected as the test set. This embodiment uses Adaptive Perceptual Graph Convolution (APSC), Heterogeneous Spatiotemporal Graph Convolution (HSGC), and Computational Fluid Dynamics (CFD) algorithms as comparison methods. The regional wind power prediction results obtained using this invention are shown in Table 1. Table 2 shows the computationally friendly performance improvement of the provincial coordinated prediction mechanism obtained by applying this invention. Figure 5 The embodiments demonstrate the multi-step prediction results of wind vector fields by comparing the present invention with the Ensemble Kalman Filter (EnKF). Figure 6 The results of wind power prediction at wind farms throughout Jiangsu Province using this invention are presented. Figure 7 This displays the predicted total wind power for the entire province. (From...) Figure 5 As shown, the wind vector field prediction method of this invention applied to the provincial central dispatch platform exhibits accuracy advantages in multi-step prediction, and can better fit the pixel distribution of the real wind vector field compared to the EnKF algorithm. Figure 6 As shown, this invention (shown as Proposed in the figure) demonstrates accuracy advantages across various terrain regions throughout the province (sea, plains, lake areas, coastal areas, etc.). Figure 7 As shown, this invention also has an accuracy advantage in predicting the total wind power of a province. From the overall fitting situation and the maximum prediction error section, it has an accuracy advantage.
[0109] Table 1. Wind power prediction results of this invention at representative wind farms and across the province.
[0110]
[0111] Table 2 Improved computational friendliness of the provincial wind power coordinated prediction mechanism
[0112]
[0113] In summary, the regional wind power coordination and prediction method provided by this invention, which focuses on efficient data fusion between a central platform and wind farms, is based on a federated separation learning mechanism. It decomposes the wind vector field and local wind power prediction tasks. At the central platform, it extracts the spatiotemporal dynamic features of the wind vector field using stacked Fourier operators. Utilizing wind vector field simulation results and wind speed measurements from regional wind farms, it achieves data assimilation based on dense-sparse adaptive convolution. This scheme can be applied to regional power dispatch and control with a large number of wind farms connected to the grid. It can be deployed at the regional power dispatch and control center and all wind farms in the region, fusing wind vector field model data and farm measurement data to predict regional wind power. It can provide regional wind power prediction results for 1 to 10 hours, thereby guiding the regional power dispatch and control center to adjust the regional power balance plan based on the prediction results, reducing reserve capacity, and promoting grid power supply and renewable energy consumption.
[0114] Another embodiment of the present invention provides a regional wind power coordinated prediction system, including a central platform and several wind farms. The central platform deploys a central platform sidewind vector field prediction model, and each wind farm deploys a local wind speed prediction model. The central platform and the wind farms repeat the following operations until the central platform sidewind vector field prediction model and all wind farm local wind speed prediction models converge:
[0115] Each wind farm uses a local wind speed prediction model for forward propagation, and differential privacy noise is applied to the activation features of the final layer of the local wind speed prediction model to obtain A. i,t Activation feature A i,t and predicted targets Both are reported to the central platform, where d is the time window to be predicted;
[0116] The central platform utilizes the wind vector field from the previous time period and the activation features collected from various wind farms {A} i,t} i∈[1,N] Wind vector field simulation and data assimilation are performed, where N is the total number of wind farms, and the loss function is based on the wind vector field prediction at the central platform. Calculate the wind vector field prediction loss to obtain the gradient information of the activated feature set.
[0117] The gradient information is then distributed sequentially to the corresponding wind farms; the central platform then updates the weights W of the wind vector field prediction model based on the gradient information obtained from backpropagation. t S ;
[0118] Each wind farm station uses gradient information dA distributed by the central platform. i,t Gradient calculation based on local wind speed prediction model of wind farm site Then, the gradient information is calibrated based on the differential privacy algorithm to obtain the calibrated gradient information. according to Update the weights of the local wind speed prediction model for wind farms.
[0119] Each wind farm also deploys a power prediction model, which obtains multi-step wind power prediction results based on local wind speed prediction results and a wind speed-power conversion model.
[0120] The construction and updating methods of the central platform side wind vector field prediction model and the local wind speed prediction model of the wind farm, as well as the specific implementation process of wind vector field simulation and data assimilation on the central platform, can be referred to the relevant descriptions in the above method embodiments, and will not be repeated here.
Claims
1. A regional wind power coordinated prediction method for data fusion between a central platform and wind farms, characterized in that, The method includes the following steps: S1: Each wind farm uses a local wind speed prediction model for forward propagation, and differential privacy noise is applied to the activation features of the final layer of the local wind speed prediction model to obtain... Activation features and predicted targets Both were reported to the central platform, including The time window to be predicted; S2: The central platform utilizes the wind vector field from the previous time period and the activation features collected from various wind farms. Perform wind vector field simulation and data assimilation. Given the total number of wind farms, the loss function is based on the prediction of the side wind vector field of the central platform. Calculate the wind vector field prediction loss to obtain the gradient information of the activated feature set. The gradient information is then distributed sequentially to the corresponding wind farms; the central platform then updates the weights of the wind vector field prediction model based on the gradient information obtained from backpropagation. The specific prediction model for the crosswind vector field of the central platform is as follows: in, For the first The measured wind speeds at each wind farm site are shown in the superscript. This indicates that the data is encrypted using a differential privacy-preserving algorithm. and These are the wind vector field simulation results and the data assimilation results, respectively. Indicates word embedding operation; Indicates from the 1st floor to the 2nd floor. Execute Fourier neural operators sequentially ; Indicates data assimilation methods; The central platform implements wind vector field simulation, specifically including: For the spacetime wind vector field ,in For the spatial location index of the wind vector field, For temporal location, based on prior physical knowledge, the spatiotemporal motion of the wind vector field can be constructed using the Navier-Stokes equations: in, Indicates the fluid viscosity coefficient. Represents a time-invariant scalar matrix. Indicates uncertain motion; Introducing stacked Fourier neural operators to achieve iterative solution: in, To iteratively update the wind vector field feature map of the nth layer, This represents the Fourier neural operator model of the nth layer. and This represents the weights and biases of the nth layer; Indicates the first Fourier neural operators for layers; Utilizing the property that time-domain partial differential equations are equivalent to frequency-domain coefficient products, the single-layer Fourier neural operator is solved using Fast Fourier Transform: in, Indicates the spatial extent of the wind vector field. This represents a learnable filter function. Then it is its index. and This is the Fast Fourier Transform and its inverse transform; Indicates the integrated filter function; S3: Each wind farm station uses the gradient information distributed by the central platform. Gradient calculation based on local wind speed prediction model of wind farm site Then, the gradient information is calibrated based on the differential privacy algorithm to obtain the calibrated gradient information. ;based on Update the weights of the local wind speed prediction model for wind farms. ; S4: Repeat steps S1-S3 until the wind vector field prediction model of the central platform and the local wind speed prediction model of all wind farms converge; finally, each wind farm trains a wind speed-power conversion model based on the local wind speed prediction results to obtain the multi-step wind power prediction results.
2. The method according to claim 1, characterized in that, In step S1, the local wind speed prediction model for the wind farm is as follows: in, and The first Predicted and measured wind speeds at each wind farm site. For the first A local Transformer prediction model for a wind farm.
3. The method according to claim 2, characterized in that, In step S2, the data assimilation performed by the central platform specifically includes: Collect wind measurement data and their spatial coordinates from wind farm stations within the region, and crop the spatial neighborhood of the wind vector field simulation results to characterize the local micro-meteorological features of the wind farm stations: in, The results of the wind vector field simulation are as follows. and The first Longitude and latitude indices of each wind farm station The size of the spatial neighborhood; Constructing key features using the difference method: in, It is the first Local area of a wind farm Location index; This indicates a local area. The number of pixels in; The pixel-level key feature aggregation is achieved using a partitioning aggregation algorithm to obtain the data assimilation results: in, and These are the wind vector field simulation results and the data assimilation results, respectively. It is a convolutional neural network. For the corresponding wind vector field pixels The aggregation space.
4. The method according to claim 3, characterized in that, In step S2, the central platform updates the weights of the wind vector field prediction model as follows: in, for Weights of the wind vector field prediction model at any given time. Update the parameters for the model weights.
5. The method according to claim 4, characterized in that, In step S3, the wind farm calibration gradient information is as follows: in, Indicates the noise level. This represents the noise modulus applied to the gradient information. This indicates that the expected value is 0 and the variance is 0. The normal distribution This represents a matrix consisting entirely of 1s.
6. The method according to claim 5, characterized in that, In step S3, the wind farm updates the weights of its local wind speed prediction model, as follows: in, Update the parameters for the model weights.
7. The method according to claim 6, characterized in that, In step S4, each wind farm station trains a wind speed-power conversion model based on the local wind speed prediction results to obtain multi-step wind power prediction results, as detailed below: in, and The first Predicted power values and measured input values for each wind farm station. For the first Local wind speed-power conversion model for a wind farm. This is digit-wise multiplication.
8. A regional wind power coordinated prediction system, characterized in that, The system comprises a central platform and several wind farms. The central platform deploys a central platform crosswind vector field prediction model, and each wind farm deploys a local wind speed prediction model. The central platform and wind farms repeat the following operations until the central platform crosswind vector field prediction model and all wind farm local wind speed prediction models converge: Each wind farm uses a local wind speed prediction model for forward propagation, and differential privacy noise is applied to the activation features of the final layer of the local wind speed prediction model to obtain... Activation features and predicted targets Both were reported to the central platform, including The time window to be predicted; The central platform utilizes the wind vector field from the previous time period and the activation features collected from various wind farms. Perform wind vector field simulation and data assimilation. Given the total number of wind farms, the loss function is based on the prediction of the side wind vector field of the central platform. Calculate the wind vector field prediction loss to obtain the gradient information of the activated feature set. The gradient information is then distributed sequentially to the corresponding wind farms; the central platform then updates the weights of the wind vector field prediction model based on the gradient information obtained from backpropagation. The specific prediction model for the crosswind vector field of the central platform is as follows: in, For the first The measured wind speeds at each wind farm site are shown in the superscript. This indicates that the data is encrypted using a differential privacy-preserving algorithm. and These are the wind vector field simulation results and the data assimilation results, respectively. Indicates word embedding operation; Indicates from the 1st floor to the 2nd floor. Execute Fourier neural operators sequentially ; Indicates data assimilation methods; The central platform implements wind vector field simulation, specifically including: For the spacetime wind vector field ,in For the spatial location index of the wind vector field, For temporal location, based on prior physical knowledge, the spatiotemporal motion of the wind vector field can be constructed using the Navier-Stokes equations: in, Indicates the fluid viscosity coefficient. Represents a time-invariant scalar matrix. Indicates uncertain motion; Introducing stacked Fourier neural operators to achieve iterative solution: in, To iteratively update the wind vector field feature map of the nth layer, This represents the Fourier neural operator model of the nth layer. and This represents the weights and biases of the nth layer; Indicates the first Fourier neural operators for layers; Utilizing the property that time-domain partial differential equations are equivalent to frequency-domain coefficient products, the single-layer Fourier neural operator is solved using Fast Fourier Transform: in, Indicates the spatial extent of the wind vector field. This represents a learnable filter function. Then it is its index. and This is the Fast Fourier Transform and its inverse transform; Indicates the integrated filter function; Each wind farm station uses gradient information distributed by the central platform. Gradient calculation based on local wind speed prediction model of wind farm site Then, the gradient information is calibrated based on the differential privacy algorithm to obtain the calibrated gradient information. ;according to Update the weights of the local wind speed prediction model for wind farms. ; Each wind farm also deploys a power prediction model, which obtains multi-step wind power prediction results based on local wind speed prediction results and a wind speed-power conversion model.