Regional wind power coordination prediction method and system for central platform-wind power station data fusion

Abstract

The invention discloses a central platform-wind power station data fusion-oriented regional wind power coordination prediction method and system. According to the method, on the basis of a Fourier neural operator, a physical space-time law and other uncertain motion laws contained in a wind vector field are extracted, and Fourier approximation loss is used for constraint; for space-time wind vector field simulation results and regional wind power plant station power measurement, data assimilation is carried out through a dense-sparse adaptive convolutional network, and wind field simulation results are corrected by using measurement truth value data; data fusion requirements of a regional center platform and an edge wind power station are considered, safe and rapid solution of center side wind field simulation and edge side local power prediction is realized by using separation federal learning, and regional wind power prediction accuracy, calculation friendliness and data privacy protection capability are cooperatively improved. The method can be applied to a regional power dispatching regulation and control center platform for grid connection of a large number of wind power stations, and regional wind power prediction results can be provided for 1-10 hours.

Description

Technical Field

[0001] The present invention belongs to the technical field of development and utilization of renewable energy, and particularly relates to a regional wind power coordinated prediction method and system for data integration between a central platform and a wind farm station. Background Art

[0002] With the large-scale access of wind turbines to the regional power grid, due to the continuous characteristics of numerical meteorological field errors, the "spatial accumulation" problem is prone to occur in regional wind power errors, which in turn affects the 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 regional wind power prediction level, reduce the grid reserve capacity and auxiliary costs, and then improve the safe and stable operation level of the power grid. Summary of the Invention

[0003] Object of the Invention: The present invention proposes a regional wind power coordinated prediction method and system for data integration between a central platform and a wind farm station, which effectively alleviates the problem of cumulative regional wind power prediction errors and provides effective data support for improving the regional power balance level and promoting the economic and safe operation of the power grid.

[0004] Technical Solution: To achieve the above object of the invention, the regional wind power coordinated prediction method for data integration between a central platform and a wind farm station proposed by the present invention includes the following steps:

[0005] S1: Each wind farm station performs forward propagation using a local wind speed prediction model, and applies differential privacy noise to the activation features of the final layer of the local wind speed prediction model to obtain A i,t , and the activation feature A i,t and the prediction target are reported to the central platform together, where d is the time window to be predicted;

[0006] S2: The central platform uses the wind vector field of the previous time period and the activation features {A i,t} i∈[1,N] collected from each wind farm station to perform wind vector field simulation and data assimilation. N is the total number of wind farm stations. Based on the loss function of wind vector field prediction on the central platform side, calculate the wind vector field prediction loss to obtain the gradient information of the activation feature set and distribute the gradient information to the corresponding wind farm stations in turn; the central platform then updates the weight W t S of the wind vector field prediction model according to the gradient information obtained by backpropagation;

[0007] S3: Each wind farm station calculates the gradient i,t according to the gradient information dA distributed by the central platform, in combination with the local wind speed prediction model of the wind farm station.Furthermore, calibrate the gradient information based on the differential privacy algorithm to obtain the calibrated gradient information Based on Update the weights of the local wind speed prediction model of the wind farm

[0008] S4: Repeat the above steps S1 - S3 until the wind vector field prediction model of the central platform and all local wind speed prediction models of the wind farms converge; finally, each wind farm trains a wind speed - power conversion model according to 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 of the wind farm is specifically as follows:

[0010]

[0011] Among them, and v i are the predicted wind speed and the measured wind speed of the i - th wind farm respectively, and f Traf,i is the local Transformer prediction model of the i - th wind farm.

[0012] Furthermore, in step S2, the wind vector field prediction model on the central platform side is specifically as follows:

[0013]

[0014] Among them, the superscript * indicates that the data is encrypted by the differential privacy protection algorithm, and are the wind vector field simulation and data assimilation results respectively, and P represents the word embedding operation; represents performing the Fourier neural operator f a sequentially from the first layer to the n afo,n ; represents the data assimilation method.

[0015] Furthermore, in step S2, the implementation of the wind vector field simulation on the central platform specifically includes:

[0016] For the spatio - temporal wind vector field where c=(x, y) is the spatial position index of the wind vector field, and t is the temporal position. Based on the physical prior knowledge, the spatio - temporal motion law of the wind vector field can be constructed by the Navier - Stokes equation:

[0017]

[0018] Among them, γ represents the fluid viscosity coefficient, and f(c) represents the non - time - varying scalar matrix, represents the uncertain motion;

[0019] Introduce the stacked Fourier neural operator to achieve iterative solution:

[0020]

[0021] Among them, is the characteristic map of the wind vector field for iterative updating of the nth layer, represents the Fourier neural operator model of the nth layer, w n and b n represent the weights and biases of the nth layer; f afo,n represents the Fourier neural operator of the nth layer;

[0022] Utilize the property that the time-domain partial derivative is equivalent to the product of frequency-domain coefficients, and solve the single-layer Fourier neural operator through the fast Fourier transform:

[0023]

[0024] Among them, D represents the spatial range of the wind vector field, k φ represents the learnable filtering function, x k is its index, and are the fast Fourier transform and its inverse transform; R φ represents the integrated filtering function.

[0025] Furthermore, in step S2, the data assimilation by the central platform specifically includes:

[0026] Collect the measured wind data and its spatial coordinates of the wind measurement towers in the wind farm within the regional range, and crop the spatial neighborhood of the wind vector field simulation result as the characterization of the local micro-meteorological characteristics of the wind farm:

[0027]

[0028] Among them, is the wind vector field simulation result, and are the longitude index and latitude index of the i-th wind farm respectively, and r is the size of the spatial neighborhood;

[0029] Use the difference method to construct key features:

[0030]

[0031] Among them, (m,n) is the position index of the local area S of the i-th wind farm i ; |·| represents the number of pixels in the local area S i ;

[0032] Use the partition aggregation algorithm to achieve pixel-level key feature aggregation and obtain the data assimilation result:

[0033]

[0034] s.t. T∈{S i} i∈[1,N] ,(m,n)∈{T1∩...∩T |T|}

[0035] Wherein, and are the wind vector field simulation and data assimilation results respectively, and f Conv (·) is a convolutional neural network, and T is the aggregation space corresponding to the wind vector field pixel (m,n).

[0036] Furthermore, in step S2, the central platform updates the weights of the wind vector field prediction model as follows:

[0037]

[0038] Wherein, WtS is the weight of the wind vector field prediction model at time t, and η is the model weight update parameter.

[0039] Furthermore, in step S3, the wind farm calibrates the gradient information as follows:

[0040]

[0041] Wherein, σ represents the noise scale, C′ represents the noise modulus applied to the gradient information, represents a normal distribution with an expectation of 0 and a variance of σ 2 C′I, and I represents the all-ones matrix.

[0042] Furthermore, in step S3, the wind farm updates the weights of the local wind speed prediction model as follows:

[0043]

[0044] Wherein, η is the model weight update parameter.

[0045] Furthermore, in step S4, 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 as follows:

[0046]

[0047] Wherein, and P i are the power prediction value and the input measured value of the i-th wind farm respectively, and f MLP,i is the wind speed-power conversion model of the i-th wind farm, is the bitwise multiplication.

[0048] The present invention also provides a regional wind power coordinated prediction system, including a central platform and several wind farm stations. The central platform deploys a wind vector field prediction model on the central platform side, and each wind farm station deploys a local wind speed prediction model. The central platform and the wind farm stations repeatedly perform the following operations until the wind vector field prediction model on the central platform side and the local wind speed prediction models of all wind farm stations converge:

[0049] Each wind farm station uses the local wind speed prediction model for forward propagation, and applies differential privacy noise to the activation features of the final layer of the local wind speed prediction model to obtain A i,t , and the activation feature A i,t and the prediction target are reported to the central platform together, where d is the time window to be predicted;

[0050] The central platform uses the wind vector field of the previous time period and the activation features {A i,t} i∈[1,N] collected from each wind farm station to perform wind vector field simulation and data assimilation. N is the total number of wind farm stations. Based on the loss function of the wind vector field prediction on the central platform side calculate the wind vector field prediction loss to obtain the gradient information of the activation feature set and distribute the gradient information to the corresponding wind farm stations in turn; the central platform then updates the weight W of the wind vector field prediction model according to the gradient information obtained by backpropagation t S ;

[0051] Each wind farm station calculates the gradient according to the gradient information dA i,t distributed by the central platform, in combination with the local wind speed prediction model of the wind farm station and then calibrates the gradient information based on the differential privacy algorithm to obtain the calibrated gradient information According to update the weight of the local wind speed prediction model of the wind farm station

[0052] Each wind farm station also deploys a power prediction model, and the power prediction model obtains the multi-step wind power prediction result based on the local wind speed prediction result and the wind speed-power conversion model.

[0053] Beneficial effects: (1) The present invention proposes a coordinated prediction mechanism for a regional central platform and a wind farm station. The regional central platform and the wind farm station respectively implement wind vector field prediction and local power prediction, and realize data integration between the central platform and the edge wind farm station nodes through federated separation learning, thereby synergistically improving the regional wind power prediction accuracy, privacy protection, and computing efficiency; (2) The present invention uses a stacked Fourier neural operator to extract the spatio-temporal dynamic features of the wind vector field, and can effectively solve the N-S equation through an iterative method, synergistically leveraging the high generalization ability of the physical model and the strong generalization fitting ability of deep learning; (3) The present invention combines the wind vector field simulation results and the wind speed measurements of the regional wind farm stations to carry out data assimilation, efficiently integrating dense wind field data and sparse wind speed measurement data, making the wind field data derivation more in line with the actual situation, and effectively avoiding the accumulation of multi-step prediction errors in the wind field; (4) The present invention meets the actual needs of power grid dispatching automation, can be applied to the regional power dispatching control center, effectively alleviates the problem of cumulative regional wind power prediction errors, and provides data support for improving the regional power balance level and promoting the economic and safe operation of the power grid. Brief Description of the Drawings

[0054] Figure 1 It is a schematic diagram of the regional wind power coordinated prediction method proposed by the present invention;

[0055] Figure 2 It is a schematic diagram of each participating element in the regional wind power prediction of the present invention;

[0056] Figure 3 It is a schematic diagram of the method for extracting spatio-temporal dynamic features of the wind vector field based on the Fourier operator proposed by the present invention;

[0057] Figure 4 It is a schematic diagram of the multi-source data assimilation method proposed by the present invention;

[0058] Figure 5 It is the predicted result of the wind vector field image in the embodiment of the present invention;

[0059] Figure 6 It is the predicted result of the regional wind power in the embodiment of the present invention;

[0060] Figure 7 It is the predicted result of the total regional wind power in the embodiment of the present invention. Detailed Embodiments

[0061] The technical solutions of the present invention will be further described below with reference to the drawings and embodiments.

[0062] How to grasp the spatio-temporal derivation characteristics of the wind field is the core factor determining the regional wind power prediction level. Compared with traditional methods that only rely on historical wind field spatio-temporal dynamic modeling to capture future wind field derivation characteristics, how to introduce the data of wind speed measurement devices in wind farms to participate in data assimilation is the key to improving the spatio-temporal derivation performance modeling of the wind field; in addition, the accuracy of wind field simulation can be improved by introducing the wind vector field. Therefore, realizing the efficient simulation of the wind vector field and data assimilation, coordinating the regional central dispatching platform and wind farms to achieve information interconnection, and jointly improving prediction accuracy, privacy protection, and computational friendliness are the key points and difficulties in the current research on regional wind power prediction.

[0063] Based on the above analysis, the present invention proposes a regional wind power coordinated prediction method for data integration between the central platform and wind farms. Drawing on the federated learning mechanism, the regional power grid dispatching center platform (also known as the regional central dispatching platform, hereinafter referred to as the central platform, dispatching center, or dispatching platform) is used as the federated learning central server, and wind farms distributed at various locations are used as federated learning clients, also known as edge wind farms. The central platform implements large-scale wind vector field simulation and data assimilation for ubiquitous power measurement; the edge wind farms use the numerical weather prediction issued by the central platform to achieve local efficient power prediction. The overall architecture is as Figure 1 shown, and each participating element is as Figure 2 shown. Taking the time section t as the reference position, the regional wind power coordinated prediction mechanism is divided into the following 3 communication steps. In all data communication links, all transmitted data is encrypted by the differential privacy protection algorithm.

[0064] S1: Each edge wind farm uses the local wind speed prediction model for forward propagation, and applies differential privacy noise to the activation features of the final layer of the local prediction model to obtain A i,t , and reports the activation feature A i,t and the prediction target to the regional dispatching platform together, where d is the time window to be predicted.

[0065] S2: The central platform uses the wind vector field in the previous time period (the time window length is p) and the activation features {A i,t} i∈[1,N] collected from each wind farm to implement wind vector field simulation and data assimilation. Based on the loss function of the wind vector field prediction on the central platform side, calculate the wind vector field prediction loss, obtain the gradient information of the activation feature set, and distribute the gradient information to the corresponding wind farms in turn. The central platform then updates the weights of the wind vector field prediction model according to the gradient information obtained by backpropagation:

[0066]

[0067] Among them, Wt S is the weight of the wind vector field prediction model at time t, and η is the model weight update parameter.

[0068] S3: The wind farm station calculates the gradient according to the gradient information dA distributed by the central platform i,t , combines it with the local prediction model of the wind farm station to calculate the gradient Furthermore, 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, C′ represents the noise modulus applied to the gradient information, represents a normal distribution with an expectation of 0 and a variance of σ 2 C′I, and I represents the all-ones matrix. Finally, update the local model of the wind farm station:

[0071]

[0072] Repeat the above steps S1 - S3 until the wind vector field prediction models of the regional dispatch center platform and all local wind farm station prediction models 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 station then trains a wind speed - power conversion model based on the local wind speed prediction results to obtain the multi-step wind power prediction results:

[0073]

[0074] where, and P i are the power prediction value and the input measured value of the i-th wind farm station respectively, and f MLP,i is the local wind speed - power conversion model of the i-th wind farm station, is the bitwise multiplication.

[0075] According to the embodiment of the present invention, in step S1, the local wind speed prediction model of the wind farm station is constructed as follows:

[0076]

[0077] where, and are the predicted wind speed and the measured wind speed of the i-th wind farm station respectively. f Traf,i is the local Transformer prediction model of the i-th wind farm station. The local prediction loss of the wind farm station can be designed as:

[0078] In step S2, the wind vector field prediction model (numerical weather prediction model) on the central platform side is constructed as follows:

[0079]

[0080] Where the superscript * indicates that the data is encrypted by the differential privacy protection algorithm. Denotes from the first layer to the nth a Sequentially execute the Fourier neural operator f afo,n ; Denotes the data assimilation method. The loss function for wind vector field prediction on the central platform side Is designed as follows:

[0081]

[0082] Where Denotes the wind vector field prediction loss, Denotes the data assimilation loss, Denotes the Fourier approximation loss. λ1 and λ2 are the coefficients of the data assimilation loss and the Fourier loss respectively. Denotes the wind speed field.

[0083] In step S2, for wind field simulation, referring to Figure 3 , input the historical wind vector field. The central platform uses the frequency domain mode separation mechanism of the stacked Fourier neural operator to extract the temporal partial differential dynamic law of the wind vector field, obtain the numerical simulation result of the wind vector field, and use the Fourier approximation loss to constrain the spatio-temporal physical law of the wind vector field, specifically including:

[0084] For the spatio-temporal wind vector field Where c = (x, y) is the spatial position index of the wind vector field, and t is the temporal position. Based on physical prior knowledge, the spatio-temporal motion law of the wind vector field can be constructed by the Navier-Stokes (N-S) equation:

[0085]

[0086] Where γ represents the fluid viscosity coefficient, f(c) represents a non-time-varying scalar matrix, Denotes other uncertain motions caused by elevation compensation, energy transfer, etc.

[0087] Considering the computational dissipation and non-linear fitting ability brought by directly solving the N-S equation, introduce the stacked Fourier neural operator f afo,n To achieve iterative solution:

[0088]

[0089] Where To iteratively update the wind vector field feature map of the nth layer, represents the Fourier neural operator model of the nth layer, w n and b n represents the weight and bias of the nth layer, and P represents the word embedding operation.

[0090] Since the partial differential dynamic law of the wind vector field in the time domain is difficult to extract effectively, the characteristic that the time domain partial differential is equivalent to the product of the frequency domain coefficients can be used to solve the single-layer Fourier neural operator through fast Fourier transform:

[0091] Where D represents the spatial range of the wind vector field, k φ represents a learnable filter function, x k is its index. and is the fast Fourier transform and its inverse transform. φ represents the integrated filter function.

[0092] In step S2, for data assimilation, the dense wind vector field and the sparse wind speed measurement correction results are fused through a dense-sparse adaptive convolutional network to achieve efficient data assimilation. Figure 4 , the specific method is as follows:

[0093] Collect the wind tower measurement data and their spatial coordinates within the regional wind farm station, and clip the spatial neighborhood of the wind vector field simulation results to characterize the local micrometeorological characteristics of the wind farm station:

[0094]

[0095] in, is the wind vector field simulation result, and are the longitude and latitude indexes of the i-th wind farm station, r is the size of the spatial neighborhood, and N is the total number of wind farm stations in the region.

[0096] To accurately describe the differences in temporal and spatial phase and modulus between the actual wind speed measurements at wind farms and the wind vector field, a differential method is used to construct key features:

[0097]

[0098] Among them, (m,n) is the local area S of the i-th wind farm station i The position index of |·| represents the local area S i The number of pixels in .

[0099] The partition aggregation algorithm is used to achieve pixel-level key feature aggregation and obtain the data assimilation results:

[0100]

[0101] st T∈φ{S i} i∈[1,N] ,(m,n)∈{T1∩...∩T |T|}

[0102] in, and are wind vector field simulation and data assimilation results, respectively, f Conv (·) is a convolutional neural network, and T is the aggregation space corresponding to the wind vector field pixels (m,n).

[0103] The present invention takes into account the needs of data privacy encryption and computational efficiency. It performs gradient feature encryption processing on communication data through the horizontal federation mechanism of separated federated learning. The central platform uses the federated averaging algorithm to process the gradient features reported by each wind farm station, realizing efficient data integration and iterative solution between the central platform and the edge wind farm station model.

[0104] In order to verify the performance of the method proposed in the present invention, the following experiments were carried out in the examples. The area used for testing in this example is from Jiangsu Province. The entire province of Jiangsu includes 156 wind farms, including 40 offshore wind farms and 126 onshore wind farms, with a total installed capacity of 22.6GW. The wind vector field used in the example was selected from the ERA-5 meteorological reanalysis data set developed by the European Centre for Meteorological Forecasts, which provides high-precision wind vector field data with a geographical area (23°N-43°N, 110°E-130°E) and a geographical resolution of 0.25°, and a time resolution of 1 hour. The measurement and simulation elevation of the wind field is 1000hPa (approximately 111.8m).

[0105] The embodiment uses the standard absolute reference error to evaluate the wind power prediction results:

[0106]

[0107] in, and P t The predicted value and actual value of wind power at time t respectively. t is the time step of the test set, P c The installed capacity of the wind farm.

[0108] For the ERA-5 meteorological reanalysis dataset and the in-situ measurement data of wind farms across Jiangsu Province, temporal synchronization matching is performed. The data from June 2022 to December 2023 are selected as the training set (13,140 samples), and the 8,760 samples from January 2024 to December 2024 are used as test samples. In this embodiment, Adaptive Perceptual Graph Convolution (APSC), Heterogeneous Spatiotemporal Graph Convolution (HSGC), and Computational Fluid Dynamics (CFD) algorithms are used as comparison methods. The regional wind power prediction results obtained using the present invention are shown in Table 1. Table 2 shows the improved computational friendliness of the provincial coordinated prediction mechanism obtained by applying the present invention. Figure 5 Shows the multi-step prediction results of the wind vector field in the comparison between the present invention and the Ensemble Kalman Filter (EnKF) in the embodiment. Figure 6 Shows the wind power prediction results of the present invention at wind farms across Jiangsu Province Figure 7 Then shows the prediction of the total wind power prediction for the whole province. Figure 5 As shown, the wind vector field prediction method of the present invention applied to the provincial central dispatching platform shows an accuracy advantage in multi-step prediction, and can better fit the pixel distribution of the real wind vector field compared with the EnKF algorithm. Figure 6 As shown, the present invention (shown as Proposed in the figure) shows an accuracy advantage in different terrain regions of the whole province (offshore, plain, lake area, coastal, etc.). Figure 7 As shown, the present invention also has an accuracy advantage in the prediction of the total provincial wind power. It has an accuracy advantage in terms of the overall fitting situation and the maximum prediction error section.

[0109] Table 1 Wind power prediction results of the present invention at representative wind farms and across the whole province

[0110]

[0111] Table 2 Improvement in computational friendliness of the provincial wind power coordinated prediction mechanism

[0112]

[0113] In summary, the regional wind power coordinated prediction method provided by the present invention for efficient data integration between the central platform and wind farms is based on the federated separation learning mechanism, realizes the splitting of the wind vector field-local wind power prediction task, extracts the spatio-temporal dynamic characteristics of the wind vector field based on the stacked Fourier operator on the central platform, and uses the simulation results of the wind vector field and the wind speed measurements of regional wind farms to perform data assimilation based on dense-sparse adaptive convolution. This solution can be applied to the regional power dispatching and control of a large number of wind farms connected to the grid, and is deployed correspondingly in the regional power dispatching control center and all wind farms in the region. By integrating the wind vector field model data and the station measurement data to predict the regional wind power, it can provide regional wind power prediction results for 1 to 10 hours, and then guide the regional power dispatching control center to adjust the regional power balance plan according to the prediction results, reduce the reserve capacity, and promote the power supply guarantee of the power grid and the consumption of new energy.

[0114] Another embodiment of the present invention further provides a regional wind power coordinated prediction system, including a central platform and several wind farms. The central platform deploys a wind vector field prediction model on the central platform side, and each wind farm deploys a local wind speed prediction model. The central platform and the wind farms repeatedly perform the following operations until the wind vector field prediction model on the central platform side and the local wind speed prediction models of all wind farms converge:

[0115] Each wind farm uses the local wind speed prediction model for forward propagation, and applies differential privacy noise to the activation features of the final layer of the local wind speed prediction model to obtain A i,t , and the activation feature A i,t and the prediction target are reported to the central platform together, where d is the time window to be predicted;

[0116] The central platform uses the wind vector field of the previous time period and the activation features {A i,t} i∈[1,N] collected from each wind farm to perform wind vector field simulation and data assimilation. N is the total number of wind farms. Based on the loss function of the wind vector field prediction on the central platform side calculate the wind vector field prediction loss to obtain the gradient information of the activation feature set

[0117] and sequentially distribute the gradient information to the corresponding wind farms; the central platform then updates the weight W of the wind vector field prediction model according to the gradient information obtained by backpropagation t S ;

[0118] Each wind farm combines the gradient information dA i,t distributed by the central platform, and calculates the gradient in combination with the local wind speed prediction model of the wind farm Furthermore, 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 of the wind farm station.

[0119] Each wind farm station is also deployed with a power prediction model, and the power prediction model obtains the multi-step wind power prediction result based on the local wind speed prediction result and the wind speed-power conversion model.

[0120] For the construction and update methods of the wind vector field prediction model on the central platform side and the local wind speed prediction model of the wind farm station, as well as the specific implementation process of the wind vector field simulation and data assimilation on the central platform, reference can be made to the relevant descriptions in the above method embodiments, which will not be elaborated here.

Claims

1. A regional wind power coordinated prediction method for data integration between the central platform and wind farms, characterized in that The method includes the following steps: S1: Each wind farm uses the local wind speed prediction model for forward propagation, and applies differential privacy noise to the activation features of the final layer of the local wind speed prediction model to obtain A i,t , and the activation feature A i,t and the prediction target are reported to the central platform together, where d is the time window to be predicted; S2: The central platform uses the wind vector field of the previous time period and the activation features {A i,t} i∈[1,N] collected from each wind farm station to perform wind vector field simulation and data assimilation. N is the total number of wind farm stations. Based on the loss function of the wind vector field prediction on the central platform side calculate the wind vector field prediction loss to obtain the gradient information of the activation feature set and distribute the gradient information to the corresponding wind farm stations in sequence; the central platform then updates the weights W of the wind vector field prediction model according to the gradient information obtained by backpropagation t S ; S3: Each wind farm calculates the gradient according to the gradient information dA distributed by the central platform i,t , combined with the local wind speed prediction model of the wind farm Furthermore, 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 of the wind farm S4: Repeat the above steps S1 - S3 until the central platform wind vector field prediction model and all wind farm local wind speed prediction models 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, wherein In step S1, the local wind speed prediction model of the wind farm is as follows: Among them, and v i are the predicted wind speed and the measured wind speed of the i-th wind farm respectively, and f Traf,i is the Transformer prediction model local to the i-th wind farm.

3. The method according to claim 2, characterized in that, In step S2, the central platform side wind vector field prediction model is as follows: Among them, the superscript * indicates that the data is encrypted by the differential privacy protection algorithm. and are the wind vector field simulation and data assimilation results respectively, and P represents the word embedding operation. denotes from the first layer to the nth a sequentially execute the Fourier neural operator f afo,n ; represents the data assimilation method.

4. The method according to claim 3, characterized in that In step S2, the central platform implements wind vector field simulation, which specifically includes: For the spatio-temporal wind vector field where \(c=(x,y)\) is the spatial position index of the wind vector field and \(t\) is the temporal position. Based on physical prior knowledge, the spatio-temporal motion law of the wind vector field can be constructed using the Navier-Stokes equation: where γ represents the fluid viscosity coefficient, and f(c) represents a time-invariant scalar matrix, represents the uncertain motion; Introduce a stacked Fourier neural operator to achieve iterative solution: Among them, is to iteratively update the wind vector field feature map of the nth layer, represents the Fourier neural operator model of the nth layer, w n and b n represent the weights and biases of the nth layer; f afo,n represents the Fourier neural operator of the nth layer; Utilize the property that the time - domain partial derivative is equivalent to the product of frequency - domain coefficients, and solve the single - layer Fourier neural operator through the fast Fourier transform: Among them, D represents the spatial range of the wind vector field, and k φ represents a learnable filtering function, and x k is its index. And are the fast Fourier transform and its inverse transform; R φ represents the integrated filtering function.

5. The method according to claim 4, wherein In step S2, the central platform conducts data assimilation, which specifically includes: Collect the measured wind data and its spatial coordinates of the wind measurement towers of the wind farms within the regional range, and crop the spatial neighborhood of the wind vector field simulation results as the local micro - meteorological characteristics representing the wind farms: Among them, is the simulation result of the wind vector field, and are respectively the longitude index and latitude index of the i-th wind farm, and r is the size of the spatial neighborhood; Construct key features using the difference method: where (m, n) is the position index of the local area S of the i-th wind farm i ; |·| represents the number of pixels in the local area S i ; Utilize the partition aggregation algorithm to achieve pixel - level key feature aggregation to obtain the data assimilation results: Among them, and are the wind vector field simulation and data assimilation results respectively, and f Conv (·) is a convolutional neural network, and T is the aggregation space corresponding to the wind vector field pixel (m, n).

6. The method according to claim 5, wherein In step S2, the central platform updates the weights of the wind vector field prediction model as follows: Among them, W t S is the weight of the wind vector field prediction model at time t, and η is the model weight update parameter.

7. The method according to claim 6, wherein In step S3, the wind farm calibrates the gradient information as follows: where, σ represents the noise scale, and C′ represents the noise modulus applied to the gradient information. denotes a normal distribution with an expectation of 0 and a variance of σ 2 C′I, where I represents the all-ones matrix.

8. The method according to claim 7, characterized in that In step S3, the wind farm updates the weights of the local wind speed prediction model, specifically as follows: Where η is the model weight update parameter.

9. The method according to claim 1, characterized in that In step S4, 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, specifically as follows: Among them, and P i are the predicted power value and the measured input value of the i-th wind farm respectively, and f MLP,i is the wind speed-power conversion model of the i-th wind farm locally, is bitwise multiplication.

10. A regional wind power coordinated prediction system, characterized in that, It includes a central platform and several wind farms. The central platform deploys a central platform side wind vector field prediction model, and each wind farm deploys a local wind speed prediction model. The central platform and the wind farms repeatedly perform the following operations until the central platform side wind vector field prediction model and all wind farm local wind speed prediction models converge: Each wind farm station performs forward propagation using a local wind speed prediction model, and adds differential privacy noise to the activation features of the final layer of the local wind speed prediction model to obtain A i,t , and the activation features A i,t and the prediction target are reported to the central platform together, where d is the time window to be predicted; The central platform utilizes the wind vector field of the previous time period and the activation features {A i,t} i∈[1,N] collected from each wind farm station to perform wind vector field simulation and data assimilation. N is the total number of wind farm stations. Based on the loss function of the wind vector field prediction on the central platform side calculate the wind vector field prediction loss to obtain the gradient information of the activation feature set and distribute the gradient information to the corresponding wind farm stations in sequence; the central platform then updates the weights W of the wind vector field prediction model according to the gradient information obtained by backpropagation t S ; Each wind farm station calculates the gradient according to the gradient information dA distributed by the central platform i,t , combined with the local wind speed prediction model of the wind farm station Furthermore, 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 of the wind farm station Each wind farm also deploys a power prediction model, and the power prediction model obtains the multi - step wind power prediction results based on the local wind speed prediction results and the wind speed - power conversion model.

Images