A wind speed field prediction method and device based on mechanism prior and data-driven fusion

Abstract

The application provides a wind speed field prediction method and device based on mechanism priori and data driving fusion, and belongs to the technical field of wind speed field prediction. The method acquires historical observed wind speed field, future priori wind speed field and geographical auxiliary information and pre-processes; adopts a wind speed field prediction model to determine a total residual prediction result; the wind speed field prediction model comprises a system deviation branch, a boundary topography deviation branch, an instantaneous disturbance branch and a fusion module; the system deviation branch extracts wind speed field evolution characteristics under multiple time scales to generate a system deviation residual component prediction result; the boundary topography deviation branch extracts local structure deviation characteristics related to wind direction, sea-land boundary and topography conditions to generate a boundary topography residual component prediction result; the instantaneous disturbance branch generates an instantaneous disturbance residual component prediction result according to historical wind speed field time variation characteristics; the fusion module weightedly fuses each residual component to correct the future priori wind speed field to obtain a future prediction wind speed field, and improves accuracy and continuity.

Description

Technical Field

[0001] This invention belongs to the field of wind speed field prediction technology, and in particular relates to a wind speed field prediction method and device based on the fusion of mechanism prior and data-driven approaches. Background Technology

[0002] Wind speed field forecasting is a fundamental step in wind power forecasting, wind energy resource assessment, wind farm operation scheduling, and local meteorological analysis. For offshore wind farms, coastal areas, and areas with complex terrain, wind speed not only varies over time but is also influenced by land-sea distribution, topographic relief, underlying surface conditions, and local weather processes, exhibiting significant spatiotemporal variation characteristics. Accurate future wind speed field results provide fundamental data support for wind farm output assessment, operation planning, and wind resource analysis.

[0003] In practical applications, the wind speed field in the target area typically exhibits both large-scale background variations and localized detailed variations. Wind speeds change rapidly near land-sea boundaries, and complex terrain areas are prone to local acceleration, shading, or disturbances. Short-term weather fluctuations can also cause abrupt changes at local grid points. The combined effect of these factors can lead to biases in localized predictions, manifesting as uneven changes near boundaries, larger errors in complex terrain areas, or unstable tracking of short-term fluctuations. This negatively impacts the reliability of subsequent wind power forecasting and dispatch analysis.

[0004] Existing wind speed field prediction methods mainly include mechanistic model methods and data-driven methods. Mechanistic model methods are typically based on numerical weather prediction, dynamic models, or parametric models to simulate large-scale wind field evolution and have a certain ability to express the physical background. Data-driven methods are typically based on machine learning or deep learning, learning the spatiotemporal variation patterns through historical wind speed field data to predict future wind speed fields, and have certain advantages in fitting nonlinear relationships.

[0005] However, in scenarios such as offshore wind farms, coastal boundaries, and complex terrain, the predictive stability of existing methods is still affected by field conditions. Mechanistic models are easily limited by grid scale, terrain representation, and boundary conditions in characterizing local fine-grained variations; data-driven methods still need to balance the physical rationality and spatial continuity of prediction results under conditions of varying sample distribution, local abrupt changes, and complex spatial correlations. In particular, when the wind speed field simultaneously includes long-term trends, short-term fluctuations, and local disturbances, the prediction results are prone to local anomalies, discontinuous field shapes, or uneven error distribution. Summary of the Invention

[0006] The purpose of this invention is to provide a wind speed field prediction method and device based on the fusion of mechanistic prior and data-driven approaches. This method can improve the accuracy, local adaptability and spatial continuity of future wind speed field predictions for offshore wind farms, coastal areas and complex terrain areas by combining historical wind speed field observations and geographic auxiliary information with the future prior wind speed field predictions provided by the mechanistic model.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a wind speed field prediction method based on the fusion of mechanistic priors and data-driven approaches, comprising: Acquire historical wind speed fields, future prior wind speed fields output by the mechanistic model, and geographic auxiliary information for the target area; Based on the historical wind speed field, the future prior wind speed field, and the geographic auxiliary information, a wind speed field prediction model is used to determine the total residual prediction result; wherein, the wind speed field prediction model includes a system bias branch, a boundary terrain bias branch, an instantaneous disturbance branch, and a fusion module; The system bias branch is used to extract wind speed field evolution features at different time scales from the historical observed wind speed field and the future prior wind speed field, and to generate system bias residual component prediction results based on the wind speed field evolution features. The boundary topography deviation branch is used to extract local structural deviation features related to wind direction, land-sea boundary and topography conditions based on the historical wind speed field, the future a priori wind speed field and the spatial correlation features in the geographic auxiliary information, and to generate boundary topography residual component prediction results based on the local structural deviation features. The instantaneous disturbance branch is used to extract local disturbance features based on the temporal variation characteristics of the historical observed wind speed field, and to generate instantaneous disturbance residual component prediction results based on the local disturbance features; The fusion module is used to perform weighted fusion of the prediction results of the system deviation residual component, the prediction results of the boundary terrain residual component, and the prediction results of the instantaneous disturbance residual component to obtain the total residual prediction result. The future prior wind speed field is corrected based on the total residual prediction result to obtain the future predicted wind speed field of the target area.

[0008] Furthermore, the wind speed field prediction model is obtained in the following way: Acquire historical wind speed fields, future prior wind speed fields output by the mechanism model, geographic auxiliary information, and real future wind speed fields from the training samples; Using the future prior wind speed field as the prediction reference field, a target residual field is constructed based on the difference between the actual future wind speed field and the future prior wind speed field. The historical wind speed field, the future prior wind speed field, and the geographic auxiliary information in the training samples are used as model inputs, and the target residual field is used as the residual prediction target. The model outputs the total residual prediction result, including the system bias branch, the boundary terrain bias branch, the instantaneous disturbance branch, and the fusion module. The predicted wind speed field is obtained based on the total residual prediction result and the future prior wind speed field, and the training loss is constructed based on the predicted wind speed field, the real future wind speed field and the prediction results of each residual component. Using the training loss as the optimization objective, the parameters of the system bias branch, the boundary terrain bias branch, the instantaneous disturbance branch, and the fusion module are optimized to obtain the wind speed field prediction model.

[0009] Furthermore, the geographic auxiliary information includes at least one of the following: land-sea distribution, topographic elevation, underlying surface information, and prevailing wind direction; The training loss includes a base prediction loss and a branch supervision loss. The base prediction loss is constructed based on the difference between the predicted wind speed field and the true future wind speed field. The branch supervision loss is constructed in the following manner: The target residual field is subjected to time smoothing or low-pass filtering to obtain the system bias monitoring residual; A boundary terrain region mask is constructed based on the geographic auxiliary information, and the boundary terrain deviation supervision residual is extracted from the target residual field based on the boundary terrain region mask. The remaining residual after deducting the system deviation supervision residual and the boundary terrain deviation supervision residual from the target residual field, or the high-frequency component of the target residual field, is determined as the instantaneous disturbance supervision residual; The system deviation branch loss is constructed based on the difference between the predicted results of the system deviation residual components and the system deviation supervision residual; the boundary terrain deviation branch loss is constructed based on the difference between the predicted results of the boundary terrain residual components and the boundary terrain deviation supervision residual; and the instantaneous disturbance branch loss is constructed based on the difference between the predicted results of the instantaneous disturbance residual components and the instantaneous disturbance supervision residual. The branch supervision loss is determined based on the system deviation branch loss, the boundary terrain deviation branch loss, and the instantaneous disturbance branch loss.

[0010] Furthermore, the system bias branch generates the prediction results of the system bias residual components, including: The historical wind speed field, the future prior wind speed field, and the geographic auxiliary information are encoded to obtain a temporal feature tensor. The time-series feature tensors are respectively input We obtain the outputs of each time expert from different time scales. The temporal feature tensor is pooled, and the pooling result is sequentially input into a multilayer perceptron and a normalized exponential function to obtain a temporal expert weight vector:

[0011] in: And satisfy: ; The outputs of each time expert are weighted and fused according to the time expert weight vector, and the prediction result of the system bias residual component is generated by mapping the system bias component output:

[0012] in, Represents the temporal feature tensor. Indicates the number of experts at that time. Indicates the first The output of a time-based expert This indicates a pooling operation. Represents a multilayer perceptron mapping. Represents the normalized exponent mapping, Indicates the first The time expert weight corresponding to each time expert This indicates the output mapping of the system bias component. This represents the prediction result of the system deviation residual component.

[0013] Furthermore, the boundary terrain deviation branch generates the prediction results of the boundary terrain residual components, including: Based on the prevailing wind direction, land-sea boundary, and topographic distribution in the aforementioned geographic auxiliary information, the target grid points are determined. The corresponding directional candidate neighborhood set ; Obtain the target grid points Local characteristics The directional candidate neighborhood set Candidate Neighborhood Points Local characteristics The target grid points With the candidate neighbor points Relative azimuth information between and geographical information differences ; According to the target grid points Local characteristics The candidate neighborhood points Local characteristics The relative orientation information and the geographical information differences Determine the candidate neighborhood points Corresponding directional neighborhood weights:

[0014] The target grid point is obtained by aggregating the local features and geographical information differences of the candidate neighborhood points according to the directional neighborhood weights. Corresponding aggregate features:

[0015] Based on the aggregated features corresponding to each target grid point, the boundary terrain residual component prediction result is generated by mapping the boundary terrain deviation component output:

[0016] in, Represents target grid points The corresponding set of directional candidate neighborhoods, Represents target grid points Local characteristics, Indicates candidate neighborhood points Local characteristics, Represents target grid points With candidate neighborhood points Relative orientation information between them Represents target grid points With candidate neighborhood points Geographical information differences between them Represents the directional candidate neighborhood set The candidate neighborhood index used for summation. This represents the neighborhood weight calculation function. Indicates candidate neighborhood points The corresponding directional neighborhood weights, This represents a mapping function that reflects the differences in local characteristics and geographic information of candidate neighboring points. Represents target grid points The corresponding aggregation features, Indicates the spatial grid size of the target area. This indicates the output mapping of the boundary terrain deviation component. This represents the prediction result of the boundary terrain residual components.

[0017] Furthermore, the instantaneous disturbance branch generates the prediction results of the instantaneous disturbance residual components, including: Based on the differences in wind speed fields at adjacent historical moments in the historical wind speed field, a time difference feature is constructed:

[0018] Determination and Time Alignment mechanism: prior field or background field estimation And based on the historical wind speed field observed at the current moment. Estimation of the prior field or background field of the aforementioned mechanism The difference between them is used to construct local deviation characteristics:

[0019] The historical wind speed field, the time difference feature, and the local deviation feature are input into the disturbance capture mapping to obtain the instantaneous disturbance feature:

[0020] Based on the instantaneous disturbance characteristics, the prediction result of the instantaneous disturbance residual component is generated by mapping the instantaneous disturbance component output:

[0021] in, Indicates time The corresponding time difference features, and These represent the wind speed fields at adjacent historical moments in the historical wind speed field. Indicates the current prediction time. Indicates a historical moment index. Indicates the length of the historical input sequence. Representation of time Alignment mechanism: prior field or background field estimation. Indicates local deviation characteristics. This represents the historical wind speed field. This represents the set of time difference features formed by various time difference features. This represents the local disturbance modeling function. Indicates the characteristics of instantaneous disturbances. This indicates the output mapping of instantaneous disturbance components. This indicates the prediction result of the instantaneous disturbance residual component.

[0022] Furthermore, the fusion module performs weighted fusion of the prediction results of the system bias residual components, the prediction results of the boundary terrain residual components, and the prediction results of the instantaneous disturbance residual components to obtain the total residual prediction result, including: Based on the intermediate expressions of the system deviation branch, the boundary terrain deviation branch, and the instantaneous disturbance branch. , and Generate branch fusion weights:

[0023] The prediction results of the system bias residual component, the boundary terrain residual component, and the instantaneous disturbance residual component are weighted and fused based on the branch fusion weights to obtain the total residual prediction result:

[0024] Based on the total residual prediction result, the future prior wind speed field is residually corrected to obtain the future predicted wind speed field of the target area:

[0025] in, , and These represent the intermediate expressions for the system deviation branch, the boundary terrain deviation branch, and the instantaneous disturbance branch, respectively. Indicates feature splicing, This indicates a pooling operation. Represents a multilayer perceptron mapping. Represents the normalized exponent mapping, , and These are the branch fusion weights corresponding to the system deviation branch, the boundary terrain deviation branch, and the instantaneous disturbance branch, respectively. , This indicates the total residual prediction result. , and These represent the prediction results of the system bias residual component, the boundary terrain residual component, and the instantaneous disturbance residual component, respectively. This indicates element-wise multiplication or channel-based broadcast multiplication. This indicates the predicted future wind speed field. This represents the future prior wind speed field. Indicates the current prediction time. This indicates the prediction step size.

[0026] In a second aspect, the present invention provides a wind speed field prediction device based on the fusion of mechanistic priors and data-driven approaches, comprising: The data acquisition module is used to acquire historical wind speed fields, future prior wind speed fields output by the mechanism model, and geographic auxiliary information for the target area. The residual prediction module is used to determine the total residual prediction result based on the historical observed wind speed field, the future prior wind speed field, and the geographic auxiliary information using a wind speed field prediction model; wherein, the wind speed field prediction model includes a system bias branch, a boundary terrain bias branch, an instantaneous disturbance branch, and a fusion module. The system bias branch is used to extract wind speed field evolution features at different time scales from the historical observed wind speed field and the future prior wind speed field, and to generate system bias residual component prediction results based on the wind speed field evolution features. The boundary topography deviation branch is used to extract local structural deviation features related to wind direction, land-sea boundary and topography conditions based on the historical wind speed field, the future a priori wind speed field and the spatial correlation features in the geographic auxiliary information, and to generate boundary topography residual component prediction results based on the local structural deviation features. The instantaneous disturbance branch is used to extract local disturbance features based on the temporal variation characteristics of the historical observed wind speed field, and to generate instantaneous disturbance residual component prediction results based on the local disturbance features; The fusion module is used to perform weighted fusion of the prediction results of the system deviation residual component, the prediction results of the boundary terrain residual component, and the prediction results of the instantaneous disturbance residual component to obtain the total residual prediction result. The residual correction module is used to perform residual correction on the future prior wind speed field based on the total residual prediction result, so as to obtain the future predicted wind speed field of the target area.

[0027] In a third aspect, the present invention provides an electronic device including a processor and a memory, the processor being configured to execute a computer program stored in the memory to implement a wind speed field prediction method based on the fusion of mechanistic priors and data-driven approaches.

[0028] In a fourth aspect, the present invention provides a computer-readable storage medium storing at least one instruction that, when executed by a processor, implements a wind speed field prediction method based on the fusion of mechanistic priors and data-driven approaches.

[0029] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By acquiring historical wind speed fields, future prior wind speed fields, and geographic auxiliary information, and using a wind speed field prediction model that includes system bias branches, boundary topographic bias branches, instantaneous disturbance branches, and a fusion module to generate a total residual prediction result, the future prior wind speed field is then corrected based on the total residual prediction result. This means that the wind speed field prediction process no longer relies entirely on the data-driven model to directly generate a complete future wind speed field, but instead performs bias correction based on the future prior wind speed field given by the mechanistic model. This approach can balance the background wind field representation capability of the mechanistic model and the local correction capability of the data-driven model, thereby improving the accuracy, local adaptability, and spatial continuity of future wind speed field prediction in the target area.

[0030] 2. By constructing a target residual field using the difference between the real future wind speed field and the future prior wind speed field, and using this target residual field as the residual prediction target to train the wind speed field prediction model, the model training process revolves around the prediction bias in the future prior wind speed field. At the same time, by performing unified spatial grid mapping, time index alignment, and normalization processing on historical observed wind speed fields, future prior wind speed fields, and geographic auxiliary information, it is possible to ensure that data from different sources and of different types correspond to the input under the same prediction task, which is conducive to improving data consistency and prediction stability in the model training and inference process.

[0031] 3. Wind speed field evolution features are extracted through multi-timescale time experts in the system bias branch. Local structural bias features are aggregated based on directional candidate neighborhoods, relative orientation information, and geographic information differences through the boundary topography bias branch. Instantaneous disturbance features are extracted by combining time difference features and local bias features through the instantaneous disturbance branch. This enables residual modeling for overall background wind field shift, local structural bias caused by land-sea boundaries and complex terrain, short-term abrupt changes, and high-frequency disturbances. Furthermore, the prediction results of each residual component are adaptively fused through the fusion module, which can improve the ability of wind speed field prediction results to characterize different error types in complex geographical environments. Attached Figure Description

[0032] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a flowchart illustrating a wind speed field prediction method based on the fusion of mechanism prior and data-driven approaches according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the residual back-addition prediction principle based on the mechanistic prior reference field in an embodiment of the present invention; Figure 3 This is a schematic diagram of error attribution decomposition and three-branch modeling in an embodiment of the present invention; Figure 4This is a flowchart illustrating the technical route of an embodiment of the present invention; Figure 5 This is a structural block diagram of a wind speed field prediction device based on the fusion of mechanism prior and data-driven methods according to an embodiment of the present invention. Figure 6 This is a structural block diagram of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0033] The present invention will now be described in detail with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other.

[0034] The following detailed description is exemplary and intended to provide further detailed explanation of the invention. Unless otherwise specified, all technical terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. The terminology used in this invention is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention.

[0035] Example 1 This invention provides a wind speed field prediction method based on the fusion of mechanistic priors and data-driven approaches, such as... Figure 1 As shown, the method includes steps S1, S2 and S3.

[0036] In one specific implementation, this embodiment takes a coastal offshore wind farm and its surrounding complex terrain area as the target region, and provides an implementation method that can be directly applied in engineering without changing the overall technical concept of the invention. The wind speed field within the target region exhibits multiple time-scale characteristics, including short-term fluctuations, periodic changes, and trend evolution. Spatially, it exhibits spatial correlations between different regions and is easily affected by land-sea boundaries, topographic relief, underlying surface differences, and local disturbances. Therefore, this embodiment does not directly generate a complete future wind speed field from a data-driven model. Instead, it uses the future prior wind speed field output by the mechanistic model as the prediction reference field, models the prediction error of this future prior wind speed field relative to the actual wind speed field, and applies the predicted residual correction to the future prior wind speed field to obtain the future predicted wind speed field for the target region.

[0037] like Figure 2 As shown, embodiments of the present invention use a future priori wind speed field As a baseline field for prediction, and combined with historical wind speed fields and geographic auxiliary information Modeling the residuals yields the predicted residual field. Then predict the residual field With future a priori wind field Perform residual back-addition to output the final predicted wind speed field. Therefore, the problem of predicting the overall future wind speed field is transformed into the problem of correcting residuals based on the future prior wind speed field.

[0038] Step S1: Obtain the historical wind speed field, the future prior wind speed field output by the mechanism model, and the geographic auxiliary information of the target area, and preprocess the historical wind speed field, the future prior wind speed field, and the geographic auxiliary information. In this embodiment of the invention, the historical wind speed field refers to wind speed field data from multiple consecutive historical moments in the target area, used to characterize the spatiotemporal evolution of wind speed in the target area before the current prediction moment. The future prior wind speed field is the prior wind speed field output by the mechanistic model for future moments. The mechanistic model can be a numerical weather prediction model, a dynamic model, or a parametric model that can provide background information on the future wind field. Geographic auxiliary information is used to characterize the spatial environmental features of the target area, which includes at least land-sea distribution, topographic elevation, and prevailing wind direction information, and may also include underlying surface information.

[0039] In this embodiment of the invention, the steps for preprocessing historical wind speed fields, future prior wind speed fields, and geographic auxiliary information include: Historical wind speed fields, future a priori wind speed fields, and geographic auxiliary information are uniformly mapped to the same spatial grid. Time index alignment and normalization are performed on historical wind speed fields, future a priori wind speed fields, and geographic auxiliary information. Historical wind speed fields are represented as follows:

[0040] in:

[0041] The future a priori wind field is represented as:

[0042] in, Indicates the current prediction time. Indicates the length of the historical input sequence. Indicates the prediction step size. Indicates the spatial grid size of the target area. Indicates the number of wind field variable channels; When historical wind speed fields and future a priori wind speed fields are represented using two-dimensional wind vector fields. , respectively, correspond to the east-west wind speed component and the north-south wind speed component; Geographic auxiliary information includes at least one of the following: land-sea distribution, topographic elevation, underlying surface information, and prevailing wind direction.

[0043] Furthermore, during preprocessing, spatial registration can be performed on historical wind speed fields, future prior wind speed fields, and geographic auxiliary information to ensure that all types of data are located under the same spatial grid coordinates; temporal index alignment can be performed on all types of data to match the historical input sequence corresponding to the historical wind speed field, the prediction time corresponding to the future prior wind speed field, and the target area corresponding to the geographic auxiliary information under the same prediction task; and normalization processing can be performed on all types of data to ensure that they are under a unified dimension, so that they can be directly input into the subsequent wind speed field prediction model.

[0044] In one specific implementation, if there are missing measurement points in the historical wind speed field, neighborhood interpolation or temporal interpolation is used to fill in the gaps. If there are outliers in the historical wind speed field or the future prior wind speed field, threshold truncation and smoothing are used to reduce the impact of outliers on residual prediction. After preprocessing, the historical wind speed field, the future prior wind speed field, and geographic auxiliary information form the model input in the same batch.

[0045] Step S2: Based on historical wind speed fields, future prior wind speed fields, and geographic auxiliary information, the total residual prediction result is determined using a wind speed field prediction model; wherein, the wind speed field prediction model includes a system bias branch, a boundary terrain bias branch, an instantaneous disturbance branch, and a fusion module. In this embodiment of the invention, the wind speed field prediction model is used to predict the residuals of wind speed field deviations that are not accurately characterized by the future prior wind speed field, using the future prior wind speed field as the prediction reference field. The model input consists of preprocessed historical observed wind speed fields, future prior wind speed fields, and geographic auxiliary information, and the model output is the total residual prediction result.

[0046] like Figure 3 As shown, the prediction errors are processed according to their formation characteristics into systematic bias components, boundary topographic bias components, and instantaneous disturbance components. The systematic bias component corresponds to the overall shift of the background wind field and multi-timescale evolution mismatch; the boundary topographic bias component corresponds to local structural biases caused by land-sea boundaries, topographic relief, and underlying surface differences; and the instantaneous disturbance component corresponds to short-term rapid changes and local high-frequency fluctuations. By modeling these three types of branches, different types of wind speed field prediction errors can be processed in their corresponding paths.

[0047] The systematic bias branch is used to extract wind speed field evolution characteristics at different time scales from historical observed wind speed fields and future a priori wind speed fields, and generate systematic bias residual component prediction results based on wind speed field evolution characteristics. The boundary topography deviation branch is used to extract local structural deviation features related to wind direction, land-sea boundary and topographic conditions based on the spatial correlation features in historical wind speed fields, future a priori wind speed fields and geographic auxiliary information, and to generate boundary topography residual component prediction results based on the local structural deviation features. The instantaneous disturbance branch is used to extract local disturbance features based on the temporal variation characteristics of historical wind speed fields, and to generate instantaneous disturbance residual component prediction results based on the local disturbance features; The fusion module is used to perform weighted fusion of the prediction results of the system deviation residual component, the boundary terrain residual component, and the instantaneous disturbance residual component to obtain the total residual prediction result; In this embodiment of the invention, the wind speed field prediction model is obtained in the following manner: Acquire historical wind speed fields, future prior wind speed fields output by the mechanism model, geographic auxiliary information, and real future wind speed fields from the training samples; Using the future prior wind speed field as the prediction reference field, the target residual field is constructed based on the difference between the actual future wind speed field and the future prior wind speed field. The historical wind speed field, future prior wind speed field and geographic auxiliary information in the training sample are used as model input, and the target residual field is used as the residual prediction target. The model outputs the total residual prediction result, including the system bias branch, the boundary terrain bias branch, the instantaneous disturbance branch and the fusion module. The predicted wind speed field is obtained based on the total residual prediction result and the future prior wind speed field, and the training loss is constructed based on the predicted wind speed field, the actual future wind speed field and the prediction results of each residual component. Using the training loss as the optimization objective, the parameters of the system bias branch, boundary terrain bias branch, instantaneous disturbance branch, and fusion module are optimized to obtain the wind speed field prediction model.

[0048] Specifically, during the training phase, the actual future wind speed field is assumed to be:

[0049] The target residual field of the true future wind speed field relative to the future prior wind speed field is defined as:

[0050] in, Represents the target residual field. Represents the actual future wind speed field. This represents the future prior wind speed field. The model does not directly learn the real future wind speed field. Instead, it is the learning objective residual field. Prediction results After obtaining the total residual prediction results, the final predicted wind speed field is expressed as:

[0051] in, This indicates the predicted wind speed field. This represents the total residual prediction result.

[0052] To enable the system bias branch, boundary terrain bias branch, and instantaneous disturbance branch to learn the corresponding residual features, this invention further constructs supervised residuals corresponding to each branch.

[0053] The target residual field is subjected to time smoothing or low-pass filtering to obtain the system bias monitoring residual.

[0054] A boundary topographic region mask is constructed based on geographic auxiliary information, and the boundary topographic deviation supervision residual is extracted from the target residual field based on the boundary topographic region mask. The geographic auxiliary information includes at least one of the following: land-sea distribution, topographic elevation, underlying surface information, and prevailing wind direction.

[0055] The remaining residual after deducting the system deviation supervision residual and the boundary terrain deviation supervision residual from the target residual field, or the high-frequency components of the target residual field, are determined as the instantaneous disturbance supervision residual.

[0056] To avoid the system bias branch, boundary terrain bias branch, and instantaneous disturbance branch performing indiscriminate learning solely under the total residual objective, this invention further establishes a branch constraint mechanism, causing different branches to be biased towards learning different types of residual features. The correspondence between the above three branches and the three types of residuals is not determined solely by the branch names, but is jointly defined by the input features of each branch, the construction method of the supervised residuals, and the training loss. The branch constraint mechanism may include at least one of branch supervised loss, region mask constraint, and temporal low-frequency / high-frequency decomposition constraint.

[0057] In a preferred embodiment, the target residual field Perform time smoothing or low-pass filtering to obtain the system bias monitoring residual. :

[0058] Furthermore, a boundary topographic region mask is constructed based on at least one of the following: land-sea boundary, topographic elevation gradient, underlying surface difference, and prevailing wind direction. And based on the boundary terrain region mask, from the target residual field Extracting boundary terrain deviation monitoring residuals :

[0059] Furthermore, the target residual field Subtracting system deviation monitoring residuals Monitoring residuals of boundary topographic deviation The remaining portion is used as the transient disturbance monitoring residual. :

[0060] Alternatively, the target residual field The high-frequency components are used as instantaneous disturbances to monitor the residuals. :

[0061] Therefore, the system bias branch uses low-frequency, long-term, or background residuals as the monitoring basis, the boundary topography bias branch uses the spatial structural residuals in the boundary topography region as the monitoring basis, and the instantaneous disturbance branch uses high-frequency components or residual disturbance residuals as the monitoring basis.

[0062] In a preferred embodiment, the total residual prediction result can be obtained by directly adding the prediction results of the three residual components:

[0063] in, This indicates the prediction results of the system bias residual components. This indicates the prediction results of the boundary topographic residual components. This represents the prediction result of the instantaneous disturbance residual component.

[0064] In another preferred embodiment, to enhance adaptability to different scenarios, the total residual prediction result is represented by an adaptive weighted fusion method as follows:

[0065] in, The fusion weights for the three branches satisfy the following:

[0066] symbol This indicates element-wise multiplication or channel-based broadcast multiplication.

[0067] In this embodiment of the invention, the step of generating the prediction result of the system deviation residual component in step S2 includes: Encoding historical wind speed fields, future prior wind speed fields, and geographic auxiliary information yields a temporal feature tensor:

[0068] Input the temporal feature tensors respectively We obtained the outputs of time experts at different time scales:

[0069] Generate a time expert weight vector based on the time series feature tensor:

[0070] in:

[0071] And it satisfies:

[0072] The outputs of each time expert are weighted and fused according to the time expert weight vector, and the prediction results of the system bias residual component are generated by mapping the system bias component output:

[0073] in, Represents the input feature encoding mapping. This represents the historical wind speed field. Indicates the future a priori wind speed field. Represents geographic auxiliary information, Indicates the current prediction time. Indicates the length of the historical input sequence. Indicates the prediction step size. Indicates the number of experts at that time. Indicates the first A time-based expert mapping Indicates the first The output of a time expert This indicates a pooling operation. Represents a multilayer perceptron mapping. Represents the normalized exponent mapping, Indicates the first The time expert weight corresponding to each time expert This indicates the output mapping of the system bias component. This represents the prediction result of the system bias residual component.

[0074] In this embodiment of the invention, the systematic bias branch is used to model the overall trend shift, multi-timescale evolution mismatch, and background calibration error between the future prior wind speed field and the actual wind speed field. Different time experts can correspond to different time-sampling scales, different time-receptive fields, or different time-attention ranges, thereby extracting wind speed field evolution features at different time scales. This is achieved through time expert weight vectors. The system bias branch can adaptively determine the strength of the expert's role at different times based on the current input state.

[0075] In this embodiment of the invention, the step of generating the boundary terrain residual component prediction result by the boundary terrain deviation branch includes: Target grid points are determined based on the prevailing wind direction, land-sea boundaries, and topographic distribution from the geographic auxiliary information. The corresponding directional candidate neighborhood set Specifically, with target grid pointsi Centered on the preset spatial radius r Internal determination of the basic candidate neighborhood set :

[0076] in, j Indicates candidate grid points, d(i,j) Represents target grid points i With candidate grid points j The spatial distance between them.

[0077] Based on target grid points i The corresponding prevailing wind direction determines the wind direction unit vector. e i And based on the target grid points i Point to candidate grid points j Direction vector v ij Determine candidate grid points j Relative to target grid point i The upstream or downstream relationship. When candidate grid points j When located within the upstream or downstream range corresponding to the prevailing wind direction, candidate grid points will be... j Add to wind direction candidate neighborhood set N i wind In a preferred embodiment, the wind direction candidate neighborhood set N i wind Represented as:

[0078] in, This indicates the wind direction screening threshold. Represents the unit vector of wind direction With direction vector The cosine of the angle between them. Therefore, upstream and / or downstream neighboring points related to the direction of the prevailing wind propagation can be selected.

[0079] Determine the land-sea boundary grid points based on the land-sea distribution, and identify the target grid points. i Is it located near a land-sea boundary? (Target grid point) i The distance between the land and sea boundary is less than or equal to a preset boundary distance threshold. τ b At that time, from the basic candidate neighborhood set Adjacent grid points located on both sides of the land-sea boundary are selected to form a candidate neighborhood set for the boundary. N i bndIn a preferred embodiment, the boundary candidate neighborhood set N i bnd Represented as:

[0080] in, d b (i) Represents target grid points i Distance to the land and sea border, land(i) and land(j) Representing the target grid points respectively i and candidate grid points j The corresponding land and sea attributes.

[0081] Target grid points are calculated based on the terrain elevation field. i With the basic candidate neighborhood set Candidate grid points j The terrain elevation difference or terrain gradient difference between the points is used to select candidate grid points that meet the preset conditions, forming a terrain candidate neighborhood set. In a preferred embodiment, the terrain candidate neighborhood set Represented as:

[0082] in, T i and T j Representing the target grid points respectively i and candidate grid points j The terrain elevation at the location, T i and T j Representing the target grid points respectively i and candidate grid points j The topographic elevation gradient at the location, τ T Indicates the threshold for terrain elevation difference. τ T This represents the terrain gradient difference threshold.

[0083] The wind direction candidate neighborhood set N i wind Boundary candidate neighborhood set N i bnd and terrain candidate neighborhood set Perform a union operation on at least one of the following to obtain the target grid points.i The corresponding directional candidate neighborhood set :

[0084] in, N i wind Candidate neighborhoods used to characterize the relationship with the propagation of the prevailing wind direction N i bnd Used to characterize candidate neighborhoods on both sides of the land-sea boundary. This is used to characterize candidate neighborhoods related to topographic elevation differences or topographic gradient differences. Thus, a directional candidate neighborhood set is formed. It is not determined solely by a fixed spatial distance, but rather by a combination of wind direction propagation relationships, land-sea boundary relationships, and topographic differences, thereby enabling the boundary topographic deviation branch to focus on aggregating local neighborhood information related to the boundary topographic deviation.

[0085] Get target grid points Local characteristics Targeted candidate neighborhood set Candidate Neighborhood Points Local characteristics Target grid points With candidate neighborhood points Relative azimuth information between and geographical information differences ; Based on target grid points Local characteristics Candidate Neighborhood Points Local characteristics Relative azimuth information and geographical information differences Determine candidate neighborhood points Corresponding directional neighborhood weights:

[0086] The target grid points are obtained by aggregating the local features and geographic information differences of candidate neighborhood points based on directional neighborhood weights. Corresponding aggregate features:

[0087] Based on the aggregated features corresponding to each target grid point, the boundary terrain residual component prediction result is generated by mapping the boundary terrain deviation component output:

[0088] in, Represents target grid points The corresponding set of directional candidate neighborhoods, Represents target grid points Local characteristics, Indicates candidate neighborhood points Local characteristics, Represents target grid points With candidate neighborhood points Relative orientation information between them Represents target grid points With candidate neighborhood points Geographical information differences between them Represents the set of directional candidate neighborhoods The candidate neighborhood index used for summation. This represents the neighborhood weight calculation function. Indicates candidate neighborhood points The corresponding directional neighborhood weights, This represents a mapping function that reflects the differences in local characteristics and geographic information of candidate neighboring points. Represents target grid points The corresponding aggregation features, Indicates the spatial grid size of the target area. This indicates the output mapping of the boundary terrain deviation component. This represents the prediction results of the boundary terrain residual components.

[0089] In this embodiment of the invention, the boundary topography deviation branch is used to model local structural deviations under the influence of land-sea boundaries, complex terrain, and prevailing wind direction. Oriented candidate neighborhood set. The construction simultaneously considers wind direction, land-sea boundaries, and topographic distribution, enabling target grid points to aggregate information from candidate neighboring points with directional propagation relationships or geographical relevance. Relative orientation information. Used to characterize the directional relationship between the target grid point and candidate neighboring points, geographic information differences. Aggregated features are used to characterize the differences between the two in terms of marine and land properties, topographic elevation, or underlying surface conditions. It can be used to reflect local structural deviations related to wind direction, land-sea boundaries, and topographic conditions.

[0090] In this embodiment of the invention, the step of generating the prediction result of the instantaneous disturbance residual component by the instantaneous disturbance branch includes: Based on the differences in wind speed fields between adjacent historical moments in historical observations, time difference characteristics are constructed:

[0091] Determine the current prediction time Alignment mechanism: prior field or background field estimation ; wherein, the Prioritize using mechanistic models for the current prediction time. Output or back-calculate the background wind speed field, analysis field, or assimilated background field; when the mechanistic model does not directly provide the current prediction time. For the corresponding background field, based on the mechanism model, the background wind speed field output at adjacent time points is interpolated or aligned with the time index to obtain the result with the current prediction time. Aligned background field; when the corresponding mechanistic background field cannot be obtained, time smoothing or low-pass filtering is performed on the historical wind speed field to obtain the current prediction time. Background field estimation.

[0092] Based on historical wind speed field observations at the current moment Estimation of the prior field or background field of the aforementioned mechanism The difference between them is used to construct local deviation characteristics:

[0093] By inputting historical wind speed fields, temporal difference features, and local bias features into the disturbance capture mapping, instantaneous disturbance features are obtained:

[0094] Based on the instantaneous disturbance characteristics, the prediction results of the instantaneous disturbance residual components are generated by mapping the output of the instantaneous disturbance components:

[0095] in, Indicates time The corresponding time difference features, and These represent the wind speed fields at adjacent historical moments in the historical wind speed field. Indicates the current prediction time. Indicates a historical moment index. Indicates the length of the historical input sequence. Representation of time Alignment mechanism: prior field or background field estimation. Indicates local deviation characteristics. This represents the historical wind speed field. This represents the set of time difference features formed by various time difference features. This represents the local disturbance modeling function. Indicates the characteristics of instantaneous disturbances. This indicates the output mapping of instantaneous disturbance components. This represents the prediction result of the instantaneous disturbance residual component.

[0096] In this embodiment of the invention, the transient disturbance branch is used to capture local high-frequency variations, short-term abrupt changes, and non-stationary disturbance information in the wind speed field. Temporal difference features. Local bias characteristics are used to enhance the model's sensitivity to the rate of change of historical wind speed fields. This is used to characterize the local differences between the current observed wind speed field and the currently aligned mechanistic prior field or background field estimate. It involves inputting historical observed wind speed fields, temporal difference features, and local bias features into the perturbation capture map. This allows us to obtain instantaneous disturbance characteristics for predicting the instantaneous disturbance residual components.

[0097] In this embodiment of the invention, the step of the fusion module weightedly fusing the prediction results of the system deviation residual component, the boundary terrain residual component, and the instantaneous disturbance residual component to obtain the total residual prediction result includes: Based on the intermediate expressions of the system bias branch, boundary terrain bias branch, and instantaneous disturbance branch. , and Generate branch fusion weights:

[0098] The prediction results of the system bias residual component, the boundary topography residual component, and the instantaneous disturbance residual component are weighted and fused based on the branch fusion weights to obtain the total residual prediction result:

[0099] Based on the total residual prediction results, the future prior wind speed field is corrected using residuals to obtain the future predicted wind speed field for the target area:

[0100] in, , and These represent the intermediate expressions for the system bias branch, the boundary terrain bias branch, and the instantaneous disturbance branch, respectively. Indicates feature splicing, This indicates a pooling operation. Represents a multilayer perceptron mapping. Represents the normalized exponent mapping, , and These are the branch fusion weights corresponding to the system bias branch, the boundary terrain bias branch, and the instantaneous disturbance branch, respectively. , This represents the total residual prediction result. , and These represent the prediction results for the system bias residual components, the boundary topography residual components, and the instantaneous disturbance residual components, respectively. This indicates element-wise multiplication or channel-based broadcast multiplication. This indicates the predicted wind speed field. Indicates the future a priori wind speed field. Indicates the current prediction time. This indicates the prediction step size.

[0101] like Figure 4 As shown, in the complete technical process, the input data includes historical observed wind speed fields. Future a priori wind speed field and geographic auxiliary information After unified preprocessing, the above input data forms the model input corresponding to the prediction task. During the training phase, a target residual field is constructed based on the real future wind speed field and the future prior wind speed field. Subsequently, the prediction results of each residual component are obtained through a branch prediction module that includes a system bias branch, a boundary terrain bias branch, and an instantaneous disturbance branch. The total residual prediction result is obtained through an adaptive fusion module. Finally, the future prior wind speed field is residual corrected based on the total residual prediction result to obtain the final predicted wind speed field.

[0102] In this embodiment of the invention, in order to improve the accuracy, boundary rationality, terrain adaptability and spatial continuity of the prediction results, and to make the system deviation branch, boundary terrain deviation branch and instantaneous disturbance branch correspond to different types of residual modeling tasks, the wind speed field prediction model can be jointly optimized by supervised loss, branch supervised loss and multiple prior constraints during the training phase.

[0103] First, define the basic prediction loss as:

[0104] in, Indicates the number of samples. Indicates the first The predicted wind speed field corresponding to each training sample. Indicates the first The real future wind speed field corresponding to each training sample.

[0105] To constrain different branches to learn corresponding types of residual features, a branch supervision loss is defined. The systematic bias branch loss is expressed as:

[0106] The boundary terrain deviation branch loss is expressed as:

[0107] The instantaneous disturbance branch loss is expressed as:

[0108] in, They represent the first n The prediction results of the system bias residual components, the boundary terrain bias components, and the instantaneous disturbance residual components corresponding to each training sample; They represent the first n The system bias supervision residual, boundary terrain bias supervision residual, and instantaneous disturbance supervision residual are corresponding to each training sample. The system bias supervision residual, boundary terrain bias supervision residual, and instantaneous disturbance supervision residual can be obtained from the target residual field through time-domain low-frequency / high-frequency decomposition, region mask extraction, or residual residual decomposition.

[0109] Based on the difference between the predicted results of the system bias residual components and the system bias monitoring residual, a system bias branch loss is constructed; based on the difference between the predicted results of the boundary terrain bias components and the boundary terrain bias monitoring residual, a boundary terrain bias branch loss is constructed; based on the difference between the predicted results of the instantaneous disturbance residual components and the instantaneous disturbance monitoring residual, an instantaneous disturbance branch loss is constructed.

[0110] The branch supervision loss is determined based on the system deviation branch loss, the boundary terrain deviation branch loss, and the instantaneous disturbance branch loss.

[0111] To constrain the rationality of the predicted changes in the relationship between the land and sea boundaries, let... For the set of adjacent mesh pairs at the land-sea boundary, the land-sea boundary constraint term is defined as follows:

[0112] in, Weighting coefficients for boundary locations, and These represent the boundary adjacent grid points respectively. and grid points The corresponding predicted wind speed field value, and These represent the boundary adjacent grid points respectively. and grid points The corresponding actual wind speed field value.

[0113] To enhance the predictive field's response to the influence of complex terrain, the terrain elevation field is assumed to be... Its spatial gradient is Then the terrain prior constraint term is defined as:

[0114] in, This is the terrain adjustment coefficient. Represents the terrain elevation field at grid points Spatial gradient at that location This indicates the predicted wind speed field at the grid points. Spatial gradient at that location This represents the true wind speed field at the grid points. Spatial gradient at a given location.

[0115] To suppress isolated noise and unreasonable local jumps that may occur in the prediction field, a spatial continuity constraint term is further defined:

[0116] in, For point A set of spatially adjacent grids.

[0117] In the implementation where the total residual prediction result is obtained by directly adding the prediction results of the three residual components, a decomposition consistency loss can also be introduced to constrain the consistency of expression between the three types of residual components and the total residual field. Its definition is:

[0118] in, Indicates the first The total residual prediction result corresponding to each training sample , , They represent the first The prediction results of the system bias residual components, boundary terrain residual components, and instantaneous disturbance residual components for each training sample.

[0119] In the implementation where the total residual prediction result is obtained by directly adding the prediction results of the three residual components, the joint loss function of this embodiment is expressed as follows:

[0120] in, These are the weight coefficients for each loss term. In the implementation where the total residual prediction result is obtained using an adaptive weighted fusion method, the joint loss function includes the basic prediction loss, system bias branch loss, boundary terrain bias branch loss, instantaneous disturbance branch loss, land-sea boundary constraint term, terrain prior constraint term, and spatial continuity constraint term. During training, the joint loss function is used as the optimization objective for updating model parameters, and the parameters of the system bias branch, boundary terrain bias branch, instantaneous disturbance branch, and fusion module are simultaneously optimized through backpropagation.

[0121] Step S3: Based on the total residual prediction results, perform residual correction on the future prior wind speed field to obtain the future predicted wind speed field of the target area.

[0122] In this embodiment of the invention, the inference stage directly utilizes the preprocessed historical wind speed field, the future prior wind speed field, and geographic auxiliary information input to train a wind speed field prediction model, outputting the total residual prediction result. Subsequently, the future prior wind speed field is residual-corrected based on the total residual prediction result to obtain the future predicted wind speed field for the target area in future time periods. This future predicted wind speed field can be a scalar wind speed field or a wind vector component field including east-west and north-south wind speed components. When the predicted variable is a wind vector component field, the residual correction applies to each wind speed component separately; when the predicted variable is a scalar wind speed field, the residual correction is directly used for scalar field prediction.

[0123] In practical operation, the latest historical wind speed field of the target area, the future prior wind speed field output by the mechanistic model, and geographical auxiliary information can be input to output the predicted wind speed field for the target time period. The prediction results can be used for wind power prediction, wind farm operation scheduling, short-term wind resource assessment, and wind speed field visualization. In areas with land-sea boundaries and complex terrain, this implementation method can reduce local anomalies and field discontinuities, and improve the stability of engineering applications.

[0124] Example 2 like Figure 5 As shown, based on the same inventive concept as the above embodiments, the present invention also provides a wind speed field prediction device based on the fusion of mechanism prior and data-driven methods, comprising: The data acquisition module is used to acquire historical wind speed fields, future prior wind speed fields output by the mechanism model, and geographic auxiliary information for the target area. The residual prediction module is used to determine the total residual prediction result based on historical wind speed fields, future prior wind speed fields, and geographic auxiliary information using a wind speed field prediction model. The wind speed field prediction model includes a system bias branch, a boundary topography bias branch, an instantaneous disturbance branch, and a fusion module. The systematic bias branch is used to extract wind speed field evolution characteristics at different time scales from historical observed wind speed fields and future a priori wind speed fields, and generate systematic bias residual component prediction results based on wind speed field evolution characteristics. The boundary topography deviation branch is used to extract local structural deviation features related to wind direction, land-sea boundary and topographic conditions based on the spatial correlation features in historical wind speed fields, future a priori wind speed fields and geographic auxiliary information, and to generate boundary topography residual component prediction results based on the local structural deviation features. The instantaneous disturbance branch is used to extract local disturbance features based on the temporal variation characteristics of historical wind speed fields, and to generate instantaneous disturbance residual component prediction results based on the local disturbance features; The fusion module is used to perform weighted fusion of the prediction results of the system deviation residual component, the boundary terrain residual component, and the instantaneous disturbance residual component to obtain the total residual prediction result; The residual correction module is used to correct the future prior wind speed field based on the total residual prediction results, so as to obtain the future predicted wind speed field of the target area.

[0125] Example 3 like Figure 6 As shown, the present invention also provides an electronic device 100 for implementing a wind speed field prediction method based on the fusion of mechanism priors and data-driven approaches; The electronic device 100 includes a memory 101, at least one processor 102, a computer program 103 stored in the memory 101 and executable on at least one processor 102, and at least one communication bus 104.

[0126] The memory 101 can be used to store the computer program 103. The processor 102 implements the wind speed field prediction method based on the fusion of mechanism prior and data-driven approach in Embodiment 1 by running or executing the computer program stored in the memory 101 and calling the data stored in the memory 101.

[0127] The memory 101 may primarily include a program storage area and a data storage area. The program storage area may store the operating system, application programs required for at least one function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created based on the use of the electronic device 100 (such as audio data), etc. In addition, the memory 101 may include non-volatile memory, such as hard disk, RAM, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other non-volatile solid-state storage device.

[0128] At least one processor 102 may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Processor 102 may be a microprocessor or any conventional processor. Processor 102 is the control center of electronic device 100, connecting various parts of electronic device 100 via various interfaces and lines.

[0129] The memory 101 in the electronic device 100 stores multiple instructions to implement a wind speed field prediction method based on the fusion of mechanism prior and data-driven methods. The processor 102 can execute multiple instructions to achieve: acquiring the historical observed wind speed field of the target area, the future prior wind speed field output by the mechanism model, and geographic auxiliary information. Based on historical wind speed fields, future a priori wind speed fields, and geographical auxiliary information, a wind speed field prediction model is used to determine the total residual prediction results. The systematic bias branch is used to extract wind speed field evolution characteristics at different time scales from historical observed wind speed fields and future a priori wind speed fields, and generate systematic bias residual component prediction results based on wind speed field evolution characteristics. The boundary topography deviation branch is used to extract local structural deviation features related to wind direction, land-sea boundary and topographic conditions based on the spatial correlation features in historical wind speed fields, future a priori wind speed fields and geographic auxiliary information, and to generate boundary topography residual component prediction results based on the local structural deviation features. The instantaneous disturbance branch is used to extract local disturbance features based on the temporal variation characteristics of historical wind speed fields, and to generate instantaneous disturbance residual component prediction results based on the local disturbance features; The fusion module is used to perform weighted fusion of the prediction results of the system deviation residual component, the boundary terrain residual component, and the instantaneous disturbance residual component to obtain the total residual prediction result; Based on the total residual prediction results, the future prior wind speed field is corrected using residuals to obtain the future predicted wind speed field for the target area.

[0130] Example 4 If the modules / units integrated in the electronic device 100 are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, and read-only memory (ROM).

[0131] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0132] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0133] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0134] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0135] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0136] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A wind speed field prediction method based on the fusion of mechanism priors and data-driven approaches, characterized in that, include: Acquire historical wind speed fields, future prior wind speed fields output by the mechanistic model, and geographic auxiliary information for the target area; Based on the historical wind speed field, the future prior wind speed field, and the geographic auxiliary information, a wind speed field prediction model is used to determine the total residual prediction result; wherein, the wind speed field prediction model includes a system bias branch, a boundary terrain bias branch, an instantaneous disturbance branch, and a fusion module; The system bias branch is used to extract wind speed field evolution features at different time scales from the historical observed wind speed field and the future prior wind speed field, and to generate system bias residual component prediction results based on the wind speed field evolution features. The boundary topography deviation branch is used to extract local structural deviation features related to wind direction, land-sea boundary and topography conditions based on the historical wind speed field, the future a priori wind speed field and the spatial correlation features in the geographic auxiliary information, and to generate boundary topography residual component prediction results based on the local structural deviation features. The instantaneous disturbance branch is used to extract local disturbance features based on the temporal variation characteristics of the historical observed wind speed field, and to generate instantaneous disturbance residual component prediction results based on the local disturbance features; The fusion module is used to perform weighted fusion of the prediction results of the system deviation residual component, the prediction results of the boundary terrain residual component, and the prediction results of the instantaneous disturbance residual component to obtain the total residual prediction result. The future prior wind speed field is corrected based on the total residual prediction result to obtain the future predicted wind speed field of the target area.

2. The wind speed field prediction method based on the fusion of mechanism prior and data-driven approaches according to claim 1, characterized in that, The wind speed field prediction model is obtained in the following way: Acquire historical wind speed fields, future prior wind speed fields output by the mechanism model, geographic auxiliary information, and real future wind speed fields from the training samples; Using the future prior wind speed field as the prediction reference field, a target residual field is constructed based on the difference between the actual future wind speed field and the future prior wind speed field. The historical wind speed field, the future prior wind speed field, and the geographic auxiliary information in the training samples are used as model inputs, and the target residual field is used as the residual prediction target. The model outputs the total residual prediction result, including the system bias branch, the boundary terrain bias branch, the instantaneous disturbance branch, and the fusion module. The predicted wind speed field is obtained based on the total residual prediction result and the future prior wind speed field, and the training loss is constructed based on the predicted wind speed field, the real future wind speed field and the prediction results of each residual component. Using the training loss as the optimization objective, the parameters of the system bias branch, the boundary terrain bias branch, the instantaneous disturbance branch, and the fusion module are optimized to obtain the wind speed field prediction model.

3. The wind speed field prediction method based on the fusion of mechanism prior and data-driven approaches according to claim 2, characterized in that, The geographic auxiliary information includes at least one of the following: land and sea distribution, topographic elevation, underlying surface information, and prevailing wind direction; The training loss includes a base prediction loss and a branch supervision loss. The base prediction loss is constructed based on the difference between the predicted wind speed field and the true future wind speed field. The branch supervision loss is constructed in the following manner: The target residual field is subjected to time smoothing or low-pass filtering to obtain the system bias monitoring residual; A boundary terrain region mask is constructed based on the geographic auxiliary information, and the boundary terrain deviation supervision residual is extracted from the target residual field based on the boundary terrain region mask. The remaining residual after deducting the system deviation supervision residual and the boundary terrain deviation supervision residual from the target residual field, or the high-frequency component of the target residual field, is determined as the instantaneous disturbance supervision residual; The system deviation branch loss is constructed based on the difference between the predicted results of the system deviation residual components and the system deviation supervision residual; the boundary terrain deviation branch loss is constructed based on the difference between the predicted results of the boundary terrain residual components and the boundary terrain deviation supervision residual; and the instantaneous disturbance branch loss is constructed based on the difference between the predicted results of the instantaneous disturbance residual components and the instantaneous disturbance supervision residual. The branch supervision loss is determined based on the system deviation branch loss, the boundary terrain deviation branch loss, and the instantaneous disturbance branch loss.

4. The wind speed field prediction method based on the fusion of mechanism prior and data-driven approaches according to claim 1, characterized in that, The systematic bias branch generates the prediction results of the systematic bias residual components, including: The historical wind speed field, the future prior wind speed field, and the geographic auxiliary information are encoded to obtain a temporal feature tensor. The time-series feature tensors are respectively input We obtain the outputs of each time expert from different time scales. The temporal feature tensor is pooled, and the pooling result is sequentially input into a multilayer perceptron and a normalized exponential function to obtain a temporal expert weight vector: in: And satisfy: ; The outputs of each time expert are weighted and fused according to the time expert weight vector, and the prediction result of the system bias residual component is generated by mapping the system bias component output: in, Represents the temporal feature tensor. Indicates the number of experts at that time. Indicates the first The output of a time-based expert This indicates a pooling operation. Represents a multilayer perceptron mapping. Represents the normalized exponent mapping, Indicates the first The time expert weight corresponding to each time expert. This indicates the output mapping of the system bias component. This represents the prediction result of the system deviation residual component.

5. The wind speed field prediction method based on the fusion of mechanism prior and data-driven approaches according to claim 1, characterized in that, The boundary terrain deviation branch generates the prediction results of the boundary terrain residual components, including: Based on the prevailing wind direction, land-sea boundary, and topographic distribution in the aforementioned geographic auxiliary information, the target grid points are determined. The corresponding directional candidate neighborhood set ; Obtain the target grid points Local characteristics The directional candidate neighborhood set Candidate Neighborhood Points Local characteristics The target grid points With the candidate neighbor points Relative azimuth information between and geographical information differences ; According to the target grid points Local characteristics The candidate neighborhood points Local characteristics The relative orientation information and the geographical information differences Determine the candidate neighborhood points Corresponding directional neighborhood weights: The target grid point is obtained by aggregating the local features and geographical information differences of the candidate neighborhood points according to the directional neighborhood weights. Corresponding aggregation features: Based on the aggregated features corresponding to each target grid point, the boundary terrain residual component prediction result is generated by mapping the boundary terrain deviation component output: in, Represents target grid points The corresponding set of directional candidate neighborhoods, Represents target grid points Local characteristics, Indicates candidate neighborhood points Local characteristics, Represents target grid points With candidate neighborhood points Relative orientation information between them Represents target grid points With candidate neighborhood points Geographical information differences between them Represents the directional candidate neighborhood set The candidate neighborhood index used for summation. This represents the neighborhood weight calculation function. Indicates candidate neighborhood points The corresponding directional neighborhood weights, This represents a mapping function that reflects the differences in local characteristics and geographic information of candidate neighboring points. Represents target grid points The corresponding aggregation features, Indicates the spatial grid size of the target area. This indicates the output mapping of the boundary terrain deviation components. This represents the prediction result of the boundary terrain residual components.

6. The wind speed field prediction method based on the fusion of mechanism prior and data-driven approaches according to claim 1, characterized in that, The instantaneous perturbation branch generates the prediction results of the instantaneous perturbation residual components, including: Based on the differences in wind speed fields at adjacent historical moments in the historical wind speed field, a time difference feature is constructed: Determination and Time Alignment mechanism: prior field or background field estimation And based on the historical wind speed field observed at the current moment. Estimation of the prior field or background field of the aforementioned mechanism The difference between them is used to construct local deviation characteristics: By inputting the historical wind speed field, the time difference feature, and the local deviation feature into the disturbance capture mapping, instantaneous disturbance features are obtained: Based on the instantaneous disturbance characteristics, the prediction result of the instantaneous disturbance residual component is generated by mapping the output of the instantaneous disturbance component: in, Indicates time The corresponding time difference features, and These represent the wind speed fields at adjacent historical moments in the historical wind speed field. Indicates the current prediction time. Indicates a historical moment index. Indicates the length of the historical input sequence. Representation of time Alignment mechanism: prior field or background field estimation. Indicates local deviation characteristics. This represents the historical wind speed field. This represents the set of time difference features formed by various time difference features. This represents the local disturbance modeling function. Indicates the characteristics of instantaneous disturbances. This indicates the output mapping of instantaneous disturbance components. This indicates the prediction result of the instantaneous disturbance residual component.

7. The wind speed field prediction method based on the fusion of mechanism prior and data-driven approaches according to claim 1, characterized in that, The fusion module performs weighted fusion of the prediction results of the system deviation residual component, the prediction results of the boundary terrain residual component, and the prediction results of the instantaneous disturbance residual component to obtain the total residual prediction result, including: Based on the intermediate expressions of the system deviation branch, the boundary terrain deviation branch, and the instantaneous disturbance branch. , and Generate branch fusion weights: The prediction results of the system bias residual component, the boundary terrain residual component, and the instantaneous disturbance residual component are weighted and fused based on the branch fusion weights to obtain the total residual prediction result: Based on the total residual prediction result, the future prior wind speed field is residually corrected to obtain the future predicted wind speed field of the target area: in, , and These represent the intermediate expressions for the system deviation branch, the boundary terrain deviation branch, and the instantaneous disturbance branch, respectively. Indicates feature splicing, This indicates a pooling operation. Represents a multilayer perceptron mapping. Represents the normalized exponent mapping, , and These are the branch fusion weights corresponding to the system deviation branch, the boundary terrain deviation branch, and the instantaneous disturbance branch, respectively. , This indicates the total residual prediction result. , and These represent the prediction results of the system bias residual component, the boundary terrain residual component, and the instantaneous disturbance residual component, respectively. This indicates element-wise multiplication or channel-based broadcast multiplication. This represents the predicted future wind speed field. This represents the future prior wind speed field. Indicates the current prediction time. This indicates the prediction step size.

8. A wind speed field prediction device based on the fusion of mechanism priors and data-driven approaches, characterized in that, include: The data acquisition module is used to acquire historical wind speed fields, future prior wind speed fields output by the mechanism model, and geographic auxiliary information for the target area. The residual prediction module is used to determine the total residual prediction result based on the historical observed wind speed field, the future prior wind speed field, and the geographic auxiliary information using a wind speed field prediction model; wherein, the wind speed field prediction model includes a system bias branch, a boundary terrain bias branch, an instantaneous disturbance branch, and a fusion module. The system bias branch is used to extract wind speed field evolution features at different time scales from the historical observed wind speed field and the future prior wind speed field, and to generate system bias residual component prediction results based on the wind speed field evolution features. The boundary topography deviation branch is used to extract local structural deviation features related to wind direction, land-sea boundary and topography conditions based on the historical wind speed field, the future a priori wind speed field and the spatial correlation features in the geographic auxiliary information, and to generate boundary topography residual component prediction results based on the local structural deviation features. The instantaneous disturbance branch is used to extract local disturbance features based on the temporal variation characteristics of the historical observed wind speed field, and to generate instantaneous disturbance residual component prediction results based on the local disturbance features; The fusion module is used to perform weighted fusion of the prediction results of the system deviation residual component, the prediction results of the boundary terrain residual component, and the prediction results of the instantaneous disturbance residual component to obtain the total residual prediction result. The residual correction module is used to perform residual correction on the future prior wind speed field based on the total residual prediction result, so as to obtain the future predicted wind speed field of the target area.

9. An electronic device, characterized in that, It includes a processor and a memory, the processor being used to execute a computer program stored in the memory to implement the wind speed field prediction method based on the fusion of mechanistic priors and data-driven approaches as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores at least one instruction, which, when executed by a processor, implements the wind speed field prediction method based on the fusion of mechanistic priors and data-driven approaches as described in any one of claims 1 to 7.

Images