Ocean wind energy numerical prediction method combined with artificial intelligence and device thereof

Abstract

This invention discloses a numerical forecasting method and apparatus for marine wind energy combining artificial intelligence, belonging to the field of marine wind energy numerical forecasting technology. The method includes acquiring and assimilating initial multi-source heterogeneous data of the target sea area to generate a standardized meteorological and hydrological field. Then, a deep spatiotemporal feature extraction network is used to complete multi-scale feature learning, capturing the spatiotemporal evolution pattern of wind energy parameters. The physical laws of ocean-atmospheric motion are transformed into loss function terms to constrain network learning, constructing a physically enhanced forecasting model. A causal intervention mechanism is used to simulate the intervention effect of atmospheric circulation background on the local wind field, forming a causally interpretable forecasting model. Real-time observation data is input to deduce multi-level wind energy parameter forecast sequences, and uncertainty quantification generates probabilistic forecast products with confidence intervals. This method addresses the problems of missing physical constraints and insufficient causal interpretability in existing forecasting models, optimizing the rationality and reliability of wind energy forecasts.

Description

Technical Field

[0001] This invention belongs to the field of marine wind energy numerical forecasting technology, specifically a marine wind energy numerical forecasting method and apparatus that combines artificial intelligence. Background Technology

[0002] Currently, most numerical forecasting of marine wind energy employs purely data-driven deep learning models for forecasting calculations. Some schemes combine conventional multi-source heterogeneous data assimilation methods to construct meteorological and hydrological fields. Traditional numerical forecasting models rely on predetermined physical parameters for wind field extrapolation. This type of technology is the mainstream implementation method for marine wind energy forecasting in the industry. Purely data-driven deep learning models do not incorporate the inherent physical rules of ocean-atmosphere motion. Model training relies solely on data correlation for iteration, making forecast results prone to anomalies that violate physical laws. The feature extraction methods of traditional numerical forecasting models cannot accurately capture the long-term and short-term spatiotemporal evolution characteristics of wind energy parameters. Conventional forecasting models can only achieve correlation analysis at the data level and cannot clarify the intrinsic relationship between atmospheric circulation and local wind fields.

[0003] The forecasting models in the industry have not transformed the basic physical laws of ocean-atmosphere motion into loss function terms to constrain the learning process of deep spatiotemporal feature extraction networks. The forecasting models cannot simulate the intervention effect of different atmospheric circulation backgrounds on local wind fields through causal intervention mechanisms, making it difficult to construct wind energy numerical forecasting models with causal interpretability. This is a technical problem that current marine wind energy numerical forecasting technology needs to overcome. Summary of the Invention

[0004] This invention aims to solve at least one of the technical problems existing in the prior art; Therefore, this invention proposes a numerical forecasting method for ocean wind energy that combines artificial intelligence, comprising: Acquire initial multi-source heterogeneous data of the target sea area, and assimilate and fuse the initial multi-source heterogeneous data to generate standardized historical and current meteorological and hydrological fields; Based on a deep spatiotemporal feature extraction network, multi-scale feature learning is performed on standardized historical and current meteorological and hydrological fields to capture the long- and short-term spatiotemporal evolution patterns of wind energy-related parameters. By using the physical constraint module, the known basic physical laws of ocean-atmosphere motion are transformed into loss function terms, which constrain the learning process of the deep spatiotemporal feature extraction network and generate a prediction model with enhanced physical information. By employing a causal intervention mechanism, the intervention effect of different atmospheric circulation backgrounds on local wind fields is simulated in a forecast model enhanced with physical information, thereby constructing a wind energy numerical forecast model with causal interpretability. Real-time multi-source observation data is input into a causally interpretable wind energy numerical prediction model to extrapolate a multi-level wind energy parameter prediction sequence for the target sea area over a future period. Uncertainty quantification is performed on multi-level wind energy parameter forecast sequences to generate probability forecast products with confidence intervals.

[0005] Furthermore, the acquisition of initial multi-source heterogeneous data of the target sea area, and the assimilation and fusion processing of the initial multi-source heterogeneous data to generate standardized historical and current meteorological and hydrological fields, specifically includes: Historical and real-time data from satellite remote sensing, buoys, shore-based wind towers, reanalysis data, and numerical models were collected to form the initial multi-source heterogeneous data. Missing and outlier values ​​in the initial multi-source heterogeneous data are processed using a spatiotemporal kriging interpolation-based imputation and correction method to obtain a complete and continuous multi-source dataset. Complete and continuous multi-source datasets are uniformly interpolated onto a preset spatial latitude and longitude grid and standard pressure layer to achieve consistency in spatiotemporal resolution of data from different sources; The standardized data are corrected for deviations, using high-precision field observation data as a benchmark to eliminate systematic biases; The data after bias correction are arranged in chronological order to form a four-dimensional standardized historical and current meteorological and hydrological field containing multiple variables such as wind speed, wind direction, air temperature, air pressure, sea surface temperature, and wave height.

[0006] Furthermore, the deep spatiotemporal feature extraction network performs multi-scale feature learning on standardized historical and current meteorological and hydrological fields to capture the long- and short-term spatiotemporal evolution patterns of wind energy-related parameters, specifically including: We construct a combined network that includes a convolutional neural network, a long short-term memory network, and a self-attention mechanism as a deep spatiotemporal feature extraction network; Standardized historical and current meteorological and hydrological fields are input into the convolutional neural network part of the deep spatiotemporal feature extraction network to extract multi-scale spatial feature maps from local to regional along the spatial dimension. The multi-scale spatial feature map is unfolded in the time dimension and input into the long short-term memory network part of the deep spatiotemporal feature extraction network to learn the dynamic dependency relationship of each spatial feature point in the time series. By utilizing the self-attention mechanism of the deep spatiotemporal feature extraction network, a global correlation analysis is performed on the temporal features output by the long short-term memory network to capture the long-distance spatial interaction between wind field changes in different geographical locations. The final output is an abstract feature vector that integrates multi-scale spatial, temporal dynamics and long-range correlation features, which is used to characterize the core evolution mode of wind energy-related parameters.

[0007] Furthermore, the physical constraint module transforms known fundamental physical laws of ocean-atmospheric motion into loss function terms to constrain the learning process of the deep spatiotemporal feature extraction network, specifically including: Simplified forms of mass conservation, momentum conservation, and energy conservation applicable to the target sea area scale are selected from the fluid mechanics and thermodynamics equations as the basis of physical constraints. The abstract feature vectors output by the deep spatiotemporal feature extraction network are mapped back to the gradient or divergence fields of physical variables such as wind speed, air pressure, and temperature through a differentiable physical computation graph. Based on the physical constraints, the residuals of the physical variable fields obtained by mapping on the continuity equation and the equation of motion are calculated, and the residual norm is used as the physical conservation loss term. In the training loss function of the deep spatiotemporal feature extraction network, the physical conservation loss term and the conventional data fitting loss term are weighted and combined to form the total loss function. By optimizing the total loss function through backpropagation, the deep spatiotemporal feature extraction network can learn evolution patterns that satisfy preset physical law constraints while fitting data, thereby obtaining a prediction model with enhanced physical information.

[0008] Furthermore, the aforementioned causal intervention mechanism simulates the intervention effects of different atmospheric circulation backgrounds on local wind fields in the physically enhanced forecast model, specifically including: Identify key large-scale circulation indices that affect the target sea area and introduce them as exogenous intervention variables into the input layer of the physical information-enhanced forecasting model; In the causal intervention layer of the model, virtual interventions of different intensities are applied to the key large-scale circulation index of the input, that is, their values ​​are fixed for a specific scenario, while keeping other input variables unchanged; A forward propagation-enhanced physical information forecasting model is used to calculate the differences in the final wind energy parameter forecast values ​​output by the model under different intensities of virtual intervention. The rate of change of the difference in wind energy parameter forecasts relative to the intensity of intervention was analyzed to quantify the causal effect of key large-scale circulation indices on wind energy parameters in the target sea area. The learned causal effect strength is encoded into the model weights, enabling the model to explicitly consider the causal influence of large-scale circulation when making forecasts, thus forming a wind energy numerical forecast model with causal interpretability.

[0009] Furthermore, the step of inputting real-time multi-source observation data into a causally interpretable wind energy numerical prediction model to extrapolate a multi-level wind energy parameter forecast sequence for the target sea area over a future period specifically includes: It receives the latest real-time observation data from satellites, buoys, and wind towers, and performs assimilation, interpolation, and bias correction on the data to generate a real-time initial field. The real-time initial field is spliced ​​with the standardized historical and current meteorological and hydrological fields of historical periods to form a complete input sequence that includes historical background and real-time status; The complete input sequence is fed into a causally interpretable wind energy numerical prediction model. The model uses the output of the previous time step as part of the input of the next time step to perform multi-time step iterative forward extrapolation. During the iterative simulation, the model outputs the predicted values ​​of wind speed, wind direction, and wind power density at each grid point in the target sea area for each time step. By summarizing the outputs of all forecast time steps, a multi-level wind energy parameter forecast sequence covering wind speed, wind direction, and wind power density is formed for a period of time in the future; The process involves splicing the real-time initial field with standardized historical and current meteorological and hydrological fields from historical periods to form a complete input sequence that includes historical background and real-time status. Specifically, this includes: From the standardized historical and current meteorological and hydrological fields that have been generated, a historical data segment of a preset length is extracted before the real-time moment. The historical data contains multiple consecutive time slices. For each time slice of the extracted historical data, check its consistency with the real-time initial field in terms of spatial grid, variable definition and units to ensure that each time slice of the extracted historical data can be directly compared and stitched with the real-time initial field. The real-time initial field is treated as a new time slice and appended to the end of the historical time slice sequence, and then spliced ​​together in the time dimension. The spliced ​​complete sequence is normalized along the time axis to eliminate the magnitude differences between different time points; The final output is a complete, spatiotemporally continuous and standardized input sequence from the past to the present, which is used for forecasting and extrapolation by a wind energy numerical prediction model with causal interpretability.

[0010] Furthermore, the uncertainty quantification of the multi-level wind energy parameter forecast sequence to generate a probability forecast product with confidence intervals specifically includes: During the training phase of a wind energy numerical prediction model with causal interpretability, the error distribution of each prediction on the validation set is saved to construct an empirical distribution function of the model's prediction error. In the forecasting phase, the ensemble forecasting concept is adopted, which adds random perturbations that conform to the statistical characteristics of historical errors to the real-time initial field to generate a set of ensemble member initial fields; Each set member's initial field is input into a causally interpretable wind energy numerical prediction model, and independent iterative deductions are performed to obtain a set of wind energy parameter prediction sequences. For the forecast sequence set, at each forecast time step and at each spatial grid point, calculate the statistical quantiles of the forecast values ​​of all set members; By selecting specific high and low quantiles as upper and lower bounds, a confidence interval is added to each value of the original multi-level wind energy parameter forecast sequence to form a probability forecast product.

[0011] Furthermore, the step of unfolding the multi-scale spatial feature map in the time dimension and inputting it into the long short-term memory network part of the deep spatiotemporal feature extraction network to learn the dynamic dependencies of each spatial feature point in the time series specifically includes: The multi-scale spatial feature map output by the convolutional neural network is used to extract the features of each point along the time dimension into a time series vector according to the position of the spatial grid points. The time series vector corresponding to each spatial grid point is used as an independent input and fed into the long short-term memory network unit with shared weights of the deep spatiotemporal feature extraction network. In the long short-term memory network unit, information is selectively updated, forgotten, and output through input gate, forget gate, and output gate mechanisms, learning the temporal evolution law of each spatial feature point; The updated feature vectors of each point output by the Long Short-Term Memory network unit are rearranged according to the original spatial structure to form a feature map containing temporal dynamic information.

[0012] Furthermore, the abstract feature vector output by the deep spatiotemporal feature extraction network is mapped back to the gradient or divergence field of physical variables such as wind speed, air pressure, and temperature through a differentiable physical computation graph, specifically including: Construct a differentiable mapping network whose input is an abstract feature vector and whose output is a physical variable field corresponding to the physical meaning of standardized historical and current meteorological and hydrological fields; By utilizing the last linear transformation layer of the differentiable mapping network, high-dimensional abstract features are directly mapped into a two-dimensional spatial field of the horizontal component of wind speed, air pressure, and temperature. For the horizontal component field of wind speed obtained by mapping, the horizontal divergence field and vertical vorticity field are calculated on the preset spatial grid using a differentiable central difference operator. For the mapped pressure field, calculate its gradient field in the horizontal and vertical directions; For the mapped temperature field, calculate its horizontal gradient field. The calculated divergence field, vortex field, and gradient field together constitute the physical variable gradient or divergence field used for physical constraint calculation.

[0013] Furthermore, the present invention also includes a marine wind energy numerical forecasting device incorporating artificial intelligence, the device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor, when executing the computer program, implements the steps of the marine wind energy numerical forecasting method incorporating artificial intelligence as described above.

[0014] Compared with the prior art, the beneficial effects of the present invention are: By transforming known fundamental physical laws of ocean-atmosphere motion into loss function terms and imposing constraints on the learning process of the deep spatiotemporal feature extraction network, the network's multi-scale feature learning process can be made to conform to the physical logic of ocean-atmosphere motion. This reduces non-physical feature biases generated during data-driven learning, ensuring that the spatiotemporal evolution patterns of wind energy parameters captured by the deep spatiotemporal feature extraction network are consistent with the actual changes in the ocean-atmosphere physical field. This allows the physically enhanced forecast model to follow the basic rules of ocean-atmosphere motion during the parameter update phase, thus mitigating the interference of data noise on the feature learning process.

[0015] By employing a causal intervention mechanism to simulate the intervention effects of different atmospheric circulation backgrounds on local wind fields in physically-enhanced forecast models, the interaction path between atmospheric circulation backgrounds and local wind fields can be traced, the interference of unrelated meteorological factors on wind energy parameter forecasts can be eliminated, and the inference process of wind energy numerical forecast models can form a traceable causal logic. This ensures that the output wind energy parameter forecast sequence can truly reflect the impact of changes in atmospheric circulation backgrounds, avoids the inference bias caused by simply relying on data correlation to build forecast models, and gives the calculation process of forecast models clear causal correlation attributes. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the steps of a numerical forecasting method for ocean wind energy that incorporates artificial intelligence, as described in this invention. Figure 2 This is a flowchart of multi-scale feature learning based on a deep spatiotemporal feature extraction network; Figure 3 This is a graph showing the change in loss during the physical constraint training process; Figure 4 A graph showing the relationship between large-scale circulation index and wind speed; Figure 5 The figure shows the results of quantifying the uncertainty of ocean wind energy ensemble forecast. Detailed Implementation

[0017] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] See Figure 1 The process involves acquiring initial multi-source heterogeneous data for the target sea area and assimilating and fusing this data to generate a standardized historical and current meteorological and hydrological field. Based on a deep spatiotemporal feature extraction network, multi-scale feature learning is performed on the standardized historical and current meteorological and hydrological field to capture the long- and short-term spatiotemporal evolution patterns of wind energy-related parameters. A physical constraint module is used to transform known fundamental physical laws of ocean-atmospheric motion into loss function terms, constraining the learning process of the deep spatiotemporal feature extraction network, thereby generating a physically enhanced forecast model. A causal intervention mechanism is employed to simulate the intervention effects of different atmospheric circulation backgrounds on the local wind field in the physically enhanced forecast model, constructing a causally interpretable wind energy numerical forecast model. Real-time multi-source observation data is input into the causally interpretable wind energy numerical forecast model to extrapolate a multi-level wind energy parameter forecast sequence for the target sea area over a future period. Uncertainty quantification is performed on the multi-level wind energy parameter forecast sequence to generate a probability forecast product with confidence intervals.

[0019] In one embodiment of the present invention, information is acquired from multiple observation and simulation platforms. For example, for a target sea area in the East China Sea, the collected satellite remote sensing data includes wind field data from the GCOM-W1 satellite and sea surface temperature data from the GOES-R series satellites. Buoy data comes from minute-level wind speed, wind direction, air temperature, and air pressure recorded by several marine meteorological buoys deployed in the sea area. Shore-based anemometer data comes from 100-meter-high anemometer towers erected on coastal islands. Reanalysis data uses the ERA5 reanalysis dataset provided by the European Centre for Medium-Range Weather Forecasts (ECMWF). Numerical model data is taken from the output field of the US National Environmental Prediction Center's Global Forecast System. These data together constitute the initial multi-source heterogeneous data. In some embodiments, the satellite remote sensing data may also include high-resolution sea surface wind fields retrieved from synthetic aperture radar.

[0020] In practical implementation, missing and outlier values ​​in the initial multi-source heterogeneous data are processed using a spatiotemporal kriging interpolation-based imputation and correction method. This method establishes a local sample set for each missing or suspected outlier data point, including spatially neighboring points and temporally preceding and subsequent observations. The spatiotemporal variogram model considers both spatial distance and temporal lag. For continuous missing periods in the buoy data sequence caused by equipment failure, the spatiotemporal kriging interpolation-based imputation and correction method reconstructs a complete and reasonable numerical sequence for that period using spatially neighboring buoy data and satellite transit data within the same time period. It can be understood that the data output by the spatiotemporal kriging interpolation-based imputation and correction method is continuous and smoothly transitioning in the spatiotemporal domain, resulting in a complete and continuous multi-source dataset.

[0021] In practice, complete and continuous multi-source datasets are uniformly interpolated onto a preset spatial latitude and longitude grid and standard barospheres. The preset spatial grid covers longitudes from 120°E to 128°E and latitudes from 24°N to 30°N, with a horizontal resolution of 0.1° × 0.1°. The standard barospheres are selected as 1000 hPa, 925 hPa, 850 hPa, and 700 hPa. Satellite remote sensing sea surface wind field and sea surface temperature data are mapped from their irregular grids to a unified 0.1° latitude and longitude grid using a bilinear interpolation algorithm. Single-point observation data from buoys and anemometer towers are fused with the background field using a three-dimensional variational assimilation method to generate gridded data matching the target grid. Numerical model and reanalysis data are interpolated onto the 0.1° high-resolution target grid using a conservative remapping method, achieving consistency in spatiotemporal resolution between data from different sources.

[0022] In practice, the standardized data undergoes deviation correction, using high-precision field observation data as a benchmark. The observations from the buoy and shore-based wind towers are used as the baseline true values, and the interpolated satellite data and reanalysis data from the same time and location are compared. The systematic deviation of the satellite wind speed data relative to the buoy wind speed is calculated; this deviation is expressed as a function related to the wind speed magnitude. Optionally, one form of the deviation correction function is: , in: The representative satellite inversion wind speed value, This represents the background field average wind speed within a certain spatiotemporal window. These are the correction coefficients determined through regression analysis. For the pressure data in the reanalysis data, an empirical method of adding or subtracting a fixed constant is used to eliminate mean bias and systematic bias.

[0023] In practice, the bias-corrected data is arranged chronologically, with hourly intervals in the time dimension and covering all points of the target latitude and longitude grid in the spatial dimension. The variable dimensions include the u and v components of wind speed, wind direction, air temperature, air pressure, sea surface temperature, and significant wave height. In a specific example, the time series begins at 00:00 UTC on January 1, 2015, and ends at 23:00 UTC on December 31, 2024. Each time point corresponds to a three-dimensional spatial data volume containing all variable fields. These chronologically arranged three-dimensional data volumes can be understood as constituting a four-dimensional data field, where the four dimensions are longitude, latitude, barometric strata, and time. This data volume represents the standardized historical and current meteorological and hydrological field. In some embodiments, wave height data is provided only at the surface level, which is a single layer in the barometric strata dimension.

[0024] In one embodiment of the present invention, see [reference] Figure 2 Constructing a deep spatiotemporal feature extraction network requires integrating different types of neural network layers. The constructed deep spatiotemporal feature extraction network is a combined network including a convolutional neural network, a long short-term memory network, and a self-attention mechanism. The convolutional neural network part adopts an encoder structure, consisting of four convolutional layers. Each convolutional layer is followed by a batch normalization layer and a modified linear unit activation function. The convolutional layers use 3×3 convolutional kernels with a stride of 1. Simultaneously, 2×2 max pooling layers are used to progressively downsample the feature map to expand the receptive field. Finally, multi-scale spatial feature maps are extracted from the input standardized historical and current meteorological and hydrological fields. In some embodiments, the last layer of the convolutional neural network can output a feature map with 128 channels, reducing the spatial size to one-quarter of the original input.

[0025] In practice, standardized historical and current meteorological and hydrological fields are input into the convolutional neural network of the deep spatiotemporal feature extraction network. The dimensions of the input data are [batch size, time step, height, width, number of variables]. The convolutional neural network independently extracts spatial features from data slices at each time step. For a standardized meteorological and hydrological field describing a target sea area, its spatial dimensions are 60 grid points in height, 80 grid points in width, and 7 variables. After four layers of convolution and pooling operations, the dimensions of the output multi-scale spatial feature map become [batch size, time step, 15, 20, 128]. This multi-scale spatial feature map captures spatial structure information from local eddies to weather system scales. It can be understood that each spatial location in the multi-scale spatial feature map corresponds to a feature representation of a 4×4 region on the original latitude and longitude grid.

[0026] In practice, the multi-scale spatial feature map is unfolded along the time dimension and input into the Long Short-Term Memory (LSTM) network part of the deep spatiotemporal feature extraction network to learn the dynamic dependencies of each spatial feature point in the time series. Specifically, the multi-scale spatial feature map output by the convolutional neural network is processed according to the location of each spatial grid point, and the features of each point are extracted along the time dimension into a time series vector. For a multi-scale spatial feature map with dimensions [batch size, time step, 15, 20, 128], it is rearranged into a tensor of [batch size, 15*20, time step, 128]. Thus, each of the 300 spatial locations has a time series vector with a length equal to the time step and a feature dimension of 128. The time series vector corresponding to each spatial grid point is used as an independent input and fed into the LSM network unit of the deep spatiotemporal feature extraction network with shared weights. This means that all 300 spatial points use the same set of LSM network parameters to process their respective temporal information. In a Long Short-Term Memory (LSTM) network unit, information is selectively updated, forgotten, and output through input gates, forget gates, and output gates, learning the temporal evolution of each spatial feature point. The input gate controls how much new information is incorporated into the cell state at the current time step, the forget gate determines how much past cell state information is retained, and the output gate calculates the output feature for that time step based on the current cell state. The updated feature vectors of each point output by the LSM network unit are rearranged according to the original spatial structure to form a feature map containing temporal dynamic information, with its dimensions restored to [batch size, time step size, 15, 20, hidden layer dimension].

[0027] In practical implementation, the self-attention mechanism of the deep spatiotemporal feature extraction network is used to perform global correlation analysis on the temporal features output by the Long Short-Term Memory network. The input to the self-attention mechanism is the feature map output by the Long Short-Term Memory network, which is first linearly transformed into a query vector, a key vector, and a value vector. Attention weights between any two spatial locations are calculated to capture the long-range spatial interactions between wind field changes at different geographical locations. Optionally, the calculation of attention weights can be expressed as: , in: Indicates position Position Attention weights It is a location The query vector, It is a location The key vector, It is the dimension of the key vector. This represents the total number of spatial locations. The attention weights are then weighted and summed with the value vectors to obtain a context vector for each location that incorporates global information. In some embodiments, a multi-head self-attention mechanism can be used to perform the above calculations multiple times in parallel, and then the outputs of the multiple heads can be concatenated and linearly transformed.

[0028] In practical implementation, the final output is an abstract feature vector that integrates multi-scale spatial, temporal dynamics, and long-range correlation features. The multi-scale spatial features extracted by the convolutional neural network, the temporal dynamics modeled by the long short-term memory network, and the long-range correlation information captured by the self-attention mechanism are fused and dimensionality reduced through a fully connected layer. This fully connected layer can be understood as flattening the three-dimensional feature map (time step, spatial height, spatial width, feature channels) and mapping it to a fixed-dimensional vector, such as a one-dimensional vector of length 512. This abstract feature vector is used to characterize the core evolutionary patterns of wind energy-related parameters and serves as the input to the subsequent physical constraint module. Optionally, the training objective of the deep spatiotemporal feature extraction network is to minimize the prediction error generated after its output abstract feature vector passes through the downstream prediction head network.

[0029] In one embodiment of the present invention, simplified forms of mass conservation, momentum conservation, and energy conservation applicable to the target sea area scale are selected from fluid dynamics and thermodynamic equations as the basis of physical constraints. For mesoscale marine weather forecasting, simplified momentum equations and continuity equations including Coriolis force and horizontal pressure gradient force are typically used, such as the governing equations under the horizontal non-divergent approximation or quasi-geostrophic approximation. These equations, after discretization, serve as the mathematical expression of the basis of physical constraints. For the target sea area in the East China Sea, whose spatial scale covers 120°E to 128°E and 24°N to 30°N, with a horizontal resolution of 0.1 degrees, it falls within the scope of mesoscale marine weather forecasting. At this scale, the horizontal motion scale of the atmosphere is much larger than the vertical motion scale; therefore, the continuity equation under the horizontal non-divergent approximation is selected as the specific form of the mass conservation constraint. The expression of this equation in the local rectangular coordinate system is: ; in and These represent the components of wind speed in the east-west and north-south directions, respectively. and Let be the corresponding spatial coordinates. For the momentum conservation constraint, a simplified horizontal momentum equation considering the balance between the Coriolis force and the horizontal pressure gradient force is adopted, which takes the form: ; as well as: ; in Coriolis parameters, air density, Let be the air pressure. For the energy conservation constraint, a simplified thermodynamic energy equation is adopted, neglecting the release of latent heat of phase change. Its form is: ; in For temperature, For time. The above set of equations together constitutes the basis of specific and differentiable physical constraints applicable to the target sea area scale.

[0030] In some embodiments, for boundary layer wind field forecasting, the physical constraint basis can be a simplified form of the boundary layer momentum equation and thermodynamic energy equation that includes a turbulence parameterization scheme.

[0031] In practical implementation, a differentiable mapping network is constructed to map the abstract feature vector output by the deep spatiotemporal feature extraction network back to the physical variable field. The differentiable mapping network consists of three fully connected layers, with the last layer being a linear output layer. The input of the differentiable mapping network is the abstract feature vector output by the deep spatiotemporal feature extraction network, such as a one-dimensional vector of length 512. The output of the differentiable mapping network is two-dimensional spatial grid data of the horizontal u-component field of wind speed, the horizontal v-component field of wind speed, and the pressure and temperature fields. The output dimension is consistent with the spatial dimension of the corresponding variables in the standardized historical and current meteorological and hydrological fields. Using the last linear transformation layer of the differentiable mapping network, the high-dimensional abstract features are directly mapped to the horizontal component of wind speed, and the two-dimensional spatial fields of pressure and temperature. For the mapped horizontal component field of wind speed, the horizontal divergence field and vertical vorticity field are calculated on the preset spatial grid using the differentiable central difference operator. The central difference operator is implemented through a convolution operation with a kernel of [1,0,-1]. For the mapped pressure field, its gradient fields in the horizontal and vertical directions are calculated. The horizontal gradient field is calculated using a differentiable central difference operator, while the vertical gradient field is calculated using the difference between pressure layers. For the mapped temperature field, its horizontal gradient field is calculated. The calculated divergence field, vorticity field, and gradient field together constitute the physical variable gradient or divergence field used for physical constraint calculations.

[0032] In practical implementation, based on the physical constraints, the residuals of the mapped physical variable fields on the continuity equations and motion equations are calculated. When the law of conservation of mass is selected as one of the physical constraints, for a horizontal wind field, the simplified form of its continuity equation can be expressed as the horizontal wind field divergence being zero. The corresponding residual operator R can be defined as the calculated horizontal wind speed field divergence value. The physical conservation loss term is the average of the squares of this divergence value at all calculation points. The horizontal component of wind speed, pressure field, temperature field, and their gradient divergence field in the physical variable gradient or divergence field are substituted into the selected simplified physical equation set. It can be understood that for the mass conservation constraint, the divergence of the horizontal wind field is calculated as the residual term. For the momentum conservation constraint, the calculated local acceleration, Coriolis force term, pressure gradient force term, and wind field advection term are substituted into the equation, and the difference between the left and right sides of the equation is calculated as the momentum residual. The physical conservation loss term is the norm of these residual terms, for example, using the root mean square error form. One way to calculate the physical conservation loss term can be expressed as: , in: Representing the physical conservation loss term, K is the total number of spatial grid points in the computational domain. It is the horizontal wind speed vector at grid point a. It is the air pressure at grid point a. It is the temperature at grid point a. , , These represent the corresponding gradient or divergence fields, respectively. The residual operator, representing the fundamental equations of the physical constraints, outputs a residual vector whose L2 norm squared is used to calculate the loss. In some embodiments, the calculation can be performed separately for multiple standard pressure layers and then summed.

[0033] In practical implementation, the training loss function of the deep spatiotemporal feature extraction network combines a weighted combination of the physical conservation loss term and the conventional data fitting loss term to form the total loss function. The conventional data fitting loss term is the mean squared error between the predicted value and the actual observed value output by the deep spatiotemporal feature extraction network. Optionally, the total loss function can be defined as... ,in It is the data fitting loss term. It is a physical conservation loss term. It is a hyperparameter used to balance the weights of two items. The value is determined through cross-validation. It can be understood that the role of the physical conservation loss term is to guide the feature mapping relationship learned by the deep spatiotemporal feature extraction network, and the resulting physical variable field does not severely violate known physical laws.

[0034] In practical implementation, the total loss function is optimized through backpropagation, and the training process uses gradient-based optimization algorithms, such as the Adam optimizer. In each iteration, the gradient of the total loss function with respect to all trainable parameters of the deep spatiotemporal feature extraction network and the differentiable mapping network is calculated, and then the network parameters are updated. The physics-conserving loss term propagates the gradient through the differentiable mapping network and differentiable physics operators, ensuring that the evolutionary patterns learned internally by the deep spatiotemporal feature extraction network satisfy preset physical law constraints while fitting the data, thus obtaining a physically-enhanced prediction model. Optional, The value can start from a small value and gradually increase with the number of iterations during training. This allows the model to initially focus on fitting the data and gradually strengthen its adherence to physical constraints in later stages. During training, by setting appropriate hyperparameters and optimization algorithms, such as using a strategy of dynamically adjusting the weights of the physical constraint loss term, the total loss function can be ensured to decrease steadily. This enables the physically enhanced prediction model to effectively learn the evolutionary pattern that simultaneously satisfies data fitting and physical law constraints.

[0035] See Figure 3This is a graph showing the change in loss during the physical constraint training process, intuitively reflecting the optimization process of the model fitting data and satisfying physical constraints. The initial value of the physical conservation loss term is relatively high, decreasing the fastest in the early stages of training. As iterations progress, its value decreases significantly and eventually converges to a relatively small stable level. This reflects that the model gradually learns and satisfies physical laws such as mass conservation and momentum conservation, and the constraint effect of physical constraints significantly increases with training. The initial value of the data fitting loss term is approximately 2.08, decreasing steadily and stabilizing in the later stages. This represents the model's ability to fit real observation data, and the later convergence indicates that the model has captured the spatiotemporal evolution of the wind field quite well. The total loss function is a weighted combination of the physical conservation loss and the data fitting loss, showing a continuous downward trend and eventually converging to approximately 0.3~0.4. This reflects the model's balanced optimization result between "data fitting accuracy" and "physical rationality." Specifically, the constraints imposed by the physical constraint module are soft constraints. Its mechanism is not to force the physical conservation loss term to converge to zero to achieve precise satisfaction of the hard constraint, but rather to introduce the physical conservation loss term into the total loss function as a regularization guide signal. During the model parameter optimization process, the search direction is biased towards parameter configurations with lower physical violations under the same data fitting ability. When the data fitting loss term and the physical conservation loss term reach a dynamic equilibrium during joint optimization, the model is not simply optimizing data fitting as the main objective, but is in a Pareto optimal state of multi-objective learning. In this state, any operation that attempts to further reduce the data fitting loss will lead to a significant increase in the physical conservation loss, and vice versa. The principle of the soft constraint mechanism in weakening data noise interference is that the observation noise and random perturbations in the training data do not have compressible physical laws. When the model is constrained by the physical conservation loss, its network parameter updates will be guided away from those directions that generate non-physical feature mappings for fitting noise, so that the spatiotemporal evolution pattern learned by the model preferentially reflects the real signal components in the data that are consistent with the basic laws of ocean and atmospheric motion. The optimality criterion for the trade-off between data fitting and physical satisfaction under this soft constraint mechanism is defined as follows: During the model training process, the wind energy parameter forecast accuracy on the independent validation set is used as the final selection criterion. Within the iteration interval where the forecast accuracy on the validation set reaches the optimal level or no longer significantly improves, the model parameter configuration with the lowest physical conservation loss term is selected as the final model parameter.

[0036] In one embodiment of the invention, key large-scale circulation indices affecting the target sea area are identified. For example, for wind energy forecasting in the North Atlantic, key large-scale circulation indices may include the North Atlantic Oscillation Index (NAO), the East Atlantic Type Index (EATI), and the Scandinavian Type Index (STI). These circulation indices are typically derived from reanalysis data products from the Climate Research Centre and represent the characteristic state of large-scale pressure systems. Introducing key large-scale circulation indices as exogenous intervention variables into the input layer of a physically-enhanced forecasting model means that, in addition to standardized meteorological and hydrological field data, a vector containing multiple circulation index values ​​is additionally concatenated into the model input. In some embodiments, historical monthly average circulation indices can be selected as part of the input features. Referring to Table 1, some examples of large-scale circulation indices that may be used for forecasting in the North Atlantic are listed.

[0037] Table 1: Example Table of Large-Scale Circulation Indices Applicable to North Atlantic Wind Energy Forecasting

[0038] In practice, at the causal intervention layer of the physically-enhanced forecast model, virtual interventions of varying intensities are applied to the key large-scale circulation indices in the input. Specifically, before calculating the gradient during model forward propagation, the circulation index values ​​in the input vector are fixed to preset specific scenario values. For example, the North Atlantic Oscillation Index (NAO) is set to values ​​with standard deviations of +2, 0, and -2, while all other meteorological and hydrological variables in the input vector remain unchanged. To ensure the physical consistency of other meteorological and hydrological variables when setting the virtual intervention for the circulation index, it is not simply a matter of keeping all local variables unchanged. Instead, a conditional generation network, pre-learned from historical reanalysis data, simultaneously generates physically coordinated local meteorological field adjustments while applying specific intervention values ​​to the circulation index. Specifically, the conditional generation network takes the target circulation index intervention value and the original local meteorological field as inputs. Through an encoder-decoder structure, it outputs the adjustment amounts of meteorological variables at each spatial grid point. The encoder consists of three convolutional layers and two fully connected layers, while the decoder consists of corresponding deconvolutional layers and upsampling layers. The network is pre-trained on historical reanalysis data with the training objective of minimizing the L2 reconstruction error between the generated field and the true field, and the sum of the residuals between the adjusted local field and the mass conservation and hydrostatic equilibrium equations. During intervention, the conditional generation network is responsible for physically self-consistently covariantly adjusting the local meteorological field based on large-scale circulation changes. In each virtual intervention, the adjustment amount is superimposed on the original local input variables, ensuring that the input state constructed by the causal intervention layer physically satisfies the variable coupling constraints implicit in the atmospheric motion equations. This operation can be understood as performing a "counterfactual" setting at the model's input, simulating the possible impact on forecast results when the large-scale circulation state is in different phases under the same weather background. The difference between counterfactual intervention and conventional sensitivity analysis lies in the fact that sensitivity analysis only measures the magnitude of the model output's response to input disturbances, without concern for the physical reality of the disturbance itself or the preservation of coupling relationships between variables. By guiding the covariant adjustment of the local meteorological field with physical constraints during virtual intervention, each set of counterfactual input states satisfies variable field configuration relationships compatible with actual atmospheric motion patterns. Furthermore, counterfactual intervention is only performed during the training phase to learn the strength of causal effects and encode them into the model weights. The model input during the inference phase is real observational data, eliminating the need for any further variable intervention or forward propagation of physically enhanced forecast models with inconsistent input states. To separate the direct effect of the circulation index on the wind field and the indirect effects generated through other variables from the total causal effect under different intensities of virtual intervention, two sets of intervention calculations are run in parallel during the causal intervention process. The first set of interventions only changes the circulation index value, while simultaneously using the aforementioned conditional generation network to covariantly adjust the local meteorological field. The resulting total effect includes the direct effect of the circulation index on the wind field and the indirect contributions generated through adjustments to other variables.The second intervention, after setting the same circulation index value and simultaneously adjusting other meteorological variables, additionally maps intermediate feature variables unrelated to the circulation index to a subspace orthogonal to the circulation index before entering the feature extraction network of the physically-enhanced forecast model. This mapping is achieved through an orthogonal projection layer, which is optimized during the training phase to maximize the Hilbert-Schmidt independence criterion between the projected features and the intervention value of the circulation index, ensuring that subsequent network layers cannot recover the circulation index information from this subspace. The difference between the forecast values ​​output by the two interventions represents the direct causal effect of the circulation index on the wind energy parameters of the target sea area. The final wind energy parameter forecast values ​​output by the model differ. For the same input sample, by running the model multiple times, each time changing only the circulation index value, a set of wind speed, wind direction, or wind power density forecast fields corresponding to different circulation backgrounds is obtained.

[0039] In practical implementation, the rate of change of wind energy parameter forecast values ​​relative to the intervention intensity is analyzed to quantify the causal effect of key large-scale circulation indices on wind energy parameters in the target sea area. For a target grid point, when the North Atlantic Oscillation Index changes from -2 to +2, the 72-hour forecast wind speed at that point changes from 8.5 m / s to 10.2 m / s. The total wind speed change is 1.7 m / s, and the range of the intervention intensity change is 4 standard deviations. Optionally, the causal effect intensity can be measured by the change in forecast values ​​caused by a unit change in intervention intensity. One quantification formula is: , in: Indicates the strength of causal effect. and These represent the values ​​at which the circulation index intervention value is... and The predicted wind energy parameters output by the time model. and This represents two different intervention intensities. Encoding the learned causal effect strength into the model weights can be achieved by utilizing the causal effect strength... The weights of the fully connected layer that produces the final forecast output of the model are modulated so that the weight matrix of this layer is affected when processing features related to the circulation index. The constraints allow the model to explicitly consider the causal effects of large-scale circulation when making forecasts, resulting in a wind energy numerical forecasting model with causal interpretability. In the above formula, and Defined as the spatial average of the future wind speed scalar forecast values ​​for all spatial grid points within the target sea area at a specified forecast lead time, the numerator is the difference between two scalars, thus resolving the dimensionality matching problem when the forecast values ​​are in the spatiotemporal field. The denominator represents the difference in intervention intensity. When the value approaches zero, this invention uses a method of setting a minimum intervention intensity difference threshold of 0.01 standard deviation units to ensure numerical stability. If the actual intervention intensity difference is less than this threshold, the causal effect intensity of the grid point or the predicted variable is directly calculated. Defined as zero. The applicability of this linear approximation formula lies in the fact that the changes in model output caused by different intervention scenarios (such as the circulation index within the range of -2 to +2 standard deviations) fall within the quasi-linear response range. By comparing model outputs under multiple intervention intensities in practical operation, this linear approximation holds when the output change maintains a relatively constant ratio with respect to the intervention intensity. Its error range can be calculated by measuring the changes at different intervention intensity intervals. The difference in values ​​is used for evaluation to ensure that the estimation error of the causal effect strength introduced by the linear approximation is within an acceptable level.

[0040] The constructed causal interpretability is explicitly defined as the explicit decomposition and quantification of the large-scale circulation intervention effect at the input-output response level, rather than a layer-by-layer analysis of all nonlinear mapping links within the deep learning model. The specific implementation mechanism is as follows: In the forecasting and extrapolation phase, the input data received by the causally interpretable wind energy numerical forecasting model consists of two concatenated parts: one is the spatiotemporal sequence data of the standardized meteorological and hydrological field, and the other is the key large-scale circulation index value for the corresponding time period. The deep spatiotemporal feature extraction network within the model performs nonlinear feature transformation on the meteorological and hydrological field data to generate abstract feature vectors, maintaining the inherent complex mapping characteristics of the deep learning model. In the output phase, the gated attention module receives the large-scale circulation index value for the current time period as input. This module consists of an input layer, a fully connected hidden layer containing thirty-two neurons, a hyperbolic tangent activation function, and a linear output layer connected sequentially, outputting a set of modulation weights with the same dimension as the wind energy parameter forecast. The modulation weights are multiplied element-wise with the baseline forecast features output by the fully connected forecast layer of the deep spatiotemporal feature extraction network to obtain the final wind energy parameter forecast value. Through this architecture, the forecast value at any grid point at any time can be explicitly decomposed into two traceable components: the first component is the baseline forecast component, generated by the nonlinear characteristic transformation of historical meteorological and hydrological fields; the second component is the causal modulation component, determined by the product of the modulation weights generated by the current large-scale circulation index through the gated attention module and the baseline forecast features. The two components can be calculated and stored numerically separately, making the source of each contribution to the forecast result clearly identifiable. When the large-scale circulation index changes significantly at different times, for example, when the intensity index of the western Pacific subtropical high changes from a positive anomaly of one standard deviation to a negative anomaly, the modulation weights generated by the gated attention module change accordingly, and the value of the causal modulation component changes accordingly, ultimately reflecting a responsive adjustment in the wind energy parameter forecast sequence. The entire adjustment process, from the input of the circulation index to the generation of modulation weights and the output of forecast values, is a deterministic forward computation. The values ​​of each intermediate variable can be recorded and traced back, thus forming a traceable causal logical chain from changes in atmospheric circulation background to the output of wind energy parameter forecasts. This chain does not require physical meaning labeling for every dimension of the feature transformation within the deep spatiotemporal feature extraction network. Instead, by establishing an explicit modulation branch driven by the circulation index at the output, the influence of large-scale circulation is separated from the data-driven nonlinear feature transformation, allowing the intensity, direction, and time-varying characteristics of this influence to be independently quantified, verified, and traced.

[0041] Specifically, the learned causal effect strength is encoded into the model weights through a learnable gated attention module. This module takes the calculated causal effect strength as input and generates a set of modulated weights. The learnable gated attention module is structured as a two-layer fully connected network. The first layer contains 32 neurons and uses the hyperbolic tangent activation function, while the second layer is a linear layer whose output neuron count is identical to that of the fully connected layer responsible for the final wind energy parameter forecast in the physical information augmentation prediction model. Causal effect strength As a scalar input to this module, it is first embedded into a 32-dimensional feature space through a linear transformation, then nonlinearly mapped using a hyperbolic tangent function, and finally a set of modulation weights with the same dimension as the forecast layer's output is generated by the linear layer. The modulation weights are fused with the original fully connected layer weights of the model through element-wise multiplication. That is, when the model performs forward propagation to calculate the forecast value, the original output feature vector of the forecast layer is multiplied element-wise with this set of modulation weights to obtain the final forecast features modulated by causal effects. In the training process, this gated attention module is jointly trained end-to-end with the entire physically enhanced forecast model. Its training objective function is the total loss function, which is a weighted combination of data fitting loss and physical conservation loss. The training data organization is completely consistent with the main model, using sequence samples of standardized meteorological and hydrological fields. In this way, the gated attention module learns how to transform different causal effect intensities into appropriate modulation of the model's forecast behavior.

[0042] This set of modulation weights is then applied to the weights or features of the neural network layer responsible for outputting wind energy parameter forecasts in the physically-enhanced forecasting model, thereby achieving explicit modulation of the model's forecasting behavior. In this way, the model can adaptively adjust its internal parameters based on the current input large-scale circulation index value during forecasting, explicitly considering and reflecting the causal influence of large-scale circulation on wind energy parameters in the target sea area.

[0043] In practice, the latest real-time observation data from satellites, buoys, and wind towers are received and assimilated, interpolated, and corrected for bias to generate a real-time initial field. The data processing method used is consistent with that used when generating standardized historical and current meteorological and hydrological fields. A pre-defined length of historical data, for example 120 hours, is extracted from the generated standardized historical and current meteorological and hydrological fields, prior to the real-time moment. This historical data contains 120 consecutive time slices, each spaced one hour apart. For each time slice of the extracted historical data, its consistency with the real-time initial field in terms of spatial grid, variable definitions, and units is checked. It is understood that since both historical and real-time data originate from standardized historical and current meteorological and hydrological fields or are generated using the same processing flow, they are consistent in spatial grid, variable definitions, and units, ensuring that each time slice of the extracted historical data can be directly compared and spliced ​​with the real-time initial field. The real-time initial field is then appended as a new time slice to the end of the historical time slice sequence and spliced ​​along the time dimension to form a complete sequence of 121 time slices. In some embodiments, it is necessary to ensure that the last moment of the historical time slice is continuous with the moment of the real-time initial field before stitching. The stitched complete sequence is then normalized along the time axis. The normalization parameters use the mean and standard deviation of the historical data portion to eliminate magnitude differences between different time points. Finally, a complete, spatiotemporally continuous, and normalized input sequence from the historical data to the current moment is output, which can be used for forecasting and extrapolation by a causally interpretable wind energy numerical prediction model.

[0044] In practice, the complete input sequence is fed into a causally interpretable wind energy numerical prediction model. The model uses the output of the previous time step as part of the input for the next time step, performing multi-time-step iterative forward extrapolation. Causal intervention is continuously applied during the iterative extrapolation process. Specifically, when the model extrapolates the forecast value for each future time step, its input includes the key large-scale circulation index value for the current time step. This circulation index value is provided by an externally provided circulation index forecast or persistent forecast. At each extrapolation time step, the gated attention module of the causal intervention layer dynamically generates a set of modulation weights based on the current time step circulation index value and fuses them with the weights of the model's forecast layer, achieving real-time causal modulation of the forecast behavior for the current time step. By injecting large-scale circulation constraints into each step of the autoregressive extrapolation, the effectiveness of causal intervention is ensured not to diminish with the number of iterations, allowing the causal influence of large-scale circulation to persist throughout the entire forecast lead time. The model receives a complete input sequence containing historical background and real-time status, and simultaneously receives the key large-scale circulation index value for the current moment. The model first outputs the predicted wind speed, wind direction, and wind power density at each grid point within the target sea area for the first future time step. Then, this first-time-step forecast output is combined with the latest data from the complete input sequence and the future circulation index forecast to form a new input for extrapolating the forecast values ​​for the second time step. This iterative process is repeated until all time steps for the preset forecast lead time are completed. During the iterative extrapolation process, the model outputs the predicted wind speed, wind direction, and wind power density at each grid point within the target sea area for each extrapolated time step. The outputs from all forecast time steps are then aggregated to form a multi-level wind energy parameter forecast sequence covering wind speed, wind direction, and wind power density for a future period.

[0045] See Figure 4 This is a large-scale circulation index and wind speed relationship graph, showing the linear relationship between three North Atlantic circulation indices and wind speed in the target sea area. It represents a typical analytical result of the "causal intervention mechanism" in numerical weather prediction of ocean wind energy, used to quantify the impact of large-scale circulation on local wind fields. All three circulation indices are significantly positively correlated with wind speed; the higher the index value, the greater the corresponding wind speed. The North Atlantic Oscillation Index (NATO) has the strongest impact on wind speed, with the largest increase in wind speed per unit change in index, making it the dominant circulation factor driving local wind field changes. The East Atlantic Oscillation Index (EAO) has the weakest impact, while the Scandinavian Oscillation Index (SIO) has a moderate impact. Throughout the entire index range from -2 to 2, the three curves maintain a parallel upward trend, indicating that the linear regulation effect of different circulation types on wind speed is stable and predictable. The wind energy resource potential under different circulation backgrounds can be assessed based on the seasonal / interannual forecasts of the circulation indices.

[0046] In one embodiment of the present invention, during the training phase of the causally interpretable wind energy numerical prediction model, the error distribution of each forecast on the validation set is saved. The validation set is a portion of time-series data extracted from standardized historical and current meteorological and hydrological fields that does not participate in model parameter training. After the model training is completed, uncertainty quantification is performed during the forecast phase. The causally interpretable wind energy numerical prediction model used in this phase is the final forecast model with fixed parameters, constructed through the aforementioned physical constraint training and causal intervention mechanism. Uncertainty quantification aims to evaluate the forecast stability of the trained model in the face of inherent uncertainties in the input data. During each forecast validation, the causally interpretable wind energy numerical prediction model is used to iteratively forecast a sample in the validation set for the next 72 hours. The forecast results are compared with the actual observations, and the forecast error for each grid point, each forecast time step, and each variable (such as the U and V components of wind speed) is calculated. It can be understood that the forecast error is the forecast value minus the actual observation value. Error values ​​from all samples, all grid points, and all forecast time steps on the entire validation set are collected to form a large error sample set. An empirical distribution function for the model forecast error is then constructed based on this set. This empirical distribution function describes the frequency of the forecast error within different value ranges. Optionally, for wind speed forecast errors, the empirical distribution function can be expressed as: , in: Indicates that the error is less than or equal to The empirical probability, It represents the total number of error samples on the validation set. It is the first One error sample value, It is an indicator function that takes the value 1 when the condition inside the parentheses is true, and 0 otherwise.

[0047] In practical implementation, an ensemble forecasting approach is adopted in the forecasting phase, adding random perturbations that conform to the statistical characteristics of historical errors to the real-time initial field. The statistical characteristics of the random perturbations are estimated from the empirical distribution function of the model forecast errors constructed during the training phase, such as the mean, variance, and spatial correlation structure of the estimated errors. A set of ensemble member initial fields is generated, and the number of ensemble members can be set to 50. The specific number of ensemble members is determined according to the required probability forecast confidence level. In this embodiment, a statistical rule of thumb is adopted, namely, to stably estimate the quantiles near the upper and lower boundaries of the 90% confidence interval, at least 50 to 100 ensemble members are required. Therefore, 50 members are set as the benchmark configuration to balance computational efficiency and interval estimation stability. When extracting the perturbation vector from the multidimensional normal distribution, the dimension of this distribution is equal to the total number of degrees of freedom of the variables that need to be perturbed in the real-time initial field, that is, equal to the number of spatial grid points multiplied by the number of physical variables involved in the perturbation. In this embodiment, it is 15 multiplied by 20, totaling 300 grid points, multiplied by the wind speed. Quantity and The two variables form a 600-dimensional vector space. The spatial covariance matrix of this multidimensional normal distribution is an empirical covariance matrix obtained by maximum likelihood estimation on the error sample set during the training phase, and processed using the Tikhonov regularization method, that is, adding an identity matrix multiplied by a small regularization coefficient to the covariance matrix to ensure its invertibility and numerical stability. The perturbation amplitude does not change with the forecast lead time because this method constructs the perturbation based on the statistical characteristics of the analysis error of the initial field. Each set member independently extrapolates from the initial time along deterministic physical laws and model dynamics, and its forecast divergence naturally increases from the dynamic process of the model itself, without the need for manually setting the time-varying perturbation amplitude.

[0048] For each grid point variable in the real-time initial field, a perturbation vector is randomly drawn from a multidimensional normal distribution with a mean of 0 and a variance equal to the empirical variance of the variable's error, taking into account the spatial correlation between grid points. In some embodiments, the spatial correlation structure is described by the spatial covariance matrix of the error samples. This random perturbation vector is added to the original real-time initial field to obtain a set membership initial field. This process is repeated 50 times to generate 50 statistically equivalent set membership initial fields, each with different specific values.

[0049] In practice, the initial field of each ensemble member is input into a causally interpretable wind energy numerical prediction model for independent iterative extrapolation. The causally interpretable wind energy numerical prediction model performs the same extrapolation process for each ensemble member's initial field as for a single-point deterministic forecast, including the processing of the causal intervention layer and iterative forward calculation. Because the initial fields of each ensemble member differ, the future atmospheric state evolution path extrapolated by the model based on each initial field will also differ. After independently extrapolating the initial fields of 50 ensemble members, 50 sets of wind energy parameter forecast sequences are obtained. These 50 sets of forecast sequences constitute a wind energy parameter forecast sequence set. This set reflects the degree of dispersion in the forecast results caused by the uncertainty of the initial fields.

[0050] In practice, the statistical quantiles of all ensemble members' forecast values ​​are calculated for each forecast time step and each spatial grid point. For the wind speed forecast at the grid point of 125.5°E and 28.5°N for the next 24 hours, 50 wind speed forecast values ​​from 50 ensemble members at that time and location are collected. These 50 values ​​are arranged in ascending order. The high and low quantiles are selected as the upper and lower bounds, respectively. For example, the value ranked 40th (80th quantile) is selected as the upper bound of the confidence interval, and the value ranked 10th (20th quantile) is selected as the lower bound. In some embodiments, a 90% confidence interval can also be calculated, i.e., the 5th and 95th quantiles are used as the lower and upper bounds, respectively. A confidence interval is added to each value of the original multi-level wind energy parameter forecast sequence, which is usually given by the mean or median of the ensemble members. In the final forecast product, each grid point, each time period, and each variable (such as wind speed) has a predicted value with a possible range of values, forming a probabilistic forecast product. Optionally, the probabilistic forecast product can be graphically displayed as a "spaghetti plot" or a probability-filled plot to visually represent the uncertainty of the forecast.

[0051] See Figure 5This is a graph showing the results of uncertainty quantification in ocean wind energy ensemble forecasts, illustrating the impact of the number of ensemble members on the forecast confidence interval and coverage. The average confidence interval width decreases steadily from 3.2 m / s to approximately 1.85 m / s as the number of ensemble members increases from 10 to 100, exhibiting a monotonically decreasing trend. The largest decrease occurs in the early stage (10-40), indicating that adding a small number of ensemble members can significantly compress the interval width; the decrease slows down in the later stage (60-100), with the marginal benefit gradually decreasing. Interval coverage steadily increases from 0.75 to approximately 0.89 as the number of ensemble members increases, maintaining a consistent upward trend. The rate of coverage growth is negatively correlated with the rate of interval width decline; when the interval width compresses rapidly, the increase in coverage is more pronounced; as the width decline slows down in the later stage, the growth in coverage also tends to stabilize.

[0052] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A numerical forecasting method for ocean wind energy combining artificial intelligence, characterized in that, Includes the following steps: Acquire initial multi-source heterogeneous data of the target sea area, and assimilate and fuse the initial multi-source heterogeneous data to generate standardized historical and current meteorological and hydrological fields; Based on a deep spatiotemporal feature extraction network, multi-scale feature learning is performed on standardized historical and current meteorological and hydrological fields to capture the long- and short-term spatiotemporal evolution patterns of wind energy-related parameters. By using the physical constraint module, the known basic physical laws of ocean-atmosphere motion are transformed into loss function terms, which constrain the learning process of the deep spatiotemporal feature extraction network and generate a prediction model with enhanced physical information. By employing a causal intervention mechanism, the intervention effect of different atmospheric circulation backgrounds on local wind fields is simulated in a forecast model enhanced with physical information, thereby constructing a wind energy numerical forecast model with causal interpretability. Real-time multi-source observation data is input into a causally interpretable wind energy numerical prediction model to extrapolate a multi-level wind energy parameter prediction sequence for the target sea area over a future period. Uncertainty quantification is performed on multi-level wind energy parameter forecast sequences to generate probability forecast products with confidence intervals.

2. The numerical forecasting method for ocean wind energy combining artificial intelligence according to claim 1, characterized in that, The acquisition of initial multi-source heterogeneous data of the target sea area, and the assimilation and fusion processing of the initial multi-source heterogeneous data to generate standardized historical and current meteorological and hydrological fields, specifically includes: Historical and real-time data from satellite remote sensing, buoys, shore-based wind towers, reanalysis data, and numerical models were collected to form the initial multi-source heterogeneous data. Missing and outlier values ​​in the initial multi-source heterogeneous data are processed using a spatiotemporal kriging interpolation-based imputation and correction method to obtain a complete and continuous multi-source dataset. Complete and continuous multi-source datasets are uniformly interpolated onto a preset spatial latitude and longitude grid and standard pressure layer to achieve consistency in spatiotemporal resolution of data from different sources; The standardized data are corrected for deviations, using high-precision field observation data as a benchmark to eliminate systematic biases; The data after bias correction are arranged in chronological order to form a four-dimensional standardized historical and current meteorological and hydrological field containing multiple variables such as wind speed, wind direction, air temperature, air pressure, sea surface temperature, and wave height.

3. The numerical forecasting method for ocean wind energy combining artificial intelligence according to claim 2, characterized in that, The deep spatiotemporal feature extraction network performs multi-scale feature learning on standardized historical and current meteorological and hydrological fields to capture the long- and short-term spatiotemporal evolution patterns of wind energy-related parameters, specifically including: We construct a combined network that includes a convolutional neural network, a long short-term memory network, and a self-attention mechanism as a deep spatiotemporal feature extraction network; Standardized historical and current meteorological and hydrological fields are input into the convolutional neural network part of the deep spatiotemporal feature extraction network to extract multi-scale spatial feature maps from local to regional along the spatial dimension. The multi-scale spatial feature map is unfolded in the time dimension and input into the long short-term memory network part of the deep spatiotemporal feature extraction network to learn the dynamic dependency of each spatial feature point in the time series. By utilizing the self-attention mechanism of the deep spatiotemporal feature extraction network, a global correlation analysis is performed on the temporal features output by the long short-term memory network to capture the long-distance spatial interaction between wind field changes in different geographical locations. The final output is an abstract feature vector that integrates multi-scale spatial, temporal dynamics and long-range correlation features, which is used to characterize the core evolution mode of wind energy-related parameters.

4. The numerical forecasting method for ocean wind energy combining artificial intelligence according to claim 3, characterized in that, The physical constraint module transforms known fundamental physical laws of ocean-atmospheric motion into loss function terms, constraining the learning process of the deep spatiotemporal feature extraction network. Specifically, this includes: Simplified forms of mass conservation, momentum conservation, and energy conservation applicable to the target sea area scale are selected from the fluid mechanics and thermodynamics equations as the basis of physical constraints. The abstract feature vectors output by the deep spatiotemporal feature extraction network are mapped back to the gradient or divergence fields of physical variables such as wind speed, air pressure, and temperature through a differentiable physical computation graph. Based on the physical constraints, the residuals of the physical variable fields obtained by mapping on the continuity equation and the equation of motion are calculated, and the residual norm is used as the physical conservation loss term. In the training loss function of the deep spatiotemporal feature extraction network, the physical conservation loss term and the conventional data fitting loss term are weighted and combined to form the total loss function. By optimizing the total loss function through backpropagation, the deep spatiotemporal feature extraction network can learn evolution patterns that satisfy preset physical law constraints while fitting data, thereby obtaining a prediction model with enhanced physical information.

5. The numerical forecasting method for ocean wind energy combining artificial intelligence according to claim 4, characterized in that, The aforementioned causal intervention mechanism simulates the intervention effects of different atmospheric circulation backgrounds on local wind fields in a physically enhanced forecast model, specifically including: Identify key large-scale circulation indices that affect the target sea area and introduce them as exogenous intervention variables into the input layer of the physical information-enhanced forecasting model; In the causal intervention layer of the model, virtual interventions of different intensities are applied to the key large-scale circulation index of the input, that is, their values ​​are fixed for a specific scenario, while keeping other input variables unchanged; A forward propagation-enhanced physical information forecasting model calculates the differences in wind energy parameter forecasts output by the model under different intensities of virtual intervention. The rate of change of the difference in wind energy parameter forecasts relative to the intensity of intervention was analyzed to quantify the causal effect of key large-scale circulation indices on wind energy parameters in the target sea area. The learned causal effect strength is encoded into the model weights, enabling the model to explicitly consider the causal influence of large-scale circulation when making forecasts, thus forming a wind energy numerical forecast model with causal interpretability.

6. The numerical forecasting method for ocean wind energy combining artificial intelligence according to claim 5, characterized in that, The process of inputting real-time multi-source observation data into a causally interpretable wind energy numerical prediction model to extrapolate a multi-level wind energy parameter forecast sequence for the target sea area over a future period specifically includes: It receives the latest real-time observation data from satellites, buoys, and wind towers, and performs assimilation, interpolation, and bias correction on the data to generate a real-time initial field. The real-time initial field is spliced ​​with the standardized historical and current meteorological and hydrological fields of historical periods to form a complete input sequence that includes historical background and real-time status; The complete input sequence is fed into a causally interpretable wind energy numerical prediction model. The model uses the output of the previous time step as part of the input of the next time step to perform multi-time step iterative forward extrapolation. During the iterative simulation, the model outputs the predicted values ​​of wind speed, wind direction, and wind power density at each grid point within the target sea area for each time step. By summarizing the outputs of all forecast time steps, a multi-level wind energy parameter forecast sequence covering wind speed, wind direction, and wind power density is formed for a period of time in the future; The process involves splicing the real-time initial field with standardized historical and current meteorological and hydrological fields from historical periods to form a complete input sequence that includes historical background and real-time status. Specifically, this includes: From the generated standardized historical and current meteorological and hydrological fields, a preset length of historical data is extracted before the real-time moment, and the historical data contains multiple consecutive time slices. For each time slice of the extracted historical data, check its consistency with the real-time initial field in terms of spatial grid, variable definition and units to ensure that each time slice of the extracted historical data can be directly compared and stitched with the real-time initial field. The real-time initial field is treated as a new time slice and appended to the end of the historical time slice sequence, and then spliced ​​together in the time dimension. The spliced ​​complete sequence is normalized along the time axis to eliminate the magnitude differences between different time points; The final output is a complete, spatiotemporally continuous and standardized input sequence from the past to the present, which is used for forecasting and extrapolation by a wind energy numerical prediction model with causal interpretability.

7. The numerical forecasting method for ocean wind energy combining artificial intelligence according to claim 6, characterized in that, The process of quantifying the uncertainty of multi-level wind energy parameter forecast sequences to generate probability forecast products with confidence intervals specifically includes: During the training phase of a wind energy numerical prediction model with causal interpretability, the error distribution of each prediction on the validation set is saved to construct an empirical distribution function of the model's prediction error. In the forecasting phase, the ensemble forecasting concept is adopted, which adds random perturbations that conform to the statistical characteristics of historical errors to the real-time initial field to generate a set of ensemble member initial fields; Each set member's initial field is input into a causally interpretable wind energy numerical prediction model, and independent iterative deductions are performed to obtain a set of wind energy parameter prediction sequences. For the forecast sequence set, at each forecast time step and at each spatial grid point, calculate the statistical quantiles of the forecast values ​​of all set members; By selecting specific high and low quantiles as upper and lower bounds, a confidence interval is added to each value of the original multi-level wind energy parameter forecast sequence to form a probability forecast product.

8. The numerical forecasting method for ocean wind energy combining artificial intelligence according to claim 7, characterized in that, The process of unfolding multi-scale spatial feature maps along the time dimension and inputting them into the long short-term memory network portion of a deep spatiotemporal feature extraction network to learn the dynamic dependencies of each spatial feature point over time specifically includes: The multi-scale spatial feature map output by the convolutional neural network is used to extract the features of each point along the time dimension into a time series vector according to the position of the spatial grid points. The time series vector corresponding to each spatial grid point is used as an independent input and fed into the long short-term memory network unit with shared weights of the deep spatiotemporal feature extraction network. In the long short-term memory network unit, information is selectively updated, forgotten, and output through input gate, forget gate, and output gate mechanisms, learning the temporal evolution law of each spatial feature point; The updated feature vectors of each point output by the Long Short-Term Memory network unit are rearranged according to the original spatial structure to form a feature map containing temporal dynamic information.

9. A numerical forecasting method for ocean wind energy combining artificial intelligence according to claim 8, characterized in that, The abstract feature vector output by the deep spatiotemporal feature extraction network is mapped back to the gradient or divergence field of physical variables such as wind speed, air pressure, and temperature through a differentiable physical computation graph. Specifically, this includes: Construct a differentiable mapping network whose input is an abstract feature vector and whose output is a physical variable field corresponding to the physical meaning of standardized historical and current meteorological and hydrological fields; By utilizing the last linear transformation layer of the differentiable mapping network, high-dimensional abstract features are directly mapped into a two-dimensional spatial field of the horizontal component of wind speed, air pressure, and temperature. For the horizontal component field of wind speed obtained by mapping, the horizontal divergence field and vertical vorticity field are calculated on the preset spatial grid using a differentiable central difference operator. For the mapped pressure field, calculate its gradient field in the horizontal and vertical directions; For the mapped temperature field, calculate its horizontal gradient field. The calculated divergence field, vortex field, and gradient field together constitute the physical variable gradient or divergence field used for physical constraint calculation.

10. A marine wind energy numerical forecasting device incorporating artificial intelligence, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the numerical forecasting method for marine wind energy combined with artificial intelligence as described in any one of claims 1 to 9.

Images