Super-set deterministic weather forecasting method and device based on machine learning

Abstract

The invention discloses a super-set deterministic weather forecast method and device based on machine learning, and the method comprises the steps: obtaining multi-source meteorological data of a target region, carrying out the meshing of the multi-source meteorological data, carrying out the historical static feature analysis and dynamic feature analysis of the meshing features, obtaining the feature weight of each mode, and carrying out the recognition of the multi-source meteorological data. The method comprises the following steps: constructing a prediction sub-model according to an extreme event, obtaining enhanced numerical prediction data, obtaining posterior probability distribution of grid points through a conditional generative adversarial network and a Bayesian neural network based on the numerical prediction data, a gridding feature and a feature weight, taking a maximum probability value as a deterministic weather forecast, and calculating a confidence interval. And obtaining a joint probability product including wind speed and rainfall joint distribution and the characteristic contribution degree. According to the method, numerical forecasting set products of different mode centers are utilized to fuse probability forecasting information, deterministic weather forecasting is obtained, smoothing of extreme events is reduced, deterministic maximum value output is provided, and meanwhile good interpretability is achieved.

Description

Technical Field

[0001] This invention relates to the field of weather forecasting, and more particularly to a super-ensemble deterministic weather forecasting method and apparatus based on machine learning. Background Technology

[0002] Weather forecasting, as a key technology supporting modern social production, life, and disaster prevention and mitigation, has always relied on the advancement of numerical weather prediction. Numerical weather prediction generates deterministic forecasts by solving a set of partial differential equations of atmospheric dynamics, giving rise to the technique of quantifying forecast uncertainty ensemble forecasting. By generating multiple initial perturbation scenarios, it characterizes possible future climate states. Currently, in operational applications, major global meteorological centers provide a large number of numerical forecast products every day, including deterministic forecasts and global / regional ensemble systems. These multi-source model products constitute the so-called superensemble, which aims to reduce forecast errors by integrating the outputs of multiple centers.

[0003] However, multi-model ensembles suffer from systematic biases due to initial condition errors, differences in physical parameterization schemes, and limitations of the dynamic framework. Weather forecasting systems that perform well in mid-to-high latitudes tend to underestimate the intensity of tropical cyclones, while those that capture monsoon precipitation well show significant errors in the complex terrain of plateau regions. Furthermore, model performance dynamically changes with weather patterns, seasons, and geographical regions, making single ensemble weights unsuitable for all scenarios. While the Bayesian model averaging (BMA) method can allocate weights through historical error statistics, its fixed training period makes it difficult to capture model skill fluctuations during sudden weather events. Existing ensemble methods often employ weighted sets. While averaging or fixed-weight supersets can improve stability, they can lead to oversmoothing, weakening extreme precipitation signals and causing weak precipitation to spread to irrelevant areas. As a result, objective fusion methods such as grid fusion and subjective forecaster corrections have become the mainstream in operations. Meteorological stations generate deterministic products by manually assigning model weights, but the weights do not consider spatial and weather-related dependencies. Forecaster experience corrections can partially improve model bias, but the improvement rate is generally less than 5%, and it is highly dependent on individual ability and lacks stability. The public and industry users prefer to accept deterministic forecasts, but traditional methods cannot simultaneously provide deterministic output and uncertainty quantification.

[0004] In recent years, machine learning technology has provided a new path to solve these problems. Machine learning can learn the pattern of model error from massive historical data and dynamically adjust the ensemble strategy. Diffusion generation technology can generate 15-day global probability forecasts in a short time, especially in extreme weather and tropical cyclone forecasting. Kalman filter ensemble method improves the skill of near-surface element forecasts by 30% compared with the best single model and outperforms traditional super ensembles by 20%. The forecast error of typhoon track is lower than that of mainstream models. Machine learning can predict squall line patterns in advance through AI, which helps to issue warnings for severe convection. Although machine learning has made breakthroughs, it has not fully integrated multi-source observations such as satellites and radars with multi-model ensemble products, and cannot generate deterministic forecasts containing probabilistic information. Moreover, the output of AI meteorological big models is too smooth, which limits the ability to forecast extreme events. Therefore, how to promote the leap of weather forecasting towards intelligence and accuracy has become an urgent problem to be solved. Summary of the Invention

[0005] The purpose of this invention is to provide a method and apparatus for super-ensemble deterministic weather forecasting based on machine learning.

[0006] To achieve the above objectives, the present invention is implemented according to the following technical solution: The first aspect of this invention provides a super-ensemble deterministic weather forecasting method based on machine learning, comprising: Acquire multi-source meteorological data for the target area and preprocess the multi-source meteorological data, which includes meteorological observation data, digital elevation information, and numerical forecast ensemble products from different model centers; Principal component dimensionality reduction is performed on the high-order physical features of the multi-source meteorological data, and the dimensionality-reduced features are gridded according to the digital elevation information to obtain gridded features. Based on geographical region and seasonal information, historical static feature analysis is performed on the gridded features, and dynamic feature analysis is performed on the observation bias and weather flow patterns of the gridded features to obtain the feature weights of each model. Extreme events are classified and identified based on the meteorological observation data. Based on the atmospheric physical constraints and gridded characteristics of the extreme events, a prediction sub-model is constructed through a physical information neural network to obtain enhanced numerical prediction data. Based on the numerical prediction data, gridded features and feature weights, spatiotemporal consistency training is performed through a conditional generative adversarial network, and the posterior probability distribution of grid points is obtained through a Bayesian neural network according to the training results. The maximum probability value of the posterior probability distribution is used as the deterministic weather forecast, and the confidence interval is calculated to obtain a joint probability product including the joint distribution of wind speed and precipitation. The feature contribution is used as an additional product.

[0007] Furthermore, the method for obtaining the multi-source meteorological data includes: Numerical forecast ensemble products, meteorological observation data, and digital elevation information from different model centers within a preset time period are collected in the target area. Spatial resolution alignment of the numerical forecast ensemble products and meteorological observation data is performed using bilinear interpolation, and time series alignment is performed using linear interpolation to obtain multi-source meteorological data. The numerical forecast ensemble products include the temperature forecast field, precipitation forecast field, wind field forecast field, and ensemble dispersion product of the ensemble members. The meteorological observation data includes radar reflectivity, radar radial velocity, ground observation data, ERA5 reanalysis data, and meteorological AI large model prediction data. Based on the multi-source meteorological data, the standard deviation and variance of the corresponding data for the past 30 days are calculated according to the data type. Adaptive Z-score standardization is performed on each type of data according to the standard deviation and variance to obtain normalized multi-source meteorological data. Missing data in the multi-source meteorological data are imputed using a spatiotemporal graph convolutional network.

[0008] Furthermore, the method for obtaining the meshed features includes: Higher-order physical features are derived from the element field of multi-source meteorological data. Based on the higher-order physical features, dimensionality reduction is performed using principal component analysis algorithm to compress the dimension of the higher-order physical features to 30% of the original dimension, thus obtaining dimensionality-reduced features. The dimensionality-reduced features retain more than or equal to 90% of the feature information. The higher-order physical features include humidity field, atmospheric divergence field, convective effective potential energy, temperature advection field, vertical velocity field, and water vapor flux divergence field. The element field is the continuous distribution data of a single basic meteorological element in the spatial and temporal dimensions. Based on the DEM data of the digital elevation information, a unified spatial forecast grid is divided. The terrain elevation values ​​of the grid points are encoded into feature vectors through vector embedding based on the dimensionality reduction feature dimension. The dimensionality reduction feature and the terrain elevation feature are integrated into the corresponding grid points to obtain gridded features.

[0009] Furthermore, the method for obtaining the feature weights includes: Geographic regions of the target area are obtained based on digital elevation information, and the corresponding seasons of the geographic regions are obtained based on meteorological observation data. The gridded features are classified and statistically analyzed according to geographic regions and seasons to obtain historical static features. A static feature training dataset is generated based on the historical static features. The static feature training dataset is trained using a LightGBM model with a preset hybrid model to obtain the initial basic weights of each model. The historical static features include historical forecast error statistics, geographic regions, and seasonal types of different model centers. The real-time observation deviation between the gridded features and meteorological observation data is calculated. Based on the wind field structure of the atmospheric system and the weather flow pattern to which the pressure marker belongs in the current gridded features, the weather flow patterns are classified and statistically analyzed to obtain dynamic features. Based on the weather flow patterns, extreme meteorological events are identified through a pre-trained ResNet-50 classification network to obtain extreme events. The weather flow patterns include cyclones, anticyclones, and fronts. The Transformer model based on the preset hybrid model encodes dynamic features to obtain dynamic feature vectors. Based on the dynamic feature vectors, the feature correlation of the corresponding dynamic features of each mode set member is extracted through a self-attention mechanism to obtain the preliminary weight adjustment ratio of each mode. Based on the mode error ratio, the preliminary weight adjustment ratio of the grid area marked by extreme events is increased. Based on the preliminary weight adjustment ratio, the basic weights are corrected to obtain feature weights.

[0010] Furthermore, the method for obtaining the extreme event includes: Based on historical extreme weather events, the 95th percentile of the corresponding feature data is used as the extreme feature threshold. Extreme events are then identified using a discriminant formula based on this threshold. The discriminant formula is as follows: ; ; in For a two-dimensional Laplace operator, For spacetime points The atmospheric pressure field value at that location. The spatial grid coordinates of the target prediction area, This represents the air pressure field values ​​at the grid points adjacent to the right of the spatiotemporal point. This represents the air pressure field values ​​at the grid points adjacent to the left of the spatiotemporal point. This represents the air pressure field values ​​of the grid points adjacent to the upper side of the spatiotemporal point. This represents the air pressure field values ​​at the grid points adjacent to the lower side of the spatiotemporal point. For grid resolution, As an identifier for extreme events at a point in time and space, This is an indicator function; it takes the value 1 if the condition within the parentheses is met, and 0 if the condition is not met. The probability of extreme weather events. For the convective effective potential energy at a point in time and space, These are the temperature field values ​​at a point in time and space. This is the threshold for the probability of extreme events. The extreme pressure characteristic threshold is obtained based on the extreme characteristic threshold. The threshold for extreme convection characteristics. This represents the extreme temperature characteristic threshold. For OR operation.

[0011] Furthermore, the method for obtaining the numerical prediction data includes: The atmospheric motion partial differential equations are discretized based on grid regions and time periods identified by dynamic features and extreme events to obtain atmospheric physical constraints. These constraints are then embedded into the loss function of a physical information neural network, and a prediction sub-model for extreme regions is constructed. The expression for the physical constraint loss term of the loss function is as follows: ; ; in For physical constraints, This is the set of grid regions corresponding to extreme events. For the number of grid points, For masking functions in extreme regions, For spacetime points The wind speed field vector at that location. Let be the gradient operator, and be the set of spatial partial derivatives. For dot product operation, For spacetime points The atmospheric pressure field values, Atmospheric viscosity coefficient, Spacetime Point Coriolis force term, It is the L2 norm. For each model ensemble, forecasts are made at spatiotemporal points. The forecast standard deviation at that location The threshold for predicting uncertainty is obtained based on statistics of historical extreme events. The manifold feature vector of the current grid. This represents the characteristic vector of weather flow patterns under normal climate conditions in this region. Euclidean distance. The adjustment parameter for flow pattern similarity is obtained based on the degree of flow pattern anomaly. Based on the extreme event identifier, the extreme prediction range is identified through a masking function, and enhanced numerical prediction data is output through the corresponding prediction sub-model based on the extreme prediction range.

[0012] Furthermore, the method for obtaining the posterior probability distribution includes: The enhanced numerical prediction data is aligned according to the gridded feature dimension, and the multi-mode weighted output result is calculated through the feature weight. The structured preliminary spatiotemporal meteorological data is obtained through the diffusion model based on the multi-mode weighted output result. The structured input dataset is generated based on the preliminary spatiotemporal meteorological data, and the historical meteorological observation data of the same period is used as the real distribution reference sample. The structured input dataset is divided by stratified sampling to obtain the training set and the validation set. Spatiotemporal consistency training is performed using a conditional generative adversarial network based on the training set and reference samples. The feature weights corresponding to the gridded features are used as conditional constraint vectors. The generator of the conditional generative adversarial network is input with random noise using the conditional constraint vectors and the generator output is optimized using mean square error based on the validation set to obtain a spatiotemporally continuous fusion feature map. The fusion feature map includes the preliminary distribution parameters of each grid point. The reference sample and the fused feature map are used as the samples to be discriminated. The conditional constraint vector and the samples to be discriminated are input into the discriminator. The samples are discriminated based on the physical constraints of the distribution difference of adjacent grid points and gridding features. The discriminator performs adversarial gradient iteration on the samples until the spatiotemporal consistency error between the fused feature map and the reference sample does not decrease for five consecutive iterations, then the training ends. The fused feature map, gridded features, and feature weights that have been trained are used as input features. The numerical prediction error distribution under historical extreme and non-extreme scenarios is used as the initial prior. Based on the input features and prior distribution, the posterior distribution of network parameters is trained through a Bayesian neural network. The continuous probability ranking score and mean absolute error are used as the loss function of the Bayesian neural network until the deviation between the output result and the numerical prediction data is minimized, thus obtaining the posterior probability distribution of grid points. The posterior probability distribution includes the mean and variance of the distribution.

[0013] Furthermore, the method for obtaining the joint probability product includes: Based on the grid points, the mode of the posterior probability distribution is used as the maximum probability value to obtain a deterministic weather forecast result. The 90% confidence interval is calculated for the grid points according to the quantiles of the posterior probability distribution. The 5% and 95% quantiles of the posterior probability distribution are used as the upper and lower bounds of the confidence interval. The confidence interval of the deterministic weather forecast result is used as additional forecast information. The joint probability of different gridded feature elements is constructed based on the posterior probability distribution. The corresponding posterior probability distribution of the gridded feature elements is used as the marginal distribution through the Copula function. The joint probability density function of each feature element is fitted based on the marginal distribution. The joint probability density function includes the joint probability density function of wind speed and precipitation. A joint distribution of wind speed and precipitation is generated based on the joint probability density function of wind speed and precipitation. A joint probability heat map of wind speed and precipitation is generated based on the joint distribution of wind speed and precipitation. Key feature elements are obtained from the energy industry in the target area. The probability of key threshold combination is output based on the joint distribution of key feature elements. Based on the aforementioned feature weights and gridded features, the contribution of each feature to the deterministic weather forecast result, the joint distribution of wind speed and precipitation, and the probability of key threshold combinations is calculated using a preset SHAP analysis tool. A feature contribution visualization result is generated based on the contribution, and the feature contribution is used as an additional product. A joint probability product is obtained based on the deterministic weather forecast result, forecast additional information, and additional product.

[0014] A second aspect of the present invention provides a machine learning-based super-ensemble deterministic weather forecasting device, comprising: Data acquisition module: used to acquire multi-source meteorological data of the target area and preprocess the multi-source meteorological data, which includes meteorological observation data, digital elevation information and numerical forecast ensemble products from different model centers; Gridded feature extraction module: used to perform principal component dimensionality reduction on the high-order physical features of the multi-source meteorological data, and to gridded the dimensionality-reduced features according to the digital elevation information to obtain gridded features; Feature weight extraction module: used to perform historical static feature analysis on the gridded features based on geographical region and seasonal information, and to perform dynamic feature analysis on the observation bias and weather flow pattern of the gridded features to obtain the feature weights of each model; Extreme event prediction module: used to classify and identify extreme events based on the meteorological observation data, and to construct a prediction sub-model through a physical information neural network based on the atmospheric physical constraints and gridded characteristics of extreme events to obtain enhanced numerical prediction data; Super ensemble forecast data fusion module: used to perform spatiotemporal consistency training through a conditional generative adversarial network based on the numerical prediction data, gridded features and feature weights, and to obtain the posterior probability distribution of grid points through a Bayesian neural network based on the training results; Deterministic forecast product generation module: used to take the maximum probability value of the posterior probability distribution as a deterministic weather forecast, calculate the confidence interval, obtain a joint probability product including the joint distribution of wind speed and precipitation, and use the feature contribution as an additional product.

[0015] Compared with the prior art, the embodiments of the present invention have at least the following advantages or beneficial effects: This invention integrates numerical forecast ensemble products from different model centers, meteorological observation data, and digital elevation information. Principal component analysis is used to extract high-order physical features, which are then gridded. Static feature analysis of meteorological characteristics is performed based on geographical regions and seasons, while dynamic feature analysis is conducted on observation biases and weather patterns. The feature weights of each model, which change with weather patterns, are obtained, allowing for dynamic adjustment of these weights according to real-time scenarios, thus improving the accuracy of the integrated forecast. For extreme events such as typhoons and severe convection, classification and atmospheric physical constraints are combined with a physical information neural network to construct a predictive sub-model, enhancing the intensity of extreme weather signals and mitigating the problem of over-smoothing in ensemble forecasts. A conditional generative adversarial network is used to train the gridded features for spatiotemporal consistency, ensuring the structure of the element field. A Bayesian neural network is then used to obtain the posterior probability distribution, resulting in a deterministic forecast. Combined with confidence intervals, a joint wind speed and precipitation distribution product is output, along with additional information such as feature contribution. Based on a super-ensemble, this invention provides meteorological decision-making with multi-model dynamic integration, enhanced output for extreme events, and deterministic weather forecast products containing probabilistic information. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the steps of the machine learning-based superset deterministic weather forecasting method in this embodiment of the invention. Detailed Implementation

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

[0018] Reference Figure 1 As shown, this invention provides a super-ensemble deterministic weather forecasting method based on machine learning, including: Acquire multi-source meteorological data for the target area and preprocess the multi-source meteorological data, which includes meteorological observation data, digital elevation information, and numerical forecast ensemble products from different model centers; In the actual assessment, the forecast target area was from 00:00 on July 13th to 23:00 on July 15th, covering a geographical range of 113°E-116°E and 29°N-31°N, totaling 72 time intervals, including 30 days of historical forecast data. The radar reflectivity resolution of the meteorological observation data was 1km×1km, with an elevation angle of 0.5°, ranging from 10dBZ (clear sky) to 55dBZ (strong convective core), with one frame every 6 minutes, covering the evolution of strong convective cloud clusters, hourly precipitation, 2m air temperature, and wind speeds within the 10m / s range from 32 ground stations. The target area was assessed using SRTM. Digital elevation information (DEM) was obtained at a resolution of 30m × 30m. ERA5 reanalysis data was used as the background field for numerical weather prediction, with a resolution of 0.25° × 0.25°, air pressure of 990 hPa–1010 hPa, and relative humidity of 60%–90%. Numerical weather prediction ensemble products from three model centers were obtained from the target area. The first product had a resolution of 0.25° × 0.25°, with an average precipitation forecast error of 15% and a wind speed error of 12% over a fixed time period. The second product had a resolution of 0.5° × 0.5°, with an average precipitation forecast error of 20% and a wind speed error of 18%. The third product also had a resolution of 0.25° × 0.25°, with an average precipitation forecast error of 18% and a wind speed error of 15%. Multi-source meteorological data was obtained, and the spatial resolution of the multi-source meteorological data was aligned and unified to 0.25° × 0.25° to match the resolution of the first product and ERA5. A spatial matrix of 117 grid points was obtained through bilinear interpolation and linear interpolation. Each grid point corresponds to a unique spatiotemporal coordinate. Based on historical data from the past 30 days, adaptive Z-score standardization was used to obtain normalized multi-source meteorological data. Missing precipitation data at 10:00 and 11:00 on the 14th were interpolated using ST-GCN. The contemporaneous precipitation data of the four adjacent stations of the missing station, the topographic elevation of the grid, the precipitation data of the missing station in the two hours before (8:00 and 9:00) and the two hours after (12:00 and 13:00), and the contemporaneous ERA5 reanalysis precipitation data were used as the interpolation basis data. Based on the interpolation basis data, the ST-GCN output a complete value of 5.2mm, with a completion error MAE < 0.8mm. Based on the standardized data and physical common sense, the standardized wind speed value of 3.2 at a certain grid point at 15:00 on the 14th was judged as an anomaly. Since the precipitation was > 50mm / hour, the wind speed was > 20m / s and there was no strong convection signal, it was replaced with the median of 1.1 for the same period in the past 30 days. Spatiotemporally aligned, standardized, and missing multi-source meteorological data were obtained. Principal component dimensionality reduction is performed on the high-order physical features of the multi-source meteorological data, and the dimensionality-reduced features are gridded according to the digital elevation information to obtain gridded features. In the actual assessment, multi-source meteorological data was obtained. The data was in tensor format, corresponding to time number × 13 horizontal grids × 9 vertical grids × 18 basic features. Higher-order physical features were derived from the basic features, including convective effective potential energy, water vapor flux divergence, temperature advection, atmospheric divergence, vertical velocity, pressure gradient, and relative helicity. For a grid point, the near-surface air temperature is 35℃, the dew point temperature is 28℃, the 850hPa air temperature is 26℃, the dew point is 22℃, and the 500hPa air temperature is -5℃. Therefore, the convective effective potential energy ranges from 1200 J / kg (stable) to 2800 J / kg (extremely unstable), with an equilibrium result of 2500 J / kg. The water vapor flux divergence ranges from -0.05 × 10⁻⁶. -5 (Divergent) ~0.2×10 -5 (Strong convergence), temperature advection range is -3×10 -5 (Cold advection) ~8×10 -5 (Warm advection) A total of 8 high-order physical features were obtained. Standardization was performed feature-by-feature based on regional mean and standard deviation. The features were then selected based on a cumulative variance contribution rate ≥92%. The cumulative variance contribution rate of the first 6 principal components was calculated to be 93.5%, therefore, the first 6 principal components were retained. In the variance explanation ratio of the dimensionality-reduced features, the first principal component accounted for 49.2%, the second principal component 32.7%, the third principal component 10.1%, and the remaining components a cumulative 1.5%. The 30m DEM elevation data was vector-encoded, and fixed-dimensional embedding was used to map the elevation value of each grid to a 256-dimensional feature vector. The 6-dimensional reduced features were concatenated with the 256-dimensional terrain embedding vector according to dimension to obtain the grid feature vector for each grid point. This was then integrated according to time, horizontal grid, and vertical grid dimensions to obtain a gridded feature dataset. Based on geographical region and seasonal information, historical static feature analysis is performed on the gridded features, and dynamic feature analysis is performed on the observation bias and weather flow patterns of the gridded features to obtain the feature weights of each model. In the actual evaluation, historical static characteristics were obtained based on the historical forecast error statistics, geographical region and seasonal classification statistics of each model product obtained in step 1. The average historical forecast error rate was 15%, the average regional error was 12%, and the forecast bias and standard deviation for the three core elements—precipitation, wind speed, and temperature—were 0.18, 0.12, 0.08, and 0.13, respectively. The forecast timeliness (TS) was 0.21, and the extreme event hit rate was 0.85. The preset training parameters for the LightGBM hybrid model were: With a training rate of 0.05, tree depth of 6, 31 leaf nodes, and 100 iterations, the objective is to minimize the model-weighted prediction error. The label is the absolute error of the observed value minus the model prediction value. After training, the sum of the importance scores of the six static features is 1, with historical prediction error rate having an importance of 0.35, extreme event hit rate 0.25, geographic adaptability 0.20, and the remaining features totaling 0.20. Based on feature importance weighting and normalization, the basic weights of the three models are: Model 1 0.35, Model 2 0.32, and Model 3 0.3. 3. Combining the dynamic features of real-time weather such as flow patterns, physical quantity deviations, and extreme events, the Transformer encoding of a pre-defined hybrid model is used to encode feature correlations and correct the basic weights. A pre-trained ResNet-50 model is used to classify the current flow pattern, calculate the cosine similarity with the historical flow pattern, and calculate the mean Pearson correlation coefficient between the forecast value and grid features. The Transformer encoder has 3 layers, 8 attention heads, 256 hidden layer dimensions, and ReLU activation function. Four types of dynamic features × 3 patterns are input to form a 12-dimensional dynamic feature matrix. The encoding objective is to extract the correlation between patterns, features, and current weather. The output is the initial weight adjustment ratio for each pattern. A dynamic correction coefficient > 1 indicates that the pattern performance in the current scenario is better than the historical average, and the weight is increased; a dynamic correction coefficient < 1 indicates a decrease. The resulting dynamic correction coefficients for the three patterns are 1.12, 0.95, and 1.03, respectively. Based on the pattern error ratio, the initial weight adjustment ratio for the grid regions marked with extreme events is increased, resulting in feature weights of 0.378, 0.293, and 0.329. Extreme events are classified and identified based on the meteorological observation data. Based on the atmospheric physical constraints and gridded characteristics of the extreme events, a prediction sub-model is constructed through a physical information neural network to obtain enhanced numerical prediction data. In practical assessments, each grid point is determined to be an extreme event based on pressure field characteristics, grid features, and flow pattern probability. The 95th percentile of corresponding characteristic data from historical extreme weather events is used as the extreme characteristic threshold. Extreme events are then identified using a discriminant formula based on this threshold, where the extreme pressure characteristic threshold is -0.5 hPa / km². 2The extreme convection characteristic threshold is 2000 J / kg, and the extreme temperature characteristic threshold is 35℃, which are key conditions for strong convection in the warm sector. The ResNet extreme probability of the current grid point is obtained as 0.8. Based on the discriminant formula, the IsExtreme value of this grid point is determined to be 1, indicating a strong convection extreme region. The distribution of extreme regions in the target area is as follows: out of 117 grid points, 19 grid points meet all extreme thresholds, mainly concentrated in the southeast and northwest, corresponding to the core area of ​​strong convective cloud clusters. The remaining 98 grid points in the non-extreme region are mainly characterized by weak precipitation and clear skies. Physical constraints are constructed on the discrete atmospheric momentum conservation equation for the extreme regions, and the PINN loss function is embedded. The specific atmospheric momentum conservation equation is as follows: ; and For east-west and north-south wind speeds, For time, Atmospheric density is 1.2 kg / m³ 3 , This is a geostrophic parameter, representing the latitude. 10 -4 s -1 ,in The discretization formula for the directional momentum equation is: The forecast time step is as follows: For 1 hour, and For a distance of 27.8 km, atmospheric physical constraints are embedded into the loss function of a physical information neural network, and a prediction sub-model for extreme regions is constructed. A masking function is used to filter out grids in extreme regions with high uncertainty, and PINN enhancement is applied only to these grids. The prediction uncertainty threshold for the current region is 0.3. The prediction standard deviation of one grid (114.5°E, 30.25°N) is 0.816, therefore enhancement is needed. The PINN network structure of the prediction sub-model adapted for the current grid has a 265-dimensional input layer, a 4-layer fully connected hidden layer with 256 neurons per layer, a ReLU activation function, and a 3-layer output layer. The forecast elements are east-west wind speed, north-south wind speed, and precipitation. The training data for the prediction sub-model consists of historical observation data and current time-series model data from 19 grid points in the extreme region. The training parameters are a learning rate of 0.001, 500 iterations, and a batch size of 32, until the total loss does not decrease for 5 consecutive iterations. The output is an enhanced east-west wind speed of 12.0 m / s, which is 20% higher than the weighted value and matches the core wind speed of strong convection. The enhanced precipitation is 25.0 mm, which matches the precipitation intensity corresponding to radar reflectivity. The PINN enhanced value of the extreme region and the weighted value of the non-extreme region are stitched together by grid to form unified enhanced data for the entire region. Based on the numerical prediction data, gridded features and feature weights, spatiotemporal consistency training is performed through a conditional generative adversarial network, and the posterior probability distribution of grid points is obtained through a Bayesian neural network according to the training results. In the actual assessment, the enhanced numerical prediction data is aligned according to the gridded feature dimensions, and the multi-mode weighted output is calculated using the feature weights. Based on the multi-mode weighted output, a diffusion model is used to obtain structured preliminary spatiotemporal meteorological data. The observational data of severe convective weather in the target area in July is used as the real reference sample to extract the spatiotemporal sequences of precipitation and wind speed. Based on the preliminary spatiotemporal meteorological data, the training set (58 time periods) and the validation set (14 time periods) are split in an 8:2 ratio. Spatiotemporal stratified sampling is used to ensure that the extreme region proportions of the training and validation sets are consistent, both at 16.2%. Spatiotemporal consistency training is performed using CGAN based on the training set and the reference sample. The generator's input layer consists of a 265-dimensional conditional constraint vector and 100-dimensional random noise. Its four fully connected layers consist of 256→512→1024→2048 neurons, two convolutional layers with 3×3 kernels and a stride of 1, and an output layer with a 3-dimensional dimension. This generates spatiotemporally continuous precipitation and wind speed fusion feature maps. The discriminator uses a 3×3 kernel, three convolutional layers with a stride of 2, and two max pooling layers with a 2×2 stride. Finally, a sigmoid activation function outputs a probability value of 0 to 1. The spatiotemporal consistency error on the validation set does not decrease for five consecutive iterations and terminates after 520 iterations. The MAE of precipitation on the validation set is 1.2 mm, and the wind speed error is 0.3 m / s. Training is complete. The fused feature map, gridded features, and feature weights from the completed training are used as input features. These are then used to fit the posterior distribution probability to the BNN, where the hidden layers are three fully connected Bayesian layers: 512→256→128 neurons. The output layer outputs the mean and variance. The forecast error distribution for strong and weak convection scenarios in this region is set as an extreme region: a prior precipitation error of 2.5 mm, 1.2 mm. 2 The prior wind speed error is 0.8 m / s, 0.3 m / s. 2 In non-extreme areas: the prior precipitation error is 1.2 mm, 0.5 mm. 2 The prior wind speed error is 0.3 m / s, 0.1 2 The BNN training parameters were set with a learning rate of 0.001, using the SGD optimizer, 800 iterations, and a batch size of 32. CRPS and MAE were used as constraint losses. Variational inference was used to approximate the posterior distribution, and the prior and posterior probability distributions were approximated by minimizing the KL divergence. The posterior distribution of precipitation at one grid point was obtained as 25.0 mm, 3.0 mm. 2 The posterior distribution of wind speed is 12.0 m / s, 1.5 2 Gaussian filtering is applied to the posterior mean and variance of adjacent grids, with the filtering standard being the standard deviation of one grid. The maximum probability value of the posterior probability distribution is used as the deterministic weather forecast, and the confidence interval is calculated to obtain a joint probability product including the joint distribution of wind speed and precipitation. The feature contribution is used as an additional product.

[0019] In practical assessment, deterministic values ​​are extracted from the posterior probability distribution of the target area, and a 90% confidence interval is calculated. The mode of the posterior probability distribution is taken as the deterministic forecast value, and the meteorological element type of the corresponding grid point is matched. The mode of the precipitation characteristic element is 25.0 mm, so the deterministic precipitation forecast is 25.0 mm, indicating strong convective core precipitation. The deterministic wind speed forecast is 12.0 m / s, indicating strong convective wind. Based on the quantiles of the posterior probability distribution, the 5th quantile is taken as the lower bound, and the 95th quantile is taken as the upper bound. The calculated 90% confidence interval for precipitation is [20.0 mm, ...]. [30.0mm], wind speed 90% confidence interval [9.5m / s, 14.5m / s], classified according to the width of the confidence interval: narrow: ≤5mm / 2m / s, low uncertainty; medium: 5~10mm / 2~4m / s, medium uncertainty; wide: >10mm / 4m / s, high uncertainty. The current core grid is marked as high uncertainty. The posterior probability distributions of precipitation and wind speed are used as marginal distributions. Based on the marginal distributions, the joint probability density of precipitation and wind speed is fitted using the Copula function. The Copula joint probability density formula is: ; in The cumulative probability of precipitation edge distribution. The cumulative probability of the edge distribution of wind speed. The shape parameter represents the dependency between the feature elements. The joint probability density of the core grid is calculated to be 0.12, with the unit being 1 / (mm·m / s), thus obtaining the joint distribution of wind speed and precipitation. Based on the joint distribution of wind speed and precipitation, wind speed (0~20m / s, 1m / s interval) is used as the horizontal axis, and hourly precipitation (0~50mm, 2mm interval) is used as the vertical axis. Color depth represents the joint probability density, resulting in a wind speed and precipitation joint probability heatmap. In the target region, the energy sector mainly consists of wind power and photovoltaics. The key features of interest are wind speed and daily precipitation. Optimal operating conditions are wind speed of 5~10m / s and precipitation of 0~5mm. Under these conditions, the probability of key threshold combinations in the core grid is 12%, with non-extreme areas accounting for ≥30%. A threshold probability table for the energy sector is obtained. Based on the feature weights and gridded features, the contribution of each feature to the deterministic weather forecast result is calculated using a preset SHAP analysis tool. The preset SHAP analysis tool uses the TreeExplainer interpreter, and the feature contribution is calculated based on the Shapley value. The flow available potential energy (CAPE) contributes 35% to deterministic precipitation, 28% to deterministic wind speed, and 32% to the joint probability density. This indicates that the higher the CAPE, the stronger the strong convective instability energy, the greater the precipitation and wind speed, and the higher the joint probability. The contributions of water vapor flux divergence are 25%, 15%, and 20%, respectively, indicating that the stronger the water vapor convergence, the greater the precipitation potential, and indirectly increasing the probability of strong winds. A feature contribution heatmap is generated based on the contribution, with the horizontal and vertical axes representing grid coordinates and the color indicating the contribution of a core feature. Joint probability products are obtained based on deterministic weather forecast results, forecast supplementary information, and supplementary products, including deterministic grid feature element forecasts, precipitation forecasts, deterministic wind speed forecasts and corresponding 90% confidence intervals, as well as a wind speed and precipitation joint probability heatmap, an energy industry threshold probability table, a feature contribution spatiotemporal heatmap, and a feature contribution ranking table.

[0020] In this embodiment, the method for obtaining the multi-source meteorological data includes: Numerical forecast ensemble products, meteorological observation data, and digital elevation information from different model centers within a preset time period are collected in the target area. Spatial resolution alignment of the numerical forecast ensemble products and meteorological observation data is performed using bilinear interpolation, and time series alignment is performed using linear interpolation to obtain multi-source meteorological data. The numerical forecast ensemble products include the temperature forecast field, precipitation forecast field, wind field forecast field, and ensemble dispersion product of the ensemble members. The meteorological observation data includes radar reflectivity, radar radial velocity, ground observation data, ERA5 reanalysis data, and meteorological AI large model prediction data. Based on the multi-source meteorological data, the standard deviation and variance of the corresponding data for the past 30 days are calculated according to the data type. Adaptive Z-score standardization is performed on each type of data according to the standard deviation and variance to obtain normalized multi-source meteorological data. Missing data in the multi-source meteorological data are imputed using a spatiotemporal graph convolutional network.

[0021] In this embodiment, the method for obtaining the meshed features includes: Higher-order physical features are derived from the element field of multi-source meteorological data. Based on the higher-order physical features, dimensionality reduction is performed using principal component analysis algorithm to compress the dimension of the higher-order physical features to 30% of the original dimension, thus obtaining dimensionality-reduced features. The dimensionality-reduced features retain more than or equal to 90% of the feature information. The higher-order physical features include humidity field, atmospheric divergence field, convective effective potential energy, temperature advection field, vertical velocity field, and water vapor flux divergence field. The element field is the continuous distribution data of a single basic meteorological element in the spatial and temporal dimensions. Based on the DEM data of the digital elevation information, a unified spatial forecast grid is divided. The terrain elevation values ​​of the grid points are encoded into feature vectors through vector embedding based on the dimensionality reduction feature dimension. The dimensionality reduction feature and the terrain elevation feature are integrated into the corresponding grid points to obtain gridded features.

[0022] In this embodiment, the method for obtaining the feature weights includes: Geographic regions of the target area are obtained based on digital elevation information, and the corresponding seasons of the geographic regions are obtained based on meteorological observation data. The gridded features are classified and statistically analyzed according to geographic regions and seasons to obtain historical static features. A static feature training dataset is generated based on the historical static features. The static feature training dataset is trained using a LightGBM model with a preset hybrid model to obtain the initial basic weights of each model. The historical static features include historical forecast error statistics, geographic regions, and seasonal types of different model centers. The real-time observation deviation between the gridded features and meteorological observation data is calculated. Based on the wind field structure of the atmospheric system and the weather flow pattern to which the pressure marker belongs in the current gridded features, the weather flow patterns are classified and statistically analyzed to obtain dynamic features. Based on the weather flow patterns, extreme meteorological events are identified through a pre-trained ResNet-50 classification network to obtain extreme events. The weather flow patterns include cyclones, anticyclones, and fronts. The Transformer model based on the preset hybrid model encodes dynamic features to obtain dynamic feature vectors. Based on the dynamic feature vectors, the feature correlation of the corresponding dynamic features of each mode set member is extracted through a self-attention mechanism to obtain the preliminary weight adjustment ratio of each mode. Based on the mode error ratio, the preliminary weight adjustment ratio of the grid area marked by extreme events is increased. Based on the preliminary weight adjustment ratio, the basic weights are corrected to obtain feature weights.

[0023] In this embodiment, the method for obtaining the extreme event includes: Based on historical extreme weather events, the 95th percentile of the corresponding feature data is used as the extreme feature threshold. Extreme events are then identified using a discriminant formula based on this threshold. The discriminant formula is as follows: ; ; in For a two-dimensional Laplace operator, For spacetime points The atmospheric pressure field value at that location. The spatial grid coordinates of the target prediction area, This represents the air pressure field values ​​at the grid points adjacent to the right of the spatiotemporal point. This represents the air pressure field values ​​at the grid points adjacent to the left of the spatiotemporal point. This represents the air pressure field values ​​of the grid points adjacent to the upper side of the spatiotemporal point. This represents the air pressure field values ​​at the grid points adjacent to the lower side of the spatiotemporal point. For grid resolution, As an identifier for extreme events at a point in time and space, This is an indicator function; it takes the value 1 if the condition within the parentheses is met, and 0 if the condition is not met. The probability of extreme weather events. For the convective effective potential energy at a point in time and space, These are the temperature field values ​​at a point in time and space. This is the threshold for the probability of extreme events. The extreme pressure characteristic threshold is obtained based on the extreme characteristic threshold. The threshold for extreme convection characteristics. This represents the extreme temperature characteristic threshold. For OR operation.

[0024] In this embodiment, the method for obtaining the numerical prediction data includes: The atmospheric motion partial differential equations are discretized based on grid regions and time periods identified by dynamic features and extreme events to obtain atmospheric physical constraints. These constraints are then embedded into the loss function of a physical information neural network, and a prediction sub-model for extreme regions is constructed. The expression for the physical constraint loss term of the loss function is as follows: ; ; in For physical constraints, This is the set of grid regions corresponding to extreme events. For the number of grid points, For masking functions in extreme regions, For spacetime points The wind speed field vector at that location. Let be the gradient operator, and be the set of spatial partial derivatives. For dot product operation, For spacetime points The atmospheric pressure field values, Atmospheric viscosity coefficient, Spacetime Point Coriolis force term, It is the L2 norm. For each model ensemble, forecasts are made at spatiotemporal points. The forecast standard deviation at that location The threshold for predicting uncertainty is obtained based on statistics of historical extreme events. The manifold feature vector of the current grid. This represents the characteristic vector of weather flow patterns under normal climate conditions in this region. Euclidean distance. The adjustment parameter for flow pattern similarity is obtained based on the degree of flow pattern anomaly. Based on the extreme event identifier, the extreme prediction range is identified through a masking function, and enhanced numerical prediction data is output through the corresponding prediction sub-model based on the extreme prediction range.

[0025] In this embodiment, the method for obtaining the posterior probability distribution includes: The enhanced numerical prediction data is aligned according to the gridded feature dimension, and the multi-mode weighted output result is calculated through the feature weight. The structured preliminary spatiotemporal meteorological data is obtained through the diffusion model based on the multi-mode weighted output result. The structured input dataset is generated based on the preliminary spatiotemporal meteorological data, and the historical meteorological observation data of the same period is used as the real distribution reference sample. The structured input dataset is divided by stratified sampling to obtain the training set and the validation set. Spatiotemporal consistency training is performed using a conditional generative adversarial network based on the training set and reference samples. The feature weights corresponding to the gridded features are used as conditional constraint vectors. The generator of the conditional generative adversarial network is input with random noise using the conditional constraint vectors and the generator output is optimized using mean square error based on the validation set to obtain a spatiotemporally continuous fusion feature map. The fusion feature map includes the preliminary distribution parameters of each grid point. The reference sample and the fused feature map are used as the samples to be discriminated. The conditional constraint vector and the samples to be discriminated are input into the discriminator. The samples are discriminated based on the physical constraints of the distribution difference of adjacent grid points and gridding features. The discriminator performs adversarial gradient iteration on the samples until the spatiotemporal consistency error between the fused feature map and the reference sample does not decrease for five consecutive iterations, then the training ends. The fused feature map, gridded features, and feature weights that have been trained are used as input features. The numerical prediction error distribution under historical extreme and non-extreme scenarios is used as the initial prior. Based on the input features and prior distribution, the posterior distribution of network parameters is trained through a Bayesian neural network. The continuous probability ranking score and mean absolute error are used as the loss function of the Bayesian neural network until the deviation between the output result and the numerical prediction data is minimized, thus obtaining the posterior probability distribution of grid points. The posterior probability distribution includes the mean and variance of the distribution.

[0026] In this embodiment, the method for obtaining the joint probability product includes: Based on the grid points, the mode of the posterior probability distribution is used as the maximum probability value to obtain a deterministic weather forecast result. The 90% confidence interval is calculated for the grid points according to the quantiles of the posterior probability distribution. The 5% and 95% quantiles of the posterior probability distribution are used as the upper and lower bounds of the confidence interval. The confidence interval of the deterministic weather forecast result is used as additional forecast information. The joint probability of different gridded feature elements is constructed based on the posterior probability distribution. The corresponding posterior probability distribution of the gridded feature elements is used as the marginal distribution through the Copula function. The joint probability density function of each feature element is fitted based on the marginal distribution. The joint probability density function includes the joint probability density function of wind speed and precipitation. A joint distribution of wind speed and precipitation is generated based on the joint probability density function of wind speed and precipitation. A joint probability heat map of wind speed and precipitation is generated based on the joint distribution of wind speed and precipitation. Key feature elements are obtained from the energy industry in the target area. The probability of key threshold combination is output based on the joint distribution of key feature elements. Based on the aforementioned feature weights and gridded features, the contribution of each feature to the deterministic weather forecast result, the joint distribution of wind speed and precipitation, and the probability of key threshold combinations is calculated using a preset SHAP analysis tool. A feature contribution visualization result is generated based on the contribution, and the feature contribution is used as an additional product. A joint probability product is obtained based on the deterministic weather forecast result, forecast additional information, and additional product.

[0027] A second aspect of the present invention also provides a machine learning-based super-ensemble deterministic weather forecasting device, comprising: Data acquisition module: used to acquire multi-source meteorological data of the target area and preprocess the multi-source meteorological data, which includes meteorological observation data, digital elevation information and numerical forecast ensemble products from different model centers; Gridded feature extraction module: used to perform principal component dimensionality reduction on the high-order physical features of the multi-source meteorological data, and to gridded the dimensionality-reduced features according to the digital elevation information to obtain gridded features; Feature weight extraction module: used to perform historical static feature analysis on the gridded features based on geographical region and seasonal information, and to perform dynamic feature analysis on the observation bias and weather flow pattern of the gridded features to obtain the feature weights of each model; Extreme event prediction module: used to classify and identify extreme events based on the meteorological observation data, and to construct a prediction sub-model through a physical information neural network based on the atmospheric physical constraints and gridded characteristics of extreme events to obtain enhanced numerical prediction data; Super ensemble forecast data fusion module: used to perform spatiotemporal consistency training through a conditional generative adversarial network based on the numerical prediction data, gridded features and feature weights, and to obtain the posterior probability distribution of grid points through a Bayesian neural network based on the training results; Deterministic forecast product generation module: used to take the maximum probability value of the posterior probability distribution as a deterministic weather forecast, calculate the confidence interval, obtain a joint probability product including the joint distribution of wind speed and precipitation, and use the feature contribution as an additional product.

[0028] The above description is merely an example and illustration of the structure of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the structure of the invention or exceed the scope defined in the claims, all of which should fall within the protection scope of the present invention.

Claims

1. A machine learning-based superset deterministic weather forecasting method, characterized in that, Includes the following steps: Acquire multi-source meteorological data for the target area and preprocess the multi-source meteorological data, which includes meteorological observation data, digital elevation information, and numerical forecast ensemble products from different model centers; Principal component dimensionality reduction is performed on the high-order physical features of the multi-source meteorological data, and the dimensionality-reduced features are gridded according to the digital elevation information to obtain gridded features. Based on geographical region and seasonal information, historical static feature analysis is performed on the gridded features, and dynamic feature analysis is performed on the observation bias and weather flow patterns of the gridded features to obtain the feature weights of each model. Extreme events are classified and identified based on the meteorological observation data. Based on the atmospheric physical constraints and gridded characteristics of the extreme events, a prediction sub-model is constructed through a physical information neural network to obtain enhanced numerical prediction data. Based on the numerical prediction data, gridded features and feature weights, spatiotemporal consistency training is performed through a conditional generative adversarial network, and the posterior probability distribution of grid points is obtained through a Bayesian neural network according to the training results. The maximum probability value of the posterior probability distribution is used as the deterministic weather forecast, and the confidence interval is calculated to obtain a joint probability product including the joint distribution of wind speed and precipitation. The feature contribution is used as an additional product.

2. The machine learning-based superset deterministic weather forecasting method according to claim 1, characterized in that, The method for obtaining the multi-source meteorological data includes: Numerical forecast ensemble products, meteorological observation data, and digital elevation information from different model centers within a preset time period are collected in the target area. Spatial resolution alignment of the numerical forecast ensemble products and meteorological observation data is performed using bilinear interpolation, and time series alignment is performed using linear interpolation to obtain multi-source meteorological data. The numerical forecast ensemble products include the temperature forecast field, precipitation forecast field, wind field forecast field, and ensemble dispersion product of the ensemble members. The meteorological observation data includes radar reflectivity, radar radial velocity, ground observation data, ERA5 reanalysis data, and meteorological AI large model prediction data. Based on the multi-source meteorological data, the standard deviation and variance of the corresponding data for the past 30 days are calculated according to the data type. Adaptive Z-score standardization is performed on each type of data according to the standard deviation and variance to obtain normalized multi-source meteorological data. Missing data in the multi-source meteorological data are imputed using a spatiotemporal graph convolutional network.

3. The machine learning-based superset deterministic weather forecasting method according to claim 1, characterized in that, The method for obtaining the meshed features includes: Higher-order physical features are derived from the element field of multi-source meteorological data. Based on the higher-order physical features, dimensionality reduction is performed using principal component analysis algorithm to compress the dimension of the higher-order physical features to 30% of the original dimension, thus obtaining dimensionality-reduced features. The dimensionality-reduced features retain more than or equal to 90% of the feature information. The higher-order physical features include humidity field, atmospheric divergence field, convective effective potential energy, temperature advection field, vertical velocity field, and water vapor flux divergence field. The element field is the continuous distribution data of a single basic meteorological element in the spatial and temporal dimensions. Based on the DEM data of the digital elevation information, a unified spatial forecast grid is divided. The terrain elevation values ​​of the grid points are encoded into feature vectors through vector embedding based on the dimensionality reduction feature dimension. The dimensionality reduction feature and the terrain elevation feature are integrated into the corresponding grid points to obtain gridded features.

4. The machine learning-based superset deterministic weather forecasting method according to claim 1, characterized in that, The method for obtaining the feature weights includes: Geographic regions of the target area are obtained based on digital elevation information, and the corresponding seasons of the geographic regions are obtained based on meteorological observation data. The gridded features are classified and statistically analyzed according to geographic regions and seasons to obtain historical static features. A static feature training dataset is generated based on the historical static features. The static feature training dataset is trained using a LightGBM model with a preset hybrid model to obtain the initial basic weights of each model. The historical static features include historical forecast error statistics, geographic regions, and seasonal types of different model centers. The real-time observation deviation between the gridded features and meteorological observation data is calculated. Based on the wind field structure of the atmospheric system and the weather flow pattern to which the pressure marker belongs in the current gridded features, the weather flow patterns are classified and statistically analyzed to obtain dynamic features. Based on the weather flow patterns, extreme meteorological events are identified through a pre-trained ResNet-50 classification network to obtain extreme events. The weather flow patterns include cyclones, anticyclones, and fronts. The Transformer model based on the preset hybrid model encodes dynamic features to obtain dynamic feature vectors. Based on the dynamic feature vectors, the feature correlation of the corresponding dynamic features of each mode set member is extracted through a self-attention mechanism to obtain the preliminary weight adjustment ratio of each mode. Based on the mode error ratio, the preliminary weight adjustment ratio of the grid area marked by extreme events is increased. Based on the preliminary weight adjustment ratio, the basic weights are corrected to obtain feature weights.

5. The machine learning-based super-ensemble deterministic weather forecasting method according to claim 4, characterized in that, Methods for obtaining the extreme events include: Based on historical extreme weather events, the 95th percentile of the corresponding feature data is used as the extreme feature threshold. Extreme events are then identified using a discriminant formula based on this threshold. The discriminant formula is as follows: ； ； in For a two-dimensional Laplace operator, For spacetime points The atmospheric pressure field value at that location. The spatial grid coordinates of the target prediction area, This represents the air pressure field values ​​at the grid points adjacent to the right of the spatiotemporal point. This represents the air pressure field values ​​at the grid points adjacent to the left of the spatiotemporal point. This represents the air pressure field values ​​of the grid points adjacent to the upper side of the spatiotemporal point. This represents the air pressure field values ​​at the grid points adjacent to the lower side of the spatiotemporal point. For grid resolution, As an identifier for extreme events at a point in time and space, This is an indicator function; it takes the value 1 if the condition within the parentheses is met, and 0 if the condition is not met. The probability of extreme weather events. For the convective effective potential energy at a point in time and space, These are the temperature field values ​​at a point in time and space. This is the threshold for the probability of extreme events. The extreme pressure characteristic threshold is obtained based on the extreme characteristic threshold. The threshold for extreme convection characteristics. This represents the extreme temperature characteristic threshold. For OR operation.

6. The machine learning-based superset deterministic weather forecasting method according to claim 1, characterized in that, The method for obtaining the numerical prediction data includes: The atmospheric motion partial differential equations are discretized based on grid regions and time periods identified by dynamic features and extreme events to obtain atmospheric physical constraints. These constraints are then embedded into the loss function of a physical information neural network, and a prediction sub-model for extreme regions is constructed. The expression for the physical constraint loss term of the loss function is as follows: ； ； in For physical constraints, This is the set of grid regions corresponding to extreme events. For the number of grid points, For masking functions in extreme regions, For spacetime points The wind speed field vector at that location. Let be the gradient operator, and be the set of spatial partial derivatives. For dot product operation, For spacetime points The atmospheric pressure field values, Atmospheric viscosity coefficient, Spacetime Point Coriolis force term, It is the L2 norm. For each model ensemble, forecasts are made at spatiotemporal points. The forecast standard deviation at that location The threshold for predicting uncertainty is obtained based on statistics of historical extreme events. The manifold feature vector of the current grid. This represents the characteristic vector of weather flow patterns under normal climate conditions in this region. Euclidean distance. The adjustment parameter for flow pattern similarity is obtained based on the degree of flow pattern anomaly. Based on the extreme event identifier, the extreme prediction range is identified through a masking function, and enhanced numerical prediction data is output through the corresponding prediction sub-model based on the extreme prediction range.

7. The machine learning-based superset deterministic weather forecasting method according to claim 1, characterized in that, The method for obtaining the posterior probability distribution includes: The enhanced numerical prediction data is aligned according to the gridded feature dimension, and the multi-mode weighted output result is calculated through the feature weight. The structured preliminary spatiotemporal meteorological data is obtained through the diffusion model based on the multi-mode weighted output result. The structured input dataset is generated based on the preliminary spatiotemporal meteorological data, and the historical meteorological observation data of the same period is used as the real distribution reference sample. The structured input dataset is divided by stratified sampling to obtain the training set and the validation set. Spatiotemporal consistency training is performed using a conditional generative adversarial network based on the training set and reference samples. The feature weights corresponding to the gridded features are used as conditional constraint vectors. The generator of the conditional generative adversarial network is input with random noise using the conditional constraint vectors and the generator output is optimized using mean square error based on the validation set to obtain a spatiotemporally continuous fusion feature map. The fusion feature map includes the preliminary distribution parameters of each grid point. The reference sample and the fused feature map are used as the samples to be discriminated. The conditional constraint vector and the samples to be discriminated are input into the discriminator. The samples are discriminated based on the physical constraints of the distribution difference of adjacent grid points and gridding features. The discriminator performs adversarial gradient iteration on the samples until the spatiotemporal consistency error between the fused feature map and the reference sample does not decrease for five consecutive iterations, then the training ends. The fused feature map, gridded features, and feature weights that have been trained are used as input features. The numerical prediction error distribution under historical extreme and non-extreme scenarios is used as the initial prior. Based on the input features and prior distribution, the posterior distribution of network parameters is trained through a Bayesian neural network. The continuous probability ranking score and mean absolute error are used as the loss function of the Bayesian neural network until the deviation between the output result and the numerical prediction data is minimized, thus obtaining the posterior probability distribution of grid points. The posterior probability distribution includes the mean and variance of the distribution.

8. The machine learning-based superset deterministic weather forecasting method according to claim 1, characterized in that, A method for obtaining the joint probability product includes: Based on the grid points, the mode of the posterior probability distribution is used as the maximum probability value to obtain a deterministic weather forecast result. The 90% confidence interval is calculated for the grid points according to the quantiles of the posterior probability distribution. The 5% and 95% quantiles of the posterior probability distribution are used as the upper and lower bounds of the confidence interval. The confidence interval of the deterministic weather forecast result is used as additional forecast information. The joint probability of different gridded feature elements is constructed based on the posterior probability distribution. The corresponding posterior probability distribution of the gridded feature elements is used as the marginal distribution through the Copula function. The joint probability density function of each feature element is fitted based on the marginal distribution. The joint probability density function includes the joint probability density function of wind speed and precipitation. A joint distribution of wind speed and precipitation is generated based on the joint probability density function of wind speed and precipitation. A joint probability heat map of wind speed and precipitation is generated based on the joint distribution of wind speed and precipitation. Key feature elements are obtained from the energy industry in the target area. The probability of key threshold combination is output based on the joint distribution of key feature elements. Based on the aforementioned feature weights and gridded features, the contribution of each feature to the deterministic weather forecast result, the joint distribution of wind speed and precipitation, and the probability of key threshold combinations is calculated using a preset SHAP analysis tool. A feature contribution visualization result is generated based on the contribution, and the feature contribution is used as an additional product. A joint probability product is obtained based on the deterministic weather forecast result, forecast additional information, and additional product.

9. A machine learning-based superset deterministic weather forecasting device, used to execute the machine learning-based superset deterministic weather forecasting method according to any one of claims 1 to 8, characterized in that, The system includes: Data acquisition module: used to acquire multi-source meteorological data of the target area and preprocess the multi-source meteorological data, which includes meteorological observation data, digital elevation information and numerical forecast ensemble products from different model centers; Gridded feature extraction module: used to perform principal component dimensionality reduction on the high-order physical features of the multi-source meteorological data, and to gridded the dimensionality-reduced features according to the digital elevation information to obtain gridded features; Feature weight extraction module: used to perform historical static feature analysis on the gridded features based on geographical region and seasonal information, and to perform dynamic feature analysis on the observation bias and weather flow pattern of the gridded features to obtain the feature weights of each model; Extreme event prediction module: used to classify and identify extreme events based on the meteorological observation data, and to construct a prediction sub-model through a physical information neural network based on the atmospheric physical constraints and gridded characteristics of extreme events to obtain enhanced numerical prediction data; Super ensemble forecast data fusion module: used to perform spatiotemporal consistency training through a conditional generative adversarial network based on the numerical prediction data, gridded features and feature weights, and to obtain the posterior probability distribution of grid points through a Bayesian neural network based on the training results; Deterministic forecast product generation module: used to take the maximum probability value of the posterior probability distribution as a deterministic weather forecast, calculate the confidence interval, obtain a joint probability product including the joint distribution of wind speed and precipitation, and use the feature contribution as an additional product.

Images