A weather downscaling method based on spatio-temporal fusion and physical constraints

Abstract

The application provides a meteorological downscaling method based on space-time fusion and physical constraints, and belongs to the technical field of meteorological downscaling, and comprises the following steps: preprocessing multi-source meteorological related data and high-resolution meteorological true values, and constructing a training data set; an improved U-Net model is constructed, space-time features and multi-source auxiliary features are obtained through a multi-branch feature extraction unit, high-resolution information is recovered through fusion and decoding, and a key area is highlighted by embedding an attention reinforcement module; the model is trained through the training data set, a loss function combining terrain features and physical rules is used to optimize parameters, and differentiated constraints on prediction errors in complex regions and in violation of physical laws are realized; and target low-resolution data is input into the model after preprocessing, and high-resolution meteorological data and physical attribution results are output. The application solves the problems of insufficient multi-source meteorological data fusion, insufficient downscaling precision in complex terrain regions and lack of effective constraints on prediction errors in violation of physical laws in the prior art.

Description

Technical Field

[0001] This invention relates to the field of meteorological downscaling technology, and in particular to a meteorological downscaling method based on spatiotemporal fusion and physical constraints. Background Technology

[0002] In the field of meteorological downscaling, with the development of technologies such as satellite and radar monitoring and numerical weather prediction, it is now possible to acquire meteorological observation and forecast data with a certain spatiotemporal resolution. The processing capability of observational data has also been improved through the integrated application of quality control techniques using artificial intelligence and mathematical statistics. Regarding model building, various models for meteorological analysis already exist. For example, numerical weather prediction models, through the processing and assimilation of massive amounts of multi-source observational data, form high-precision gridded analysis fields, thereby predicting future weather evolution.

[0003] Meteorological research and applications are demanding increasingly higher precision and resolution from meteorological data. To meet the needs of fields such as urban disaster prevention and mitigation, precision meteorological services for agriculture, and watershed hydro-meteorological forecasting, downscaling techniques have become a research hotspot. Meanwhile, integrating multi-source data (such as satellite cloud images, radar echoes, topographic data, and conventional meteorological data) and combining physical mechanisms with artificial intelligence technologies to build more accurate meteorological models are important directions for future development. Similar to AI-driven meteorological applications, implementing large-scale AI weather forecasting model demonstration projects and developing short- and medium-term forecasting models that incorporate constraints from atmospheric physical mechanisms are all measures that align with this trend.

[0004] However, existing technologies still have shortcomings in meteorological downscaling. On the one hand, traditional downscaling methods struggle to fully utilize complementary information from multi-source data and cannot accurately characterize the impact of factors such as complex terrain on meteorological elements. For example, in areas with complex terrain, the lack of in-depth analysis of the relationship between terrain features and meteorological elements leads to significant discrepancies between the description of local meteorological conditions after the meteorological data resolution is improved and the actual situation. On the other hand, existing models lack effective constraint mechanisms when dealing with prediction errors that violate physical laws, resulting in reduced reliability of prediction results under certain extreme or special meteorological scenarios. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, the purpose of this invention is to provide a meteorological downscaling method based on spatiotemporal fusion and physical constraints. Through multi-branch feature extraction, spatiotemporal fusion, attention enhancement, and loss function optimization that integrates terrain and physical rules, it achieves accurate downscaling of low-resolution meteorological data, improves the prediction accuracy and physical rationality of high-resolution meteorological data in complex terrain areas and special meteorological scenarios, and can output physical attribution results to clarify influencing factors.

[0006] To achieve the above objectives, the present invention provides the following solution:

[0007] A meteorological downscaling method based on spatiotemporal fusion and physical constraints includes the following steps:

[0008] S1. Preprocess the multi-source meteorological data and high-resolution meteorological ground truth to construct a training dataset; the preprocessing includes standardization and feature extraction.

[0009] S2. Construct an improved U-Net model. The model obtains spatiotemporal features and multi-source auxiliary features through a multi-branch feature extraction unit, restores high-resolution information through fusion and decoding, and embeds an attention enhancement module to highlight key region features.

[0010] S3. Train the model using the training dataset, and optimize the parameters using a loss function that integrates terrain features and physical rules to achieve differentiated constraints on prediction errors for complex areas and areas that violate physical laws.

[0011] S4. After performing the preprocessing step S1 on the target low-resolution data, input it into the trained model and output high-resolution meteorological data and physical attribution results.

[0012] Preferably, in step S1, the multi-source meteorological data includes low-resolution meteorological data, topographic data, and multi-source observation data.

[0013] Preferably, the low-resolution meteorological data includes one or more of wind speed, temperature, and irradiance; the terrain data is digital elevation model data; and the multi-source observation data includes satellite cloud images and radar echoes.

[0014] Preferably, in step S1, the standardization is adaptive normalization, specifically:

[0015]

[0016] Where x′ represents the standardized data, x represents the original data, and μ(x) represents the normalized data. W ) represents the mean value within a sliding time window W, where W is 24 hours, and σ(x) W ) represents the standard deviation within the sliding time window W;

[0017] The feature extraction includes calculating terrain roughness, using the following formula:

[0018]

[0019] Where R is the terrain roughness, β is the slope, and α is the aspect. This is the gradient operator.

[0020] Preferably, in step S2, the improved U-Net model includes a multi-branch feature extraction unit, a spatiotemporal fusion module, an attention enhancement module, and a scene adaptive decoding unit; wherein, the multi-branch feature extraction unit includes three parallel branches:

[0021] The meteorological spatiotemporal branch extracts spatial features through 3×3 convolution and captures temporal dynamics using the Transformer self-attention module. The formula for calculating the temporal attention weights is as follows:

[0022]

[0023] Where, α t For temporal attention weights, Softmax is the normalization function, and Q is the value of the time series attention weights. t For time-series query vectors, K t Let d be the time key vector, T be the matrix transpose symbol, and d be the time key vector. k For vector dimensions;

[0024] The terrain branch uses residual blocks to process terrain-derived features, enhancing the ability to extract features from complex terrain areas.

[0025] The multi-source observation branch extracts features from satellite cloud images using 7×7 convolution and radar echoes using 5×5 convolution, adapting to the distribution characteristics of different observation data.

[0026] Preferably, the spatiotemporal fusion module is used to fuse the features output by the multi-branch feature extraction unit, and to achieve the correlation learning of spatial and temporal features through a cross-attention matrix, as shown in the formula:

[0027]

[0028] Among them, M st Let F be the spatiotemporal cross-attention weight matrix. s Let F be a spatial eigenvector, T be the matrix transpose symbol, and F be a spatial eigenvector. t Let σ be the time series feature vector at the current time step. st The spatiotemporal feature covariance is obtained through adaptive learning from the training data; n is the number of historical moments, with a value of 6; F t,i Let be the temporal feature vector of the i-th historical moment.

[0029] Preferably, the attention enhancement module is used to highlight key region features, adjusting the feature importance of different regions through spatial attention weights, as shown in the formula:

[0030]

[0031] Among them, W sHere, represents the spatial attention weights, sigmoid is the activation function, Conv is the 3×3 convolution operation, Cat is the feature concatenation operation, MaxPool is the max pooling operation, AvgPool is the average pooling operation, and F is the input feature map. This represents the terrain roughness gradient.

[0032] Preferably, the scene adaptive decoding unit is used to dynamically adjust the decoding strategy according to the meteorological scene, and determine the branch activation function through scene classification, with the following formula:

[0033]

[0034] Among them, P c Let F be the activation probability of the c-th scene, exp be the exponential function, MLP be the multilayer perceptron, and F be the activation probability of the c-th scene. stat F is the statistical feature vector of the current data. stat,i These are the reference statistical features for the i-th type of scenario.

[0035] Preferably, in step S3, the loss function for fusing terrain features and physical rules is calculated as follows:

[0036]

[0037] Where L is the comprehensive loss value, and y is the high-resolution meteorological true value. Here, R is the predicted value, tanh is the hyperbolic tangent function, and y is the predicted value. wind To predict wind speed, y wind,ref For reference wind speed, y rad To predict irradiance, y rad,max The theoretical maximum irradiance is calculated from the solar altitude angle.

[0038] Preferably, in step S4, the formula for calculating the terrain contribution of the physical attribution result is:

[0039]

[0040] Among them, C R The contribution of W to the terrain ranges from 0 to 1. s Spatial attention weights, Let y represent the terrain roughness gradient, and y represent the high-resolution meteorological ground truth. For predicted values, This represents the absolute error between the predicted value and the true value.

[0041] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0042] (1) This invention constructs an improved U-Net model containing multi-branch feature extraction units, extracts meteorological spatiotemporal features, terrain features and multi-source observation features respectively and fuses them, realizing the deep integration of multi-source meteorological related data, solving the problem that traditional methods cannot make full use of the complementary information of multi-source data, and improving the data utilization rate.

[0043] (2) This invention highlights key regional features by embedding an attention enhancement module and incorporates terrain features into the loss function for weighted optimization, thereby enhancing the ability to capture features of complex terrain regions. This makes the prediction results of high-resolution meteorological data in complex terrain regions more realistic and improves the downscaling accuracy of complex terrain regions.

[0044] (3) This invention constrains prediction errors that violate physical laws by using a loss function that integrates physical rules, and dynamically adjusts the decoding strategy by combining a scene adaptive decoding mechanism, thereby ensuring the physical rationality of the prediction results and improving the reliability of the prediction results under special meteorological scenarios. Attached Figure Description

[0045] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0046] Figure 1 This is a flowchart of a meteorological downscaling method based on spatiotemporal fusion and physical constraints according to the present invention. Detailed Implementation

[0047] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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.

[0048] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0049] Example 1

[0050] like Figure 1 As shown, this invention provides a meteorological downscaling method based on spatiotemporal fusion and physical constraints, comprising the following steps:

[0051] S1. Preprocess the multi-source meteorological data and high-resolution meteorological ground truth to construct a training dataset; the preprocessing includes standardization and feature extraction.

[0052] S2. Construct an improved U-Net model. The model obtains spatiotemporal features and multi-source auxiliary features through a multi-branch feature extraction unit, restores high-resolution information through fusion and decoding, and embeds an attention enhancement module to highlight key region features.

[0053] S3. Train the model using the training dataset, and optimize the parameters using a loss function that integrates terrain features and physical rules to achieve differentiated constraints on prediction errors for complex areas and areas that violate physical laws.

[0054] S4. After performing the preprocessing step S1 on the target low-resolution data, input it into the trained model and output high-resolution meteorological data and physical attribution results.

[0055] Specifically, in step S1, the multi-source meteorological data includes low-resolution meteorological data, topographic data, and multi-source observation data. Continuous data covering complex terrain areas (such as mountainous areas, river valleys, and urban building complexes) from the past 5 years are selected to ensure that the training set includes diverse terrain and meteorological scenarios, such as sunny days, precipitation, high temperatures, and cold waves.

[0056] Furthermore, the low-resolution meteorological data includes at least one of wind speed, temperature, and irradiance, with a resolution of 10km × 10km and a time interval of 1 hour, sourced from the output of a regional numerical weather prediction model. The topographic data is digital elevation model (DEM) data with a resolution of 30m × 30m, used to extract topographic features. The multi-source observation data includes satellite cloud images and radar echoes, where the satellite cloud images have a resolution of 1km × 1km, with one frame every 30 minutes; the radar echoes have a resolution of 0.5km × 0.5km, with one frame every 6 minutes, used to supplement refined observation information. In addition, the high-resolution meteorological ground truth data is obtained from ground-based densified automatic weather station observation data, which, after quality control (outlier removal and missing value filling), serves as the supervision label for model training.

[0057] Furthermore, in step S1, the standardization is adaptive sliding window normalization, specifically:

[0058]

[0059] Where x′ represents the standardized data, x represents the original data, and μ(x) represents the normalized data. W ) represents the mean value within a sliding time window W, where W is 24 hours, and σ(x) W ) represents the standard deviation within the sliding time window W;

[0060] The feature extraction includes calculating terrain roughness, using the following formula:

[0061]

[0062] Where R is the terrain roughness, β is the slope (ranging from 0° to 90°), and α is the aspect (ranging from 0° to 360°). The gradient operator is calculated using a 3×3 Sobel convolution kernel, with the horizontal kernel being:

[0063]

[0064] The vertical kernel is its transpose to enhance the spatial continuity of terrain features. The R value ranges from 0 to 1, with a larger value indicating more complex terrain, thus providing a basis for subsequent attention enhancement and loss weighting.

[0065] In step S2, the improved U-Net model includes a multi-branch feature extraction unit, a spatiotemporal fusion module, an attention enhancement module, and a scene adaptive decoding unit; wherein, the multi-branch feature extraction unit contains three parallel branches:

[0066] The meteorological spatiotemporal branch extracts spatial features through 3×3 convolution, capturing the spatial distribution features of low-resolution meteorological data. It also incorporates the Transformer self-attention module to capture temporal dynamics; the temporal attention weight calculation formula is as follows:

[0067]

[0068] Where, α t For temporal attention weights, Softmax is the normalization function, and Q is the value of the time series attention weights. t For time-series query vectors, K t Let d be the time key vector, T be the matrix transpose symbol, and d be the time key vector. k The vector dimension is used; the input time series length is 6, meaning that the features of the current moment are modeled using data from the previous 6 hours, through α. t The influence weight of historical moments is dynamically allocated; for example, data from the hour before a sudden precipitation event has a higher weight.

[0069] The terrain branch uses residual blocks to process terrain-derived features, enhancing the feature extraction capability for complex terrain areas. Specifically, it uses two residual blocks to process terrain roughness R and slope β. Each residual block contains two 3×3 convolutional layers with ReLU activation in between to solve the gradient vanishing problem in feature extraction for complex terrain areas.

[0070] The multi-source observation branch extracts features using 7×7 convolution for satellite cloud images and 5×5 convolution for radar echoes, adapting to the distribution characteristics of different observation data. Specifically, 7×7 convolution is used to extract large-scale cloud distribution features from satellite cloud images; 5×5 convolution is used to extract local heavy precipitation features from radar echoes. The output features are then convolved with features from other branches after unifying their dimensions using 1×1 convolution.

[0071] The spatiotemporal fusion module is used to fuse the features output by the multi-branch feature extraction unit. It achieves the correlation learning of spatial and temporal features through a cross-attention matrix, as shown in the formula:

[0072]

[0073] Among them, M st The spatiotemporal cross-attention weight matrix is ​​a 7×7 weight matrix composed of 6 historical time points and 1 current time point. Dynamic fusion of spatiotemporal features is achieved through matrix penalty. For example, when wind speed in mountainous areas is influenced by both topography and monsoon, M... st It will automatically increase the weights of the corresponding features of both. F s F is the spatial feature vector (derived from the convolution output of the meteorological spatiotemporal branch), T is the matrix transpose symbol, and F is the spatial feature vector. t Let σ be the time series feature vector at the current time step. st is the spatiotemporal feature covariance, initially set to 0.1, and adaptively learned and updated from the training data to balance the scale differences of spatiotemporal features; n is the number of historical moments, with a value of 6; F t,i Let be the temporal feature vector of the i-th historical moment.

[0074] The attention enhancement module is used to highlight key region features. It adjusts the feature importance of different regions through spatial attention weights, using the following formula:

[0075]

[0076] Among them, W s Here, represents the spatial attention weights, sigmoid is the activation function, Conv is the 3×3 convolution operation, Cat is the feature concatenation operation, MaxPool is the max pooling operation, AvgPool is the average pooling operation, and F is the input feature map. This represents the terrain roughness gradient.

[0077] The scene-adaptive decoding unit is used to dynamically adjust the decoding strategy according to the meteorological scene, and determines the branch activation function through scene classification, as shown in the formula:

[0078]

[0079] Among them, P cLet F be the activation probability of the c-th scene, exp be the exponential function, MLP be the multilayer perceptron, and F be the activation probability of the c-th scene. stat F is the statistical feature vector of the current data. stat,i These are the reference statistical features for the i-th type of scenario.

[0080] In step S3, the formula for calculating the loss function that fuses terrain features and physical rules is:

[0081]

[0082] Where L is the comprehensive loss value, and y is the high-resolution meteorological true value. Here, R is the predicted value, tanh is the hyperbolic tangent function, and y is the predicted value. wind To predict wind speed, y wind,ref For reference wind speed, specifically the 95th percentile of the training data, i.e., the extreme wind speed reference value, y rad To predict irradiance, y rad,max The theoretical maximum irradiance is calculated from the solar altitude angle.

[0083] Specifically, the physical rules constraints include wind speed constraints and irradiance constraints, where the wind speed constraint is: when the predicted wind speed y wind Less than the reference wind speed y wind,ref When, max(0,-y) wind / y wind,ref If ) is positive, the exp term amplifies the error;

[0084] The irradiance constraint is: when the predicted irradiance y rad Greater than the theoretical maximum irradiance y rad,max When, max(0,(y rad -y rad,max ) / y rda,max If ) is positive, the exp term penalizes the predicted value of the superphysical limit.

[0085] Training strategies include:

[0086] Optimizer: The Adam optimizer is used, with an initial learning rate of 0.001 and a decay of 10% every 5 rounds;

[0087] Iteration stopping condition: When the mean absolute error (MAE) of the validation set does not decrease for three consecutive rounds, and the goodness of fit (correlation coefficient R) to the high-resolution true value is achieved. 2 Stop when ) ≥ 0.85;

[0088] Regularization: Add an L2 regularization term (weight 0.0001) to avoid overfitting, and set the training batch size to 32 (to balance memory usage and convergence speed).

[0089] In step S4, the target low-resolution meteorological data (10km × 10km), after standardization and feature extraction in step S1, is stitched together with the corresponding topographic data and observation data. The model output is 1km × 1km high-resolution meteorological data (wind speed, temperature, irradiance), with the spatial grid aligned to the original low-resolution grid using bilinear interpolation, maintaining a time interval of 1 hour. The formula for calculating the topographic contribution of the physical attribution results is as follows:

[0090]

[0091] Among them, C R The topographic contribution is represented by a value ranging from 0 to 1; a larger value indicates that topography is the main source of error. s Spatial attention weights, Let y represent the terrain roughness gradient, and y represent the high-resolution meteorological ground truth. For predicted values, This represents the absolute error between the predicted value and the true value.

[0092] To verify the downscaling effect of the present invention under different meteorological factors (wind speed, light intensity) and typical landforms (plains-nearshore), an offshore wind farm in Hebei Province and a photovoltaic farm in Shandong Province were selected as experimental target areas.

[0093] Example 2

[0094] The selection of experimental target areas focused on the representativeness of typical landforms and meteorological features: a near-shore wind farm in Hebei Province has flat terrain and an elevation difference of less than 20m, with an average annual wind speed of 6-7m / s, and is significantly affected by sea-land wind circulation. The selection of this area aims to verify the ability of this invention to reproduce the fine structure of wind speed fields (such as the wind speed gradient of the coupled sea-land wind boundary layer); a photovoltaic power plant in Shandong Province has an average annual total solar radiation of 5400MJ / m². 2The diurnal and seasonal variations in cloud cover are dramatic, making this region a key area for verifying the downscaling robustness of this invention under scenarios with significant gradients in the illumination field and cloud disturbances. Regarding data sources, the input field uses ECMWF ERA5 reanalysis data with a spatial resolution of 0.25° (approximately 28km × 28km) and a temporal resolution of 1 hour, including wind speed (at a height of 10m) and total radiation (at the surface). The ground truth field uses ERA5-Land high-resolution reanalysis data with a spatial resolution of 0.1° (approximately 11km × 11km) and a temporal resolution of 1 hour. This data has been verified by ground observation stations to have an error of less than 5%, and can be used as the ground truth for the downscaling target. In terms of data partitioning, the training set covers 8784 hours from January 1, 2024 to December 31, 2024, covering the complete annual cycle to include seasonal meteorological characteristics; the independent validation set covers 2184 hours from January 1, 2025 to March 31, 2025, covering the wind speed fluctuations and radiation enhancement period from the end of winter to the beginning of spring, ensuring no data leakage in the model generalization evaluation. In the preprocessing stage, adaptive standardization (sliding window W = 24 hours) and feature extraction are performed on the input field and the ground truth field in step S1, calculating the topographic roughness of the land-sea transition zone in the nearshore wind farm area to be R≈0.1, and the plain topography of the photovoltaic farm area to be R≈0.05.

[0095] Experiments were conducted using an improved U-Net model. The multi-branch feature extraction unit contained a 64-channel convolutional layer, the spatiotemporal fusion module had 6 historical time points (n=6), and the attention enhancement module used a 3×3 convolutional kernel. In the physical constraint loss function, the reference wind speed y... wind,ref Using the 95th percentile of wind speed in the training set (10.2 m / s for the selected near-shore wind farm area), the theoretical maximum irradiance y rad,max Calculated based on solar altitude angle (the selected photovoltaic field area has a maximum of approximately 1000 W / m at noon). 2 The training setup uses the Adam optimizer with an initial learning rate of 0.001, iterating up to the validation set R. 2 The model is stabilized (approximately 30 rounds), with a batch size of 32 to balance training efficiency and model performance. To highlight the technical advantages of this invention, two existing technologies are selected for comparison: one is the traditional statistical downscaling method as Comparative Example 1, which uses a stepwise regression model combined with the statistical relationship between terrain factors and meteorological elements; the other is the conventional deep learning method as Comparative Example 2, namely the U-Net model without embedded physical constraints, whose loss function only uses MAE.

[0096] By conducting comparative tests under the above experimental design and process, the specific results of wind speed downscaling and solar intensity downscaling of the offshore wind farm and the photovoltaic farm are obtained, as shown in Tables 1 and 2 below, and are analyzed.

[0097] Table 1 Results of wind speed reduction scaling in nearshore wind farms

[0098] Evaluation indicators Example 2 Comparative Example 1 Comparative Example 2 MAE (m / s) 0.32 0.87 0.45 RMSE (m / s) 0.48 1.12 0.67 <![CDATA[R 2 ]]> 0.92 0.61 0.83 MAPE (%) 5.3 14.8 7.6

[0099] As shown in Table 1, in the wind speed downscaling task, all evaluation indicators of this invention are significantly better than the traditional statistical method in Comparative Example 1 and the conventional U-Net model in Comparative Example 2. This invention has the lowest MAE and RMSE, and R... 2 The highest value and the lowest MAPE value indicate that it has the highest accuracy in downscaling the wind speed field, especially when capturing complex wind speed gradient changes such as the transition between sea and land winds. This fully demonstrates the technical advantages of multi-branch feature extraction, spatiotemporal fusion and physical constraint mechanisms in wind speed downscaling.

[0100] Table 2 Results of Photovoltaic Electric Field Irradiance Downscaling

[0101] Evaluation indicators Example 2 Comparative Example 1 Comparative Example 2 <![CDATA[MAE(W / m 2 )]]> 28.6 76.3 41.2 <![CDATA[RMSE(W / m 2 )]]> 42.5 98.7 59.8 <![CDATA[R 2 ]]> 0.91 0.58 0.82 MAPE (%) 6.2 16.5 8.9

[0102] As shown in Table 2, in the task of downscaling illumination intensity, the evaluation metrics of this invention are significantly better than those of the traditional statistical method in Comparative Example 1 and the conventional U-Net model in Comparative Example 2. This invention has the lowest MAE and RMSE, and R... 2 The highest value and the lowest value of MAPE indicate that it has higher downscaling accuracy for the illumination field, especially in complex scenarios such as cloud disturbances, where it effectively avoids extreme errors beyond physical limits by means of physical constraint mechanisms.

[0103] Furthermore, in wind farm and photovoltaic power plant scenarios with typical plain-coastal topography, the MAE and RMSE of the downscaling results of this invention are significantly lower than those of traditional methods and conventional deep learning methods. 2 The values ​​are all no less than 0.91, indicating higher accuracy in restoring the fine structure of wind speed and the illumination gradient. At the same time, the physical constraint mechanism effectively avoids extreme errors that violate meteorological laws, and the MAPE is controlled within 7%. It has practical value for power generation prediction and operation and maintenance decision-making in engineering applications, and verifies the universality of the invention under different resource types and terrain conditions.

[0104] Therefore, the above-mentioned meteorological downscaling method based on spatiotemporal fusion and physical constraints is adopted. Through multi-branch feature extraction, spatiotemporal fusion, attention enhancement, and loss function optimization that integrates terrain and physical rules, accurate downscaling of low-resolution meteorological data is achieved. This improves the prediction accuracy and physical rationality of high-resolution meteorological data in complex terrain areas and special meteorological scenarios, and can output physical attribution results to clarify influencing factors.

[0105] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A meteorological downscaling method based on spatiotemporal fusion and physical constraints, characterized in that, Includes the following steps: S1. Preprocess the multi-source meteorological data and high-resolution meteorological ground truth to construct a training dataset; the preprocessing includes standardization and feature extraction. S2. Construct an improved U-Net model. The model obtains spatiotemporal features and multi-source auxiliary features through a multi-branch feature extraction unit, restores high-resolution information through fusion and decoding, and embeds an attention enhancement module to highlight key region features. In step S2, the improved U-Net model includes a multi-branch feature extraction unit, a spatiotemporal fusion module, an attention enhancement module, and a scene adaptive decoding unit; wherein, the multi-branch feature extraction unit comprises three parallel branches: The meteorological spatiotemporal branch extracts spatial features through 3×3 convolution and captures temporal dynamics using the Transformer self-attention module. The formula for calculating the temporal attention weights is as follows: ； in, The weights are for temporal attention, and Softmax is the normalization function. For time-series query vectors, For time key vectors, This is the matrix transpose symbol. For vector dimensions; The terrain branch uses residual blocks to process terrain-derived features, enhancing the ability to extract features from complex terrain areas. The multi-source observation branch extracts features from satellite cloud images using 7×7 convolution and radar echoes using 5×5 convolution, adapting to the distribution characteristics of different observation data. The spatiotemporal fusion module is used to fuse the features output by the multi-branch feature extraction unit. It achieves the correlation learning of spatial and temporal features through a cross-attention matrix, as shown in the formula: ； in, This is the spatiotemporal cross-attention weight matrix. For spatial feature vectors, This is the matrix transpose symbol. This is the time series feature vector at the current time. The spatiotemporal feature covariance is obtained through adaptive learning from the training data; This represents the number of historical moments, with a value of 6. For the first Each historical moment's time-series feature vector; The attention enhancement module is used to highlight key region features. It adjusts the feature importance of different regions through spatial attention weights, using the following formula: ； in, Here, represents the spatial attention weights, sigmoid is the activation function, Conv is the 3×3 convolution operation, Cat is the feature concatenation operation, MaxPool is the max pooling operation, and AvgPool is the average pooling operation. For the input feature map, For terrain roughness gradient; The scene-adaptive decoding unit is used to dynamically adjust the decoding strategy according to the meteorological scene, and determines the branch activation function through scene classification, as shown in the formula: ； in, For the first The activation probability of a scene type, where exp is an exponential function, and MLP is a multilayer perceptron mechanism. This is the statistical feature vector of the current data. For the first Reference statistical features for similar scenarios; S3. Train the model using the training dataset, and optimize the parameters using a loss function that integrates terrain features and physical rules to achieve differentiated constraints on prediction errors for complex areas and areas that violate physical laws. S4. After performing the preprocessing step S1 on the target low-resolution data, input it into the trained model and output high-resolution meteorological data and physical attribution results.

2. The meteorological downscaling method based on spatiotemporal fusion and physical constraints according to claim 1, characterized in that, In step S1, the multi-source meteorological data includes low-resolution meteorological data, topographic data, and multi-source observation data.

3. The meteorological downscaling method based on spatiotemporal fusion and physical constraints according to claim 2, characterized in that, The low-resolution meteorological data includes one or more of wind speed, temperature, and irradiance; the terrain data is digital elevation model data; and the multi-source observation data includes satellite cloud images and radar echoes.

4. The meteorological downscaling method based on spatiotemporal fusion and physical constraints according to claim 1, characterized in that, In step S1, the standardization is adaptive normalization, specifically: ； in, For standardized data, This is the original data. Sliding time window The mean within, Take 24 hours Sliding time window Within the standard deviation; The feature extraction includes calculating terrain roughness, using the following formula: ； in, For terrain roughness, For slope, Slope direction, This is the gradient operator.

5. The meteorological downscaling method based on spatiotemporal fusion and physical constraints according to claim 1, characterized in that, In step S3, the formula for calculating the loss function that fuses terrain features and physical rules is: ； in, The total loss value, For high-resolution meteorological true values, For predicted values, Let be the terrain roughness, and tanh be the hyperbolic tangent function. To predict wind speed, For reference wind speed, To predict irradiance, The theoretical maximum irradiance is calculated from the solar altitude angle.

6. The meteorological downscaling method based on spatiotemporal fusion and physical constraints according to claim 1, characterized in that, In step S4, the formula for calculating the terrain contribution of the physical attribution result is as follows: ； in, The contribution to the terrain is represented by a value ranging from 0 to 1. Spatial attention weights, For terrain roughness gradient, For high-resolution meteorological true values, For predicted values, This represents the absolute error between the predicted value and the true value.

Images