Multi-element deep learning correction method and device for numerical mode forecasting

Abstract

The embodiment of the invention provides a multi-element deep learning correction method and device for numerical mode forecasting. The method is applied to the technical field of data processing, and comprises the following steps: marking observation points, and performing loss function construction on the positions of the observation points to obtain a model optimization target; carrying out iterative updating on model parameters, and carrying out training process optimization by adopting a dynamic learning rate adjustment strategy to obtain a forecast field correction model; and applying the prediction field correction model to target numerical mode prediction data acquired in real time, performing prediction effect evaluation, and performing standardization processing and anti-standardization processing to obtain a correction prediction result. According to the method, the information extraction capability of a deep learning visual algorithm model is utilized, fine deviation correction of data of a numerical mode forecast space grid by meteorological station observation data is completed, meanwhile, multiple types of meteorological elements are fused, the deep learning model is utilized to learn mutual influences and restrictions between related elements, and the prediction precision of the numerical mode forecast space grid is improved. And finer deviation correction is completed.

Description

Technical Field

[0001] The present disclosure relates to the field of data processing, and in particular to a multi-factor deep learning correction method and device for numerical model forecasts. Background Art

[0002] Weather forecasting is a crucial factor influencing human production and daily activities, impacting all aspects of society. Effective weather forecasts can help people prepare for production and daily life, and avoid losses caused by natural disasters. With continued economic growth, the demand for higher-precision weather forecasts is increasing. Numerical model forecasting is one of the most important methods for weather forecasting, and numerical weather forecasting methods have been around for over 100 years since their introduction. In recent decades, numerical model forecasting has continued to improve and innovate, progressing towards more refined predictions and becoming an indispensable part of current weather forecasting. To address these shortcomings, the meteorological community has experimented with a variety of correction methods, including statistical correction, dynamic correction, and integrated forecasting. These methods have, to a certain extent, improved the accuracy of numerical weather forecasts.

[0003] However, because meteorological observation data primarily comes from station data collection, numerical model forecast data is spatially gridded. Conventional statistical methods and integrated forecasting are limited in their ability to correct for deviations from "point" to "surface." While dynamical correction methods use assimilation techniques to optimize model initial fields and boundary conditions to improve numerical model forecasts, their effectiveness is limited as errors accumulate over time. Furthermore, due to the complexity and unpredictability of the atmospheric system, the uncertainty of the initial fields used in numerical models simulating the atmosphere, and the inherent flaws of the models themselves, numerical model forecasts still suffer from significant errors. Summary of the Invention

[0004] The present disclosure provides a multi-factor deep learning correction method and device for numerical model forecasts, which utilizes the information extraction capability of a deep learning visual algorithm model to complete refined deviation correction of the data of the numerical model forecast spatial grid using meteorological station observation data. At the same time, it integrates multiple types of meteorological elements and uses a deep learning model to learn the mutual influence and constraints between related elements to complete more refined deviation correction.

[0005] According to a first aspect of the present disclosure, a multi-factor deep learning correction method for numerical model forecasts is provided, comprising:

[0006] Perform spatiotemporal matching and outlier screening on historical numerical model forecast data and historical meteorological station observation data to obtain sample pair time series file path sequences, and generate statistical information files based on the statistics of meteorological element characteristics and terrain characteristics;

[0007] Reading data from a time series file path sequence according to the sample, converting meteorological element features into tensors through netCDF format conversion, and performing Z-Score standardization processing using the statistical information file to obtain a meteorological feature standardized tensor;

[0008] The normalized tensor of meteorological features is input into the encoder of DeepLab V3+ neural network, feature extraction is performed through the ResNet series network, multi-scale feature fusion is performed through the dilated pyramid pooling module, and the decoder processes it to obtain a prediction result consistent with the input feature dimension;

[0009] According to the location distribution of the historical meteorological station observation data, the observation points are marked by constructing a mask tensor, and a loss function is constructed for the observation point positions to obtain the model optimization target;

[0010] According to the model optimization objectives, the model parameters are iteratively updated through the Adam optimizer, and the training process is optimized using a dynamic learning rate adjustment strategy to obtain a forecast field correction model;

[0011] The forecast field correction model is applied to the target numerical model forecast data collected in real time, and the forecast effect is evaluated by the root mean square error and mean absolute error. The corrected forecast result is obtained after normalization and denormalization.

[0012] According to a second aspect of the present disclosure, a multi-factor deep learning correction device for numerical model forecasts is provided, comprising:

[0013] The screening module is used to perform spatiotemporal matching processing and outlier screening on historical numerical model forecast data and historical meteorological station observation data, obtain the sample pair time series file path sequence, and generate statistical information files based on the statistics of meteorological element characteristics and terrain characteristics;

[0014] A reading module is used to read data from a time series file path sequence according to the sample, perform tensor conversion on meteorological element features through netCDF format conversion, perform Z-Score normalization processing using the statistical information file, and obtain a normalized tensor of meteorological features;

[0015] An extraction module is used to input the normalized tensor of meteorological features into the encoder of the DeepLab V3+ neural network, perform feature extraction through the ResNet series network, perform multi-scale feature fusion through the dilated pyramid pooling module, and process it by the decoder to obtain a prediction result consistent with the input feature dimension;

[0016] A marking module is used to mark observation points by constructing a mask tensor based on the location distribution of the historical meteorological station observation data, and to construct a loss function for the observation point location to obtain a model optimization target;

[0017] An updating module is used to iteratively update the model parameters through the Adam optimizer according to the model optimization objective, and to optimize the training process using a dynamic learning rate adjustment strategy to obtain a forecast field correction model;

[0018] The evaluation module is used to apply the forecast field correction model to the target numerical model forecast data collected in real time, evaluate the forecast effect through the root mean square error and mean absolute error, and obtain the corrected forecast result after standardization and de-standardization.

[0019] The present invention achieves high-quality preprocessing of meteorological data by performing spatiotemporal matching processing and outlier screening on historical numerical model forecast data and historical meteorological station observation data, and combining the generated samples to the time series file path sequence and statistical information file, thereby ensuring the data quality of subsequent deep learning model training. The meteorological element features are converted into tensors through netCDF format conversion, and the statistical information file is used for Z-Score standardization, so that meteorological element data of different dimensions can be processed at a unified scale, effectively improving the training effect of the model. The encoder structure of the DeepLabV3+ neural network is adopted, combined with the ResNet series network for feature extraction, and multi-scale feature fusion is performed through the hollow pyramid pooling module, which realizes the effective extraction and fusion of different scale features of meteorological data. During the model training process, the observation points are marked by constructing a mask tensor and the loss function is constructed for the observation point position, which effectively solves the problem of uneven spatial distribution of meteorological observation data. The Adam optimizer is used to iteratively update the model parameters, and the dynamic learning rate adjustment strategy is used to ensure the stability and convergence of the model training. Finally, the forecast effect was evaluated by the root mean square error and mean absolute error, and after standardization and de-standardization, accurate revised forecast results were obtained, completing high-quality forecast correction from point to surface and realizing refined correction of numerical model forecasts.

[0020] It should be understood that the contents described in the Summary of the Invention section are not intended to limit the key or important features of the embodiments of the present disclosure, nor are they intended to limit the scope of the present disclosure. Other features of the present disclosure will become readily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] The above and other features, advantages and aspects of the embodiments of the present disclosure will become more apparent with reference to the following detailed description in conjunction with the accompanying drawings. The accompanying drawings are provided for a better understanding of the present disclosure and do not constitute a limitation of the present disclosure. In the accompanying drawings, the same or similar reference numerals represent the same or similar elements, among which:

[0022] Figure 1A flowchart of a multi-factor deep learning correction method for numerical model forecasting according to an embodiment of the present disclosure is shown;

[0023] Figure 2 This is a schematic diagram of the network structure of DeepLab V3+ in the embodiment of this application;

[0024] Figure 3 A block diagram of a multi-factor deep learning correction device for numerical model forecasting according to an embodiment of the present disclosure is shown. DETAILED DESCRIPTION

[0025] To make the purpose, technical solutions, and advantages of the embodiments of the present disclosure more clear, the technical solutions in the embodiments of the present disclosure will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present disclosure, not all of the embodiments. Based on the embodiments of the present disclosure, all other embodiments obtained by ordinary technicians in this field without making any creative efforts are within the scope of protection of the present disclosure.

[0026] In this document, the term "and / or" simply describes a relationship between related objects, indicating that three possible relationships exist. For example, "A and / or B" can represent: A exists alone, A and B exist simultaneously, or B exists alone. Furthermore, the character " / " in this document generally indicates that the related objects are in an "or" relationship.

[0027] Figure 1 FIG. 1 is a flow chart of a multi-factor deep learning correction method 100 for numerical model forecasting according to an embodiment of the present disclosure. Figure 1 As shown, the method 100 includes:

[0028] S110: Performing spatiotemporal matching processing and outlier screening on historical numerical model forecast data and historical meteorological station observation data to obtain a sample pair time series file path sequence, and generating a statistical information file based on the statistics of meteorological element characteristics and terrain characteristics;

[0029] Optionally, historical numerical model forecast data and historical meteorological site observation data are paired according to time and spatial location, and observation site data outside the forecast area are removed through data verification; outlier detection is performed on the paired data, and samples containing abnormal data and serious missing measurements are eliminated to obtain valid sample data; indexes are established in chronological order based on the valid sample data to generate a sample pair time series file path sequence; meteorological element features are extracted from the historical numerical model forecast data, and terrain features are extracted from the terrain data to obtain a feature data set; statistical analysis is performed on the feature data set to calculate the mean, variance, maximum and minimum values of each feature; the mean, variance, maximum and minimum values of each feature are integrated to generate a statistical information file.

[0030] The spatiotemporal matching of historical numerical model forecast data and historical meteorological station observation data requires first clarifying the temporal resolution and spatial range of the numerical model forecast data. Numerical model forecast data are typically stored as grid points, each with clear longitude and latitude coordinates, while meteorological station observation data are discrete point observations. During the spatiotemporal matching process, the latitude and longitude information of the observation station is first compared with the numerical model forecast grid to identify the nearest model grid point for each observation station. In the temporal dimension, the observation time is aligned with the model forecast time to ensure that the time stamps of the two data types are completely consistent. For observation stations located outside the model forecast area, since corresponding model forecast data are unavailable, these stations' data must be excluded during the data verification phase. Outlier detection on the matched data is a key step in ensuring data quality. Outlier detection primarily involves two aspects: meteorological physical plausibility testing and statistical characteristic testing. In this process, reasonable value ranges are set based on the physical properties of different meteorological elements. For example, relative humidity should range from 0 to 100%, and temperature data should conform to local climate characteristics. In the statistical feature test, the historical statistical features of each element of each site (such as mean and standard deviation) are calculated, and data that deviates from the mean by more than a number of standard deviations are marked as suspected outliers. In addition, the temporal continuity of the data needs to be checked, and data points with sudden changes need to be marked. For the marked outliers, analysis and judgment are performed in combination with meteorological professional knowledge to confirm whether they are true outliers. At the same time, statistics are also needed for data missing situations. When the missing rate of a site in a specific period exceeds the set threshold, all samples of the site in that period are eliminated. Through these processes, valid sample data with reliable quality is finally obtained.

[0031] The process of constructing a sequence of sample-pair time-series file paths based on valid sample data is essentially establishing a data indexing system. This system must include key elements such as the data file storage path, time information, and spatial information. For each valid sample, the corresponding numerical model forecast data file path and observation data file path are recorded, and this information is organized in chronological order. This organization facilitates subsequent data reading and model training based on time series characteristics. To improve data reading efficiency, the path sequence is typically stored in a specific file format (such as JSON or CSV), and necessary metadata is appended to the sequence. During the feature extraction phase, relevant features must be extracted from historical numerical model forecast data and terrain data. Meteorological element features primarily include conventional meteorological elements such as temperature, humidity, air pressure, wind direction and speed, as well as their temporal derivatives (reflecting trends) and spatial derivatives (reflecting spatial gradients). Topographic features include information such as altitude, terrain slope, and terrain orientation. These features have a significant impact on the formation and development of local weather systems. The feature extraction process must consider the physical correlations between different features to ensure that the extracted features fully reflect the physical characteristics of the meteorological system.

[0032] Statistical analysis of the extracted feature data set is a preparation for subsequent data standardization. Statistical analysis involves calculating the statistics of each feature in the time and space dimensions. For each feature, its mean (reflecting the overall level of the feature), variance (reflecting the degree of fluctuation of the feature), maximum value, and minimum value (determining the range of the feature's value) need to be calculated. The calculation of these statistics needs to take into account the temporal and spatial distribution characteristics of the data to ensure that the statistical results are representative. Finally, the calculated statistics are integrated into a statistical information file, which will serve as an important basis for subsequent data standardization processing. The file needs to clearly record the various statistics of each feature, and indicate the time range and spatial range of the statistics.

[0033] It should be noted that, in the embodiment of the present application, the numerical model forecast data of the past two years (the example of the present invention adopts the WRF numerical model, and the hourly forecast of the Chinese region) and the historical weather station observation data of the same time period are collected. The observation data are analyzed, the data time series corresponding to the key weather phenomena are screened, and a suitable amount of samples are randomly extracted from each month to construct training data. Different forecast time effectiveness are divided into different groups. If WRF is a 7-day forecast data, seven training data groups can be constructed for different forecast time effectiveness and errors, so as to build and train models. The model forecast data and the site observation data are matched in time and space, and the sites outside the numerical model forecast area are removed. At the same time, data verification is performed to remove sample pairs containing abnormal data, serious missing values, etc. After matching, the sample pairs generate a time series file path sequence for quick reading and matching. The mean, variance, maximum value, minimum value of the data such as the alternative meteorological element characteristics and terrain characteristics of the full data set are counted, and a statistical information file is generated.

[0034] For example, consider obtaining WRF model temperature forecast data and meteorological station temperature observation data for a specific region over the past two years. The WRF model data is grid-point data, covering the entire forecast area with a temporal resolution of 1 hour, while the observation stations are distributed across the forecast area. Spatial matching reveals that some observation stations are located outside the forecast area boundary, and these station data need to be excluded. For stations within the forecast area, temporal matching and outlier detection are performed. Physical threshold testing reveals that some observations significantly deviate from the normal range, such as extreme values that clearly exceed local climate characteristics. These data are marked as outliers. Furthermore, it is found that some stations have continuous missing measurements during specific periods, and samples from these periods are also excluded. For valid samples that pass quality control, their file paths are recorded and indexed in chronological order. Basic features (such as the temperature value itself) and derived features (such as the temporal rate of temperature change and spatial gradient) are then extracted from the temperature field, along with corresponding topographic features. Statistical analysis is performed on these features, and various statistics are calculated and saved in a statistical information file.

[0035] S120: Read data from the time series file path sequence according to the sample, convert the meteorological element features into tensors through netCDF format conversion, perform Z-Score standardization processing using the statistical information file, and obtain a meteorological feature standardized tensor;

[0036] Optionally, numerical model forecast data, terrain data and meteorological station observation data are extracted according to the sample sequence of the time series file path; meteorological element feature data is read from the netCDF data format and converted into a first tensor data; the terrain data is clipped and interpolated according to the longitude and latitude to generate a second tensor data; the target feature data is read from the meteorological station observation data, and a third tensor data with the same dimension as the numerical model forecast data is constructed, and the missing positions are filled with preset default values; according to the mean and standard deviation in the statistical information file, the first tensor data, the second tensor data and the third tensor data are Z-Score standardized; the standardized first tensor data, the second tensor data and the third tensor data are combined to obtain a meteorological feature standardized tensor.

[0037] Specifically, data is retrieved from a sequence of time-series file paths based on samples. This sequence contains the storage path information for numerical model forecast data, terrain data, and meteorological station observation data. The numerical model forecast data records the forecast values of meteorological elements at each grid point within the forecast area, the terrain data contains the geographic information of the forecast area, and the meteorological station observation data represents the actual observed values of meteorological elements. Next, meteorological element feature data is read from a netCDF format file. NetCDF (Network Common Data Form) is a self-describing, machine-independent data format for array-type data, widely used for scientific data storage in fields such as meteorology and oceanography. The reading process first requires obtaining basic file information, including dimension definitions and variable attributes. The required meteorological element data is then extracted based on the variable names. This data is organized into multidimensional arrays, where the dimensions correspond to spatial coordinates (longitude and latitude) and time. These multidimensional arrays are converted into tensor format, resulting in the first tensor data, which contains the temporal and spatial distribution information of meteorological elements. The processing of terrain data requires consideration of spatial consistency with the numerical model forecast data. The original terrain data usually has independent spatial resolution and coverage, and needs to be clipped and interpolated according to the latitude and longitude grid of the numerical model forecast data. The clipping process first determines the latitude and longitude range of the forecast area, and then extracts the data of the corresponding area from the terrain data. Since the resolution of the terrain data may be inconsistent with the model grid, the terrain data needs to be resampled to the same resolution as the model grid through interpolation methods. Commonly used interpolation methods include bilinear interpolation, cubic spline interpolation, etc. The selection of a suitable interpolation method needs to consider computational efficiency and accuracy requirements. After clipping and interpolation, the second tensor data is obtained, which completely corresponds to the first tensor in terms of spatial dimensions.

[0038] For meteorological station observation data, the first step is to read the observed values of the target features from the data file. Observational data typically exists as discrete points and needs to be converted to gridded data with the same dimensions as the numerical model forecast data. This process begins by constructing an empty tensor of the same size as the model grid. Then, based on the latitude and longitude information of the observation site, the observation values are populated at the nearest grid point. Grid points without observation sites are filled with preset default values (such as -999 or NaN), which are handled specially in subsequent processing. This results in a third tensor with the same spatial dimensions as the first two tensors. Normalization is a crucial step in converting data of different dimensions to the same scale. Z-score normalization is a commonly used normalization method that transforms data using the mean and standard deviation. For each tensor, the mean and standard deviation of the corresponding feature are read from the statistics file and the data is transformed according to the normalization formula. The normalized data exhibits a distribution with a mean of 0 and a standard deviation of 1, which helps improve the training performance of deep learning models.

[0039] Finally, the three standardized tensor data are combined to produce the final standardized tensor of meteorological features. This combination process takes into account the physical correlations between features and rationally arranges the order of features within the tensor. This final tensor serves as input to the deep learning model and is used for subsequent forecast corrections.

[0040] It should be noted that in the embodiment of the present application, the data loader (DataLoader) is constructed. The main function of the data loader is to smoothly load the original data set into the model training. The process generally includes data format conversion, data batch loading, data processing, etc. In the example of the present invention, the construction of the data loader should include the following steps:

[0041] For the target sample pair, the numerical model forecast data, terrain data and corresponding meteorological station observation data are read separately according to the path and format. For the numerical model data, the characteristic data of each meteorological element is read from the netCDF data format and converted into a tensor; in addition, the terrain feature features are read from the terrain file, clipped and interpolated into a grid consistent with the numerical model according to the longitude and latitude, and converted into a tensor; for the meteorological station observation data, all station data of the corresponding time target element are read from the text, and a tensor of the same size as the numerical model and filled with default values (values that can be distinguished from the target value, such as -999, nan values, etc.) is constructed. The target element's neighbors are interpolated into this tensor according to the longitude and latitude of the station and the longitude and latitude of the numerical model grid, and used as the target value for training. The statistical information file is read and the meteorological element feature and terrain feature tensors are standardized. In this example, the Z-Score standardization method is used, and the calculation method is shown in the following formula:

[0042]

[0043] Where: z m，i is the data value of feature m after normalization in the i-th grid; x m，i is the value of the original feature m at the i-th grid; μ m is the mean of feature m; σ m is the standard deviation of feature m.

[0044] Define the getitem and len methods to return the input feature data and label values required by the model. Split the dataset into training and validation sets, build data loaders for each set, and set appropriate parameters such as batch_size, shuffle, and num_workers.

[0045] For example, first obtain the relevant data path at a certain moment from the sample pair time series file path sequence. Read the temperature field data at that moment from the netCDF file, including temperature values, spatial distribution information, etc., and convert it into the first tensor. At the same time, read the terrain data of the forecast area, and resample the terrain data to the same grid resolution as the temperature field through bilinear interpolation to form the second tensor. For the observation data, extract the observation value at that moment from the temperature record of the observation station, establish a grid of the same size as the temperature field, fill the temperature value of the observation station into the corresponding grid position, and fill the other positions with default values to obtain the third tensor. Then, according to the statistics of the temperature field, terrain height and other features recorded in the statistical information file, the three tensors are standardized separately. Finally, these standardized feature tensors are combined according to the physical meaning to obtain a complete feature tensor. This tensor contains relevant information such as temperature forecast values, terrain characteristics and observation values, providing input data for the subsequent correction model.

[0046] S130: The normalized tensor of meteorological features is input into the encoder of the DeepLab V3+ neural network. Feature extraction is performed through the ResNet series network. Multi-scale feature fusion is performed through the dilated pyramid pooling module. The decoder processes the tensor to obtain a prediction result consistent with the input feature dimension.

[0047] Optionally, in some embodiments, the meteorological feature normalized tensor is input into the DeepLab V3+ neural network to generate an intermediate layer low-level feature map and a last layer convolution feature map; the last layer convolution feature map is processed by four hole convolution blocks with different expansion rates, and each hole convolution block is equipped with a batch normalization layer and an activation layer to obtain a first group of feature maps; the last layer convolution feature map is subjected to global average pooling processing, and processed by a convolution block, equipped with a batch normalization layer and an activation layer to obtain a second group of feature maps; the first group of feature maps and the second group of feature maps are connected, processed by a convolution block and equipped with a batch normalization layer, an activation layer and a dropout layer to obtain a fused feature map; the intermediate layer low-level feature map is subjected to 1×1 convolution dimensionality reduction processing, the fused feature map is upsampled to make it the same size as the low-level feature map after dimensionality reduction, and the two are spliced and processed by a 3×3 convolution block; the feature map obtained by splicing and processing by the 3×3 convolution block is linearly interpolated and upsampled to obtain a prediction result consistent with the input feature dimension.

[0048] Specifically, the normalized tensor of meteorological features is input into the DeepLab V3+ neural network, and feature extraction is performed using the ResNet series network as the backbone network. During the feature extraction process, the ResNet network performs layer-by-layer processing through multiple residual blocks, generating two different levels of feature maps: intermediate low-level feature maps and final convolutional feature maps. The intermediate low-level feature maps retain detailed information about the input data and are important for accurately locating the spatial distribution of meteorological elements, while the final convolutional feature maps contain higher-level abstract features and can capture the complex relationships between meteorological elements. The final convolutional feature map is processed using dilated convolution technology, constructing dilated convolution blocks by setting four different dilation rates. Dilated convolution is a special convolution operation that expands the receptive field without increasing the number of parameters by inserting holes between the convolution kernel elements. The four different dilation rates enable the network to capture meteorological features at different scales, which is very important for processing cross-scale meteorological phenomena. Each dilated convolutional block is equipped with a batch normalization layer and an activation layer. The batch normalization layer is used to stabilize the network training process and reduce internal covariate shift, while the activation layer introduces nonlinear transformations to enhance the network's expressiveness. The outputs of these four dilated convolutional blocks constitute the first set of feature maps.

[0049] At the same time, the last layer of convolutional feature maps is subjected to global average pooling, an operation that captures the global information of the entire feature map. The features after global average pooling are processed through a convolution block, which is also equipped with a batch normalization layer and an activation layer to obtain a second set of feature maps. The introduction of global average pooling enables the network to fully utilize the contextual information of the entire forecast area, which helps to improve the accuracy of forecast corrections. The connection between the first and second sets of feature maps is achieved by splicing in the channel dimension. The connected feature maps are processed by a convolution block, which contains a convolution layer, a batch normalization layer, an activation layer, and a dropout layer. The dropout layer reduces the risk of overfitting by randomly discarding some neurons, thereby improving the generalization ability of the model. The fused feature map obtained by this series of processing integrates multi-scale features and global contextual information.

[0050] The low-level feature maps in the intermediate layer are subjected to 1×1 convolution dimensionality reduction. This operation reduces the number of feature channels, reducing the amount of computation while retaining important feature information. The fused feature map is then upsampled to match the size of the reduced low-level feature map using bilinear interpolation. The two are then concatenated and processed through a 3×3 convolution block, effectively fusing high-level semantic information with low-level detail features. The final step is to upsample the processed feature map to the same dimension as the input features using linear interpolation to obtain the final prediction result. This upsampling process ensures that spatial resolution is restored while maintaining feature continuity.

[0051] In the embodiment of the present application, DeepLab V3+ is a high-performance semantic segmentation model. The present invention has appropriately modified and adjusted the standard DeepLab V3+ network structure to give full play to its powerful multi-scale feature extraction ability, detail retention ability and high computational efficiency to achieve intelligent and efficient numerical model deviation correction function. Compared with other similar deep learning models, the correction model based on DeepLab V3+ can more accurately locate the forecast deviation area and has a faster inference speed, which is very suitable for application in numerical model correction scenarios. The complete network structure of DeepLab V3+ is as follows: Figure 2 As shown in the figure, it is a typical Encoder-Decoder structure, which mainly includes two parts: encoder and decoder. The specific structure and feature extraction calculation process are as follows: For the construction of the encoder, it mainly includes the backbone network (backbone, i.e. Figure 2The DCNN in

[15] and the Atrous Pyramid Pooling (ASPP) module are also included. There are two common backbone network architectures: the ResNet series and the improved Xception. This example uses the ResNet series. The meteorological tensor features are passed to the backbone network, and the feature extraction and calculation process is as follows:

[0052] The input is a meteorological feature tensor composed of multiple superimposed elements. After feature extraction and computation, multiple ResNet modules in the backbone network produce two outputs: one is the intermediate low-level feature maps, which are directly output to the decoder module; the other is the feature map output from the final convolutional layer, which is processed by the ASPP module. The ASPP module consists of four dilated convolution blocks with different dilation rates, each of which includes convolution, batch normalization, and activation layers. There is also a global average pooling block, which also includes pooling, convolution, batch normalization, and activation layers. The input feature maps pass through these five modules in parallel to produce five sets of feature maps. These five sets of feature maps are concatenated and passed through a 1×1 convolution block (including convolution, batch normalization, activation, and dropout layers) before being input to the decoder module. The decoder module primarily receives low-level feature maps from the intermediate layers of the backbone network in the encoder and the output of the ASPP module. The processing flow is as follows: First, a 1×1 convolution block is used to perform channel dimensionality reduction on the low-order feature map, and the feature map from ASPP is interpolated and upsampled to make it the same size as the low-order feature map. The two are then concatenated and sent to a 3×3 convolution block for processing. Finally, linear interpolation and upsampling are performed again to obtain a prediction result with the same resolution as the input feature grid.

[0053] For example, when the input meteorological feature normalized tensor contains temperature field data, it is first extracted through the ResNet network to obtain feature maps at different levels. Among them, the middle layer feature map retains the local variation characteristics of the temperature field, while the last layer feature map contains the large-scale distribution characteristics of the temperature field. The last layer feature map is processed by four dilated convolution blocks with different expansion rates (such as 1, 6, 12, and 18), enabling the network to simultaneously capture temperature variation characteristics at different spatial scales. For example, a smaller expansion rate focuses on local temperature gradients, while a larger expansion rate can capture temperature distribution patterns over a wider range. The global average pooling branch provides an overview of the temperature field in the entire forecast area, which helps to understand the large-scale temperature distribution characteristics. These features are fused and reconstructed to ultimately obtain a refined temperature field forecast correction result, which not only retains the spatial continuity of the original temperature field, but also improves the accuracy of the forecast through the fusion of multi-scale features.

[0054] S140: Based on the location distribution of historical meteorological station observation data, the observation points are marked by constructing a mask tensor, and a loss function is constructed for the observation point locations to obtain the model optimization target;

[0055] Optionally, based on the grid dimension data of the location distribution of historical meteorological station observation data, an initial mask tensor is constructed, and preset default values are filled in all grid positions; the latitude and longitude information of the observation station in the historical meteorological station observation data is matched with the latitude and longitude of the numerical model forecast grid data to obtain a matching result; based on the matching result, the grid positions with observation data are marked as 1 in the initial mask tensor, and the grid positions without observation data are marked as 0 to obtain a mask tensor; the square difference between the predicted value and the historical meteorological station observation data is weighted by the mask tensor to generate a mask weighted error; the mask weighted error is normalized by the number of grid points with a value of 1 in the mask tensor to obtain an average error value; the average error value is used as the calculation result of the loss function to obtain the model optimization target.

[0056] Specifically, a masking mechanism is used to address the spatial discontinuity of observational data. First, an initial mask tensor must be constructed. The dimensions of this tensor are consistent with the numerical model forecast grid. For example, if the forecast grid is a 200×300 two-dimensional grid, the initial mask tensor also has the same dimensions. During the construction process, all grid points are uniformly filled with a preset default value of 0, thus creating a "blank" mask template. To accurately mark the location of the observation station, the longitude and latitude information of the observation station in the historical meteorological station observation data must be matched with the longitude and latitude of the numerical model forecast grid. This matching process uses the nearest neighbor matching principle. That is, for each observation station, the closest grid point in the numerical model forecast grid is found. The matching process first calculates the distance from each observation station to all grid points, and then selects the grid point with the smallest distance as the corresponding location of the observation station. This matching method ensures that the observation data can be accurately mapped to the forecast grid.

[0057] Based on the above matching results, the initial mask tensor is updated. At the grid points corresponding to the observation sites, the mask value is set to 1, indicating that there are actual observation data at these locations; other locations retain the original value of 0, indicating that there are no observation data. The mask tensor obtained in this way is actually a binary matrix that clearly marks the spatial distribution of the observation data. Next, the mask tensor is used to weight the squared difference between the predicted value and the historical meteorological station observation data. First, the squared difference between the predicted value and the observed value at each grid point is calculated, and then these squared differences are multiplied by the value of the corresponding position in the mask tensor. This process actually selectively retains the error at the location with observation data and ignores the error at the location without observation data.

[0058] The mask-weighted error is normalized based on the number of grid points in the mask tensor with a median value of 1, which is the total number of actual observation sites. This ensures that the loss function calculation is not affected by the uneven distribution of observation sites. The resulting average error, which directly reflects the average deviation between the model's predictions and the observed values, is used as the loss function calculation result.

[0059] It should be noted that the present application has improved the loss function. In order to directly use the site data to correct and match the numerical model grid data, when performing loss calculations, the present invention uses a masking technology to process the model prediction value and the target value. Specifically, when the data loader module passes in the target value, that is, the site observation data, a tensor of the same size filled with all default values is first constructed according to the model forecast grid data, and then the site observation value neighbors are matched to the grid position corresponding to the longitude and latitude of the model grid according to the site longitude and latitude. In this way, a tensor with the same size and position matching as the forecast grid data is formed, where the grid position with the site observation value is filled with the target value of the corresponding meteorological element, and the place without the observation value is the preset default value. From this, the target tensor can quickly construct a mask tensor of the same size, that is, the place with the value is set to 1, and the position of the default value is set to 0. This mask tensor marks the position with the site observation data. Combined with this mask tensor, the calculation of the loss function can be designed. The specific formula is as follows:

[0060]

[0061] Where: H is the height of the tensor (number of rows); W is the width of the tensor (number of columns); y i，j is the true target value:

[0062] is the model prediction value; M i，j is a mask tensor, where M i，j =1 means that the data at this location has an observed value and the data is valid. i，j =0 means there is no observation at that location and the data is invalid.

[0063] First, the squared error loss Element-wise multiplication by the mask tensor M i，j , so that the contribution of those missing observation samples can be ignored. Then sum the weighted losses of all elements, with the weight being the corresponding mask value M i，jIn this example, the mask value is set to 1, indicating that the weights are the same. However, for some tasks that correct factors with a serious imbalance in the number of samples in the numerical range, different mask values can be set according to different target value intervals to alleviate the regression imbalance problem. For example, when correcting precipitation, in order to make the model pay more attention to the forecast of heavy rain level precipitation, the greater the precipitation, the larger the mask value can be set, so that the sample weight is greater and the model pays more attention to it. Finally, divide it by the total number of valid samples. The average loss is obtained. The output of the correction model is a complete two-dimensional data tensor that contains information about all grid points in the forecast area. During the training process, a mask mechanism is used to participate in the calculation of the loss function. The loss is calculated only at the grid points with observations, and the grid points without observations do not participate in the loss calculation. By iteratively optimizing the model parameters, the forecast values of the grid points with observations can be effectively corrected. Since the model learns the spatiotemporal correlation of the entire forecast area, when optimizing the grid points with observations, the grid points without observations will also be optimized "radially", thereby achieving an overall correction of the data in the entire forecast area.

[0064] For example, suppose there are multiple meteorological observation stations distributed across a forecast area, each recording temperature observations. First, construct an initial mask tensor composed entirely of zeros based on the dimensions of the forecast grid. Then, match the latitude and longitude information of each observation station to the forecast grid. For example, if a station is closest to the (100, 150)th grid point in the forecast grid, mark that position in the mask tensor as 1. Repeat this process for all observation stations, ultimately obtaining a binary mask tensor reflecting the distribution of the observation stations.

[0065] When the model predicts the temperature field, it compares the predicted result with the observed value. For example, if the predicted temperature at a certain observation station is 25 degrees Celsius and the actual observed temperature is 23 degrees Celsius, the squared error at that location is 4. This squared error is multiplied by the mask value (here, 1), and this error value is retained. At locations without an observation station, although there are predicted values, because the mask value is 0, the corresponding error does not participate in the loss function calculation. Finally, all retained error values are summed and divided by the total number of observation stations (i.e., the number of grid points with a mask value of 1) to obtain the average error. This average error serves as the value of the loss function and guides the model optimization process. Through appropriate weighting and normalization, accurate corrections are ensured for locations with observed data while preventing interference from locations without observed data on model training. This design also provides flexibility for differentiated processing of different observational elements. For example, for unevenly distributed elements such as precipitation, the mask value can be adjusted to enhance the correction effect within specific value ranges.

[0066] S150: Based on the model optimization objectives, the model parameters are iteratively updated through the Adam optimizer, and the dynamic learning rate adjustment strategy is used to optimize the training process to obtain the forecast field correction model;

[0067] Optionally, the training rounds are divided into six training rounds as one learning cycle, the initial learning rate is set to 0.001, and a training round sequence is generated; the model optimization target is input into the Adam optimizer, and the gradient value of each parameter is calculated by the back propagation algorithm to generate a parameter gradient sequence; based on the parameter gradient sequence, momentum correction and adaptive moment estimation are performed on each parameter to obtain the corrected gradient value; the model parameters are updated by the product of the corrected gradient value and the current learning rate to obtain the optimization result of this round; at the end of each learning cycle, the current learning rate is multiplied by a coefficient of 0.6, and the learning rate of the next learning cycle is attenuated; the loss value of the optimization result of this round is calculated and the convergence is judged. When the loss value is lower than the preset threshold, a forecast field correction model is constructed according to the current parameters.

[0068] Specifically, the training process is divided into multiple learning cycles, each consisting of six training cycles. At the beginning of training, the initial learning rate is set to 0.001. This value was determined based on extensive practical experience to avoid significant fluctuations in the early stages of training while ensuring sufficient parameter update speed. The generation of a training cycle sequence effectively establishes the basic framework of the training process, providing clear node identifiers for subsequent parameter updates and learning rate adjustments. After the model optimization objective is input to the Adam optimizer, the gradient of each parameter in the network is calculated using the backpropagation algorithm. Backpropagation is a core algorithm in deep learning. It uses the chain rule to calculate the partial derivative of the loss function for each parameter layer by layer, starting from the network output layer. This process generates a parameter gradient sequence containing the gradient information of all trainable parameters in the network. The Adam optimizer introduces the concepts of momentum and adaptive learning rate on top of the standard gradient descent method, making the parameter optimization process more stable and efficient.

[0069] The parameter gradient sequence undergoes momentum correction and adaptive moment estimation, two core features of the Adam optimizer. Momentum correction smoothes the direction of parameter updates by accumulating historical gradient information, preventing volatile fluctuations during the optimization process. Adaptive moment estimation automatically adjusts the update step size for each parameter based on the statistical characteristics of the parameter's historical gradients, ensuring that features with varying frequencies receive an appropriate learning rate. The combination of these two processing steps produces a corrected gradient value. This corrected gradient value is multiplied by the current learning rate to obtain the actual parameter update. This update is directly applied to the model parameters, completing a parameter optimization cycle. The optimization results are reflected in the model's predictive performance, and the effectiveness of the optimization can be evaluated through changes in the loss function. After each optimization cycle, a set of updated model parameters is obtained, which determine the model's predictive correction capability for the current cycle.

[0070] At the end of each learning cycle (i.e., six training rounds), the current learning rate is decayed by multiplying it by 0.6. This learning rate decay strategy allows the model to quickly approach the optimal solution region in the early stages of training, while performing more refined parameter adjustments in the later stages of training. The gradual reduction of the learning rate helps the model find a more accurate parameter configuration. Finally, the loss value is calculated for the results of each round of optimization and compared with a preset threshold. When the loss value drops below the threshold, it indicates that the model has achieved the expected forecast correction accuracy. At this time, the current parameter configuration is saved and the final forecast field correction model is constructed. The selection of this preset threshold requires a comprehensive consideration of the forecast accuracy requirements and computing resource constraints.

[0071] In the embodiment of the present application, the Adam adaptive learning rate optimization algorithm is used to complete the back propagation parameter update process. In order to further improve the computational efficiency and the robustness of the model, an adjustment strategy for the learning rate attenuation is set. The initial learning rate is set to 0.001, and the learning rate is reduced once every 6 rounds. The multiplier for each reduction in the learning rate is 0.6. Through the above-mentioned iterative training, model prediction, loss calculation, parameter update, and adaptive adjustment of the learning rate, the final intelligent correction model is obtained by multiple tuning. Finally, the root mean square error (RMSE) and mean absolute error (MAE) are used to evaluate the correction effect. The formula is:

[0072]

[0073] where yi is the observed value, is the predicted value and n is the number of samples.

[0074] For example, the training process is first divided into six epochs, with an initial learning rate of 0.001. In each training epoch, temperature field forecast data is input into the model to generate predictions. By comparing these predictions with the observed temperatures, the loss is calculated. The backpropagation algorithm then uses this gradient information to calculate the contribution of each parameter in the network to this loss. The Adam optimizer receives this gradient information and, based on the statistical properties of historical gradients, calculates the update direction and step size for each parameter. For example, the update amount for a convolutional layer's weight parameter depends not only on the current gradient value but also on the previously accumulated gradient momentum and the parameter's own update frequency. After completing one learning cycle (six epochs), the learning rate is reduced from 0.001 to 0.0006, at which point the model has shown initial results in correcting the temperature field. As training continues, the learning rate is further reduced, allowing the model to more accurately process detailed features in the temperature field. When the temperature prediction error on the validation set drops below a preset threshold, the final corrected temperature field model is obtained. This process fully demonstrates the adaptive characteristics of the deep learning model during training, and effectively corrects the temperature field forecast by dynamically adjusting the learning parameters.

[0075] S160: Apply the forecast field correction model to the target numerical model forecast data collected in real time, evaluate the forecast effect through the root mean square error and mean absolute error, and obtain the corrected forecast result after normalization and denormalization.

[0076] Optionally, statistical calculations are performed on the target numerical model forecast data, and normalization is performed based on the mean and standard deviation in the statistical information file to obtain normalized results of the forecast data; the normalized results of the forecast data are forward calculated through the forecast field correction model to obtain corrected forecast values; the corrected forecast values are denormalized to restore the values to the original scale to obtain preliminary corrected results; the sum of squares of the differences between the preliminary corrected results and the observed values is calculated, and the root mean square error is obtained after square root and sample size normalization; the absolute difference between the preliminary corrected results and the observed values is calculated, and the mean absolute error is obtained through sample size normalization; the root mean square error and the mean absolute error are used as evaluation indicators, and the preliminary corrected results are output as corrected forecast results.

[0077] Specifically, the target numerical model forecast data is standardized. This step uses the mean and standard deviation recorded in the statistical information file to perform Z-Score standardization. The purpose of standardization is to eliminate the dimensional differences between different meteorological elements, making the data distribution more reasonable and convenient for deep learning model processing. This process converts the original data into a standard normal distribution space with a mean of 0 and a standard deviation of 1. The standardized forecast data is input into the forecast field correction model for forward calculation. The forward calculation process includes the transmission of data between the layers of the deep learning network, transforming the input data through the trained network parameters, and finally obtaining the corrected forecast value. This process makes full use of the feature extraction and conversion capabilities learned by the model during the training phase to make targeted corrections to the input forecast data.

[0078] Since the revised forecast values are still in the standardized space, denormalization is required to restore the values to their original scale. Denormalization uses the same statistics (mean and standard deviation) as standardization, but performs the inverse transformation, mapping the values in the standardized space back to their original scale. The resulting preliminary revised results have the same dimensions and scale as the original observations, facilitating direct comparison and evaluation. To quantitatively evaluate the effectiveness of the correction, two metrics are calculated: the root mean square error (RMS) and the mean absolute error (MAE). The RMS error calculation process first calculates the sum of the squares of the differences between the preliminary revised results and the observed values, then takes the square root, and finally divides by the sample size for normalization. The RMS error is more sensitive to larger errors and can reflect unusual deviations in the forecast. Furthermore, the absolute difference between the preliminary revised results and the observed values is calculated and divided by the sample size to obtain the MAE, which reflects the average level of forecast error.

[0079] Finally, the calculated root mean square error (RMSE) and mean absolute error (MAE) are used as evaluation metrics, and the preliminary correction results are determined as the final corrected forecast results. These two evaluation metrics reflect the correction effect from different perspectives. RMSE emphasizes the penalty for outliers, while MAE provides a holistic error estimate.

[0080] For example, when receiving new temperature forecast data, the temperature field statistical features (mean and standard deviation) recorded in the statistical information file are read and the forecast temperature field is normalized. The normalized data is then fed into a trained temperature field correction model. The model uses the learned feature transformation relationship to generate a corrected temperature field forecast. This forecast result undergoes denormalization to restore it to the actual temperature value. The restored temperature forecast value is then compared with the measured temperature at the observation station, and the forecast error for each observation point is calculated. The root mean square error (RMS) is calculated by taking the sum of the squares of these errors, taking the square root, and normalizing for sample size to obtain the root mean square error (RMSE), which reflects the forecast accuracy. Simultaneously, the absolute difference between the predicted and observed temperatures is calculated and averaged to obtain the mean absolute error (MAE). These two error metrics together reflect the effectiveness of the temperature forecast correction and provide an objective basis for forecast quality assessment. In this example, each step of the data processing strictly adheres to the normalization-prediction-denormalization-evaluation process, ensuring the reliability and interpretability of the forecast correction results.

[0081] The above is an introduction to the method embodiment. The following is a further explanation of the solution disclosed in the present disclosure through an apparatus embodiment.

[0082] Figure 3 FIG. 2 shows a block diagram of a multi-factor deep learning correction device 200 for numerical model forecasting according to an embodiment of the present disclosure. Figure 3 As shown, the apparatus 200 includes:

[0083] The screening module 210 is used to perform spatiotemporal matching processing and outlier screening on the historical numerical model forecast data and the historical meteorological station observation data, obtain a sample pair time series file path sequence, and generate a statistical information file based on the statistics of meteorological element characteristics and terrain characteristics;

[0084] A reading module 220 is configured to read data from a time series file path sequence according to the sample, convert the meteorological element features into tensors through netCDF format conversion, and perform Z-Score normalization processing using the statistical information file to obtain a normalized tensor of meteorological features;

[0085] Extraction module 230, configured to input the normalized tensor of meteorological features into the encoder of DeepLab V3+ neural network, perform feature extraction through ResNet series network, perform multi-scale feature fusion through the dilated pyramid pooling module, and process the decoder to obtain a prediction result consistent with the input feature dimension;

[0086] A marking module 240 is used to mark observation points by constructing a mask tensor based on the location distribution of the historical meteorological station observation data, and to construct a loss function for the observation point locations to obtain a model optimization target;

[0087] An updating module 250 is configured to iteratively update the model parameters using an Adam optimizer according to the model optimization objective, and optimize the training process using a dynamic learning rate adjustment strategy to obtain a forecast field correction model;

[0088] Evaluation module 260 is used to apply the forecast field correction model to the target numerical model forecast data collected in real time, evaluate the forecast effect through root mean square error and mean absolute error, and obtain the corrected forecast result after normalization and denormalization.

[0089] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the described module can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here.

[0090] The above specific embodiments do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art will appreciate that various modifications, combinations, sub-combinations, and substitutions may be made based on design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure shall be included within the scope of protection of this disclosure.

Claims

1. A multi-factor deep learning correction method for numerical model forecasts, characterized by: include: Perform spatiotemporal matching and outlier screening on historical numerical model forecast data and historical meteorological station observation data to obtain sample pair time series file path sequences, and generate statistical information files based on the statistics of meteorological element characteristics and terrain characteristics; Reading data from a time series file path sequence according to the sample, converting meteorological element features into tensors through netCDF format conversion, and performing Z-Score standardization processing using the statistical information file to obtain a meteorological feature standardized tensor; The normalized tensor of meteorological features is input into the encoder of DeepLab V3+ neural network, feature extraction is performed through the ResNet series network, multi-scale feature fusion is performed through the dilated pyramid pooling module, and the decoder processes it to obtain a prediction result consistent with the input feature dimension; According to the location distribution of the historical meteorological station observation data, the observation points are marked by constructing a mask tensor, and a loss function is constructed for the observation point positions to obtain the model optimization target; According to the model optimization objectives, the model parameters are iteratively updated through the Adam optimizer, and the training process is optimized using a dynamic learning rate adjustment strategy to obtain a forecast field correction model; The forecast field correction model is applied to the target numerical model forecast data collected in real time, and the forecast effect is evaluated by the root mean square error and mean absolute error. The corrected forecast result is obtained after normalization and denormalization.

2. The multi-factor deep learning correction method for numerical model forecast according to claim 1 is characterized in that Performing spatiotemporal matching processing and outlier screening on historical numerical model forecast data and historical meteorological station observation data to obtain a sample pair time series file path sequence, and generating a statistical information file based on the statistics of meteorological element characteristics and terrain characteristics, including: Pair historical numerical model forecast data and historical meteorological station observation data according to time and space location, and remove observation station data outside the forecast area through data verification; Perform outlier detection on the paired data, eliminate samples containing abnormal data and serious missing data, and obtain valid sample data; Establishing an index in chronological order based on the valid sample data to generate a sample pair time series file path sequence; Extract meteorological element features from historical numerical model forecast data, extract terrain features from terrain data, and obtain feature data sets; Performing statistical analysis on the feature data set to calculate the mean, variance, maximum and minimum values of each feature; The mean, variance, maximum and minimum values of the features are integrated to generate a statistical information file.

3. The multi-factor deep learning correction method for numerical model forecast according to claim 1, characterized in that: The method of reading data from the time series file path sequence according to the sample, converting the meteorological element features into tensors through netCDF format conversion, and performing Z-Score standardization processing using the statistical information file to obtain a meteorological feature standardized tensor includes: Extracting numerical model forecast data, terrain data and meteorological station observation data from the time series file path sequence according to the sample; Read meteorological element characteristic data from the netCDF data format and convert it into the first tensor data; The terrain data is clipped and interpolated according to longitude and latitude to generate the second tensor data; Read the target element data from the meteorological station observation data, construct a third tensor data with the same dimension as the numerical model forecast data, and fill the missing locations with preset default values; Performing Z-Score normalization calculation on the first tensor data, the second tensor data, and the third tensor data according to the mean and standard deviation in the statistical information file; The standardized first tensor data, the second tensor data, and the third tensor data are combined to obtain a meteorological characteristic standardized tensor.

4. The multi-factor deep learning correction method for numerical model forecast according to claim 1, characterized in that: The meteorological feature normalized tensor is input into the encoder of the DeepLab V3+ neural network, feature extraction is performed through the ResNet series network, multi-scale feature fusion is performed through the hollow pyramid pooling module, and the decoder processes it to obtain a prediction result consistent with the input feature dimension, including: Input the normalized tensor of meteorological features into the DeepLab V3+ neural network to generate the intermediate low-level feature map and the final convolution feature map; Processing the last layer of convolutional feature maps through four dilated convolution blocks with different dilation rates, each dilated convolution block is equipped with a batch normalization layer and an activation layer to obtain a first set of feature maps; Performing global average pooling on the last layer of convolutional feature maps and processing them through convolution blocks, equipped with batch normalization layers and activation layers to obtain a second set of feature maps; Concatenate the first set of feature maps and the second set of feature maps, process them through a convolution block and equip them with a batch normalization layer, an activation layer, and a dropout layer to obtain a fused feature map; Performing 1×1 convolution dimensionality reduction processing on the low-level feature map of the intermediate layer, upsampling the fused feature map to make it the same size as the low-level feature map after dimensionality reduction, and concatenating the two and processing them through a 3×3 convolution block; The feature map obtained by splicing and then processed by the 3×3 convolution block is upsampled by linear interpolation to obtain a prediction result consistent with the input feature dimension.

5. The multi-factor deep learning correction method for numerical model forecast according to claim 1, characterized in that: According to the location distribution of the historical meteorological station observation data, the observation points are marked by constructing a mask tensor, and the loss function is constructed for the observation point positions to obtain the model optimization target, including: Based on the grid dimension data of the location distribution of historical meteorological station observation data, an initial mask tensor is constructed, and all grid points are filled with preset default values; Matching the longitude and latitude information of the observation site in the historical meteorological site observation data with the longitude and latitude of the numerical model forecast grid data to obtain a matching result; Based on the matching result, in the initial mask tensor, the grid points with observed data are marked as 1, and the grid points without observed data are marked as 0, to obtain a mask tensor; weighting the squared difference between the predicted value and the historical meteorological station observation data by the mask tensor to generate a masked weighted error; Normalizing the mask weighted error by the number of grid points with a median value of 1 in the mask tensor to obtain an average error value; The average error value is used as the calculation result of the loss function to obtain the model optimization target.

6. The multi-factor deep learning correction method for numerical model forecast according to claim 1, characterized in that: According to the model optimization objective, the model parameters are iteratively updated by the Adam optimizer, and the training process is optimized by adopting a dynamic learning rate adjustment strategy to obtain a forecast field correction model, including: The training rounds are divided into six training rounds as one learning cycle, the initial learning rate is set to 0.001, and a training round sequence is generated; The model optimization target is input into the Adam optimizer, and the gradient value of each parameter is calculated through the back propagation algorithm to generate a parameter gradient sequence; Based on the parameter gradient sequence, momentum correction and adaptive moment estimation are performed on each parameter to obtain a corrected gradient value; The model parameters are updated by multiplying the corrected gradient value by the current learning rate to obtain the optimization result of this round; At the end of each learning cycle, the current learning rate is multiplied by a coefficient of 0.6 to perform a decay adjustment on the learning rate of the next learning cycle; The loss value is calculated and convergence is judged for the optimization results of this round. When the loss value is lower than a preset threshold, a forecast field correction model is constructed according to the current parameters.

7. The multi-factor deep learning correction method for numerical model forecast according to claim 1, characterized in that: The forecast field correction model is applied to the target numerical model forecast data collected in real time, the forecast effect is evaluated by the root mean square error and the mean absolute error, and the corrected forecast result is obtained after normalization and denormalization, including: Calculating statistics on the target numerical model forecast data, and performing normalization processing based on the mean and standard deviation in the statistical information file to obtain a normalized result of the forecast data; Performing forward calculation on the normalized result of the forecast data through the forecast field correction model to obtain a corrected forecast value; Denormalizing the corrected predicted values to restore them to their original scales, thereby obtaining preliminary corrected results. Calculate the sum of squares of the differences between the preliminary correction results and the observed values, and obtain the root mean square error after square root extraction and sample size normalization; Calculate the absolute difference between the preliminary correction result and the observed value, and obtain the mean absolute error by sample size normalization; The root mean square error and the mean absolute error are used as evaluation indicators, and the preliminary correction result is output as a corrected forecast result.

8. A multi-factor deep learning correction device for numerical model forecasts, used to implement the multi-factor deep learning correction method for numerical model forecasts according to any one of claims 1 to 7, characterized in that: The multi-factor deep learning correction device for numerical model forecasts includes: The screening module is used to perform spatiotemporal matching processing and outlier screening on historical numerical model forecast data and historical meteorological station observation data, obtain the sample pair time series file path sequence, and generate statistical information files based on the statistics of meteorological element characteristics and terrain characteristics; A reading module is used to read data from a time series file path sequence according to the sample, perform tensor conversion on meteorological element features through netCDF format conversion, perform Z-Score normalization processing using the statistical information file, and obtain a normalized tensor of meteorological features; An extraction module is used to input the normalized tensor of meteorological features into the encoder of the DeepLab V3+ neural network, perform feature extraction through the ResNet series network, perform multi-scale feature fusion through the dilated pyramid pooling module, and process it by the decoder to obtain a prediction result consistent with the input feature dimension; A marking module is used to mark observation points by constructing a mask tensor based on the location distribution of the historical meteorological station observation data, and to construct a loss function for the observation point location to obtain a model optimization target; An updating module is used to iteratively update the model parameters through the Adam optimizer according to the model optimization objective, and to optimize the training process using a dynamic learning rate adjustment strategy to obtain a forecast field correction model; The evaluation module is used to apply the forecast field correction model to the target numerical model forecast data collected in real time, evaluate the forecast effect through the root mean square error and mean absolute error, and obtain the corrected forecast result after standardization and de-standardization.

Images