Method and system for seamless weather forecast product generation based on multi-time scale element fusion

Abstract

This invention discloses a method and system for generating seamless weather forecast products based on the fusion of multiple time-scale elements, relating to the field of weather forecasting technology. The method first acquires multi-source observational data and numerical model data, and generates short-, medium-, and long-term standardized sequences through time interpolation, matching, and sliding window resampling. Then, a multi-branch convolutional neural network is used to extract corresponding feature maps, calculate physical quantity field parameters, and encode them as physical context vectors. A weight map is generated through a dynamic gating network, and the feature maps are adaptively fused. The data is input into the forecasting network and trained using a joint loss function with physical constraints to generate a seamless forecast product. Finally, the weight strategy is dynamically optimized through continuity verification and error evaluation. The system contains corresponding functional modules, and the executable program is stored in the storage medium.

Description

Technical Field

[0001] This invention relates to the field of meteorological forecasting technology, and more specifically, to a method and system for generating seamless meteorological forecast products based on the fusion of multiple time-scale elements. Background Technology

[0002] In the field of meteorological science, accurate and seamless weather forecasting has always been the core pursuit. The requirements for the refinement and continuity of weather forecasting have reached an unprecedented level. Multi-source meteorological observation data and numerical model data constitute the cornerstone of modern weather forecasting. However, there are significant differences in the time scale between the two. Radar echoes capture instantaneous changes in convective systems at minute-level resolution, satellite cloud images are updated hourly to present large-scale cloud dynamics, while ground observation station data mostly record local meteorological elements at hourly to several-hour intervals. Although numerical models can simulate weather conditions at different future time periods, from short-term high-resolution to long-term low-resolution products, they are difficult to match with observation data in terms of time scale, forming an obstacle to information transmission.

[0003] Traditional meteorological forecasting methods often use simple time interpolation to unify the time scale when processing this data, ignoring the unique physical implications of data at different time scales. Short-term high-frequency data focuses on the details of convective system evolution, medium-term mid-frequency data reflects the movement and development trajectory of weather systems, and long-term low-frequency data outlines the macroscopic background of atmospheric circulation. This homogenization process leads to the loss of a large number of key physical signals. Furthermore, existing forecasting models mostly rely on data-driven feature fusion strategies. In areas with frequent severe convective weather, the dynamic correlation between physical quantities such as vertically integrated liquid water content and divergence field and observational data is not effectively utilized. The models cannot accurately and adaptively adjust feature weights, resulting in unsatisfactory forecast accuracy for both severe convective and stable weather areas. In addition, abrupt changes in forecasts occur frequently at the junctions of different time periods due to abrupt time scale transitions, and the error assessment system is disconnected from the feature fusion strategy, making it difficult to continuously optimize forecast results. Under these complex predicaments, developing innovative fusion methods that take into account the characteristics of multi-time-scale data and physical consistency has become an essential path to break through the bottlenecks of meteorological forecasting and move towards seamless and accurate forecasting.

[0004] Therefore, existing technologies suffer from problems such as difficulty in unifying data time scales, lack of physical constraints, poor time period connections, and difficulty in optimizing errors. Summary of the Invention

[0005] In order to overcome the problems of existing technologies, such as difficulty in unifying data time scales, lack of physical constraints, poor time period connection, and difficulty in optimizing errors, this invention discloses a method and system for generating seamless weather forecast products based on the fusion of multiple time scale elements, which can effectively solve the above-mentioned technical problems.

[0006] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0007] A method for generating seamless weather forecast products based on multi-timescale element fusion includes the following steps:

[0008] Acquire multi-source meteorological observation data and numerical model data. The multi-source meteorological observation data includes radar echo data, satellite cloud image data and ground observation station data at different time resolutions. The numerical model data includes atmospheric physical field data at different forecast lead times.

[0009] The multi-source meteorological observation data and numerical model data are subjected to time-scale unification processing to obtain standardized time series data, which includes short-term high-frequency observation sequences, medium-term mid-frequency simulation sequences, and long-term low-frequency trend sequences.

[0010] The basic features are extracted from the standardized time series data to obtain a multi-scale feature map, which includes a short-term feature map, a medium-term feature map, and a long-term feature map.

[0011] Calculate the physical field parameters corresponding to the standardized time series data, including vertical integral liquid water content, divergence field, and pseudo-equivalent potential temperature; input the physical field parameters into a physical sensing coding network to generate a physical context vector;

[0012] The physical context vector is input into a dynamic gating network to generate a dynamic weight map corresponding to the spatial location of the multi-scale feature map. The dynamic weight map reflects the credibility of features at different time scales at each spatial grid point.

[0013] The multi-scale feature map is adaptively fused using the dynamic weight map to obtain a fused feature map; wherein, the weight ratio of short-term high-frequency observation sequences is increased in strong convection regions, and the weight ratio of long-term low-frequency trend sequences is increased in stable weather regions.

[0014] The fused feature map is input into the prediction network for prediction processing. During model training, a joint loss function is used for parameter optimization. The joint loss function includes a mean squared error loss term and a physical consistency constraint term.

[0015] Seamless weather forecast products are generated based on the forecast processing results. These seamless weather forecast products include continuous time-series forecast results of 0-12 hour short-term forecasts, 12-72 hour medium-term forecasts, and 72-168 hour long-term forecasts.

[0016] Preferably, the step of performing time-scale unification processing on the multi-source meteorological observation data and numerical model data to obtain standardized time series data includes:

[0017] A time interpolation algorithm is used to unify multi-source meteorological observation data with different time resolutions into an observation sequence with a preset time interval;

[0018] The numerical model data is time-matched to extract the simulation data that is time-aligned with the observation sequence, forming a simulation sequence.

[0019] The observed and simulated sequences are resampled using the sliding window method to generate short-term high-frequency sequences, medium-term mid-frequency sequences, and long-term low-frequency sequences, forming standardized time series data.

[0020] Preferably, the step of extracting the basic features from the standardized time series data to obtain a multi-scale feature map includes:

[0021] Construct a multi-branch convolutional neural network, wherein the multi-branch convolutional neural network includes a short-term feature extraction branch, a medium-term feature extraction branch, and a long-term feature extraction branch;

[0022] Short-term high-frequency observation sequences are input into the short-term feature extraction branch, and refined evolution features of the convective system are extracted through multi-layer convolution and pooling operations to obtain a short-term feature map;

[0023] The mid-term intermediate frequency simulation sequence is input into the mid-term feature extraction branch to extract the movement and development features of the weather system, thus obtaining the mid-term feature map;

[0024] The long-term low-frequency trend sequence is input into the long-term feature extraction branch to extract atmospheric circulation background features and obtain a long-term feature map.

[0025] Preferably, the step of inputting the physical context vector into a dynamic gating network to generate a dynamic weight map corresponding to the spatial location of the multi-scale feature map includes:

[0026] The dynamic gating network consists of a fully connected layer and a sigmoid activation function, receiving a physical context vector and outputting a weight matrix with the same size as the feature map.

[0027] The weight matrix includes a short-term weight map, a medium-term weight map, and a long-term weight map, and the sum of the weights in spatial locations is 1.

[0028] In strong convection regions where the vertical integral liquid water content is higher than a preset threshold, the weight value of the corresponding position in the short-term weighted map is ≥0.7; in stable regions where the absolute value of divergence is lower than a preset threshold, the weight value of the corresponding position in the long-term weighted map is ≥0.6.

[0029] Preferably, the step of adaptively fusing the multi-scale feature map using the dynamic weight map to obtain a fused feature map includes:

[0030] The formula for calculating the fused feature map is: ;

[0031] in, This is a short-term weighted chart. For short-term feature maps, This is a mid-term weighted chart. For mid-term feature maps, For long-term weighted charts, For long-term feature maps;

[0032] Feature enhancement processing is performed on the fused feature map, and key physical signals are introduced into the original multi-scale feature map through residual connection to enhance the feature representation of the strong convection system.

[0033] Preferably, the formula for calculating the joint loss function is: ;

[0034] in, This is the mean squared error loss term, used to measure the deviation between the forecast result and the observed value; This is a physical consistency constraint term, used to constrain the statistical relationship between the forecast field and the physical quantity field; and These are the weighting coefficients, and + = 1;

[0035] The physical consistency constraints include: the correlation constraint between the precipitation field and the vertical velocity field, the total precipitation conservation constraint, and the thermodynamic relationship constraint between the temperature field and the pressure field.

[0036] Preferably, after generating a seamless weather forecast product, the method further includes:

[0037] A time continuity verification algorithm is used to smooth the transitions between different time periods in the forecast product, eliminating abrupt changes in the forecast.

[0038] Error assessment of forecast results at different time scales is performed using a sliding validation window, calculating the root mean square error of short-term forecasts, the mean absolute error of medium-term forecasts, and the trend accuracy of long-term forecasts.

[0039] The weight generation strategy of the dynamic gating network is dynamically adjusted based on the error evaluation results to optimize the fusion ratio of features at different time scales.

[0040] Preferably, the seamless weather forecast product generation system based on multi-timescale element fusion includes:

[0041] The data acquisition module is used to acquire multi-source meteorological observation data and numerical model data. The multi-source meteorological observation data includes radar echo data, satellite cloud image data and ground observation station data at different time resolutions.

[0042] The time scale unification module is used to perform time scale unification processing on the multi-source meteorological observation data and numerical model data to obtain standardized time series data containing short-term high-frequency observation sequences, medium-term mid-frequency simulation sequences and long-term low-frequency trend sequences.

[0043] A multi-scale feature extraction module is used to extract basic features from the standardized time series data to obtain short-term feature maps, medium-term feature maps, and long-term feature maps.

[0044] The physical sensing dynamic gating module includes a physical quantity calculation unit, a physical encoding unit, and a dynamic weight generation unit; the physical quantity calculation unit is used to calculate physical quantity field parameters; the physical encoding unit is used to encode the physical quantity field parameters into a physical context vector; and the dynamic weight generation unit is used to generate a dynamic weight map corresponding to the spatial location of the feature map.

[0045] The feature fusion module is used to perform adaptive fusion processing on the multi-scale feature map through the dynamic weight map to obtain the fused feature map.

[0046] The forecast generation module is used to input the fused feature map into the forecast network for prediction processing. During the training process, a joint loss function including a physical consistency constraint term is used to generate seamless weather forecast products.

[0047] The product optimization module is used to perform time continuity verification and error assessment on forecast products, and dynamically optimize the fusion strategy.

[0048] Preferably, the dynamic weight generation unit of the physical perception dynamic gating module adopts a gating network structure that includes an attention mechanism, which can dynamically adjust the weight allocation of features at different time scales according to real-time physical field parameters, and automatically enhance the weight ratio of short-term high-frequency observation data in strong convection regions.

[0049] Preferably, a computer-readable storage medium stores a computer program, the computer program including program instructions, which, when executed by a processor, cause the processor to perform the seamless weather forecast product generation method based on multi-timescale element fusion as described above.

[0050] Compared with existing technologies, the beneficial effects of this invention are as follows: Addressing the problem of inconsistent data time scales, this invention first uses a time interpolation algorithm to unify multi-source observation data such as radar echoes and satellite cloud images into observation sequences with preset time intervals. Then, it performs time matching on numerical model data with different forecast lead times to align the observation sequences. Finally, it uses a sliding window method to resample and generate standardized time series for short-term high-frequency (≤1 hour), medium-term mid-frequency (3-6 hours), and long-term low-frequency (≥12 hours), eliminating the time dimension differences between multi-source data and providing a unified and orderly data foundation for subsequent feature extraction, avoiding feature information misalignment or loss due to chaotic data scales. Regarding the lack of physical constraints… On the one hand, physical parameters such as vertical integral liquid water content, divergence field, and pseudo-equivalent potential temperature are calculated and transformed into physical context vectors through a physical sensing encoding network. This vectors drive a dynamic gating network composed of fully connected layers and a Sigmoid activation function to generate a spatially adaptive weight map (short-term weights ≥ 0.7 in strong convection regions and long-term weights ≥ 0.6 in stable regions), ensuring that feature fusion strictly conforms to meteorological physical laws. On the other hand, a joint loss function with constraints including the correlation between precipitation field and vertical velocity field, conservation of total precipitation, and thermodynamic relationship between temperature field and pressure field is introduced into the model training. This forces the forecast field to conform to the physical laws of atmospheric motion. From feature fusion to model optimization, a dual embedding of physical logic is used to solve the problem of purely data-driven approaches. To address the issue of models easily deviating from meteorological realities, and to resolve the problem of poor time-scale continuity, a time continuity verification algorithm is employed to smooth the transitions between different time periods after generating short-term (0-12 hours), medium-term (12-72 hours), and long-term (72-168 hours) forecasts. This directly eliminates abrupt forecast changes, ensuring the continuity and smoothness of time-series forecasts and preventing time-scale gaps from affecting the practicality of the forecast results. Furthermore, to address the difficulty in error optimization, a sliding verification window is used to calculate the root mean square error of the short-term forecast, the mean absolute error of the medium-term forecast, and the trend accuracy of the long-term forecast. Based on these error assessment results, the weight generation strategy of the dynamic gating network is dynamically adjusted to optimize the fusion of features at different time scales. By constructing a closed-loop iterative mechanism of error assessment, strategy adjustment, and fusion optimization, this technology addresses the shortcomings of traditional methods where errors cannot be fed back into the model and accuracy cannot be continuously improved. It is the synergistic effect of these four problem-solving paths that enables the seamless weather forecast products generated by this technology to not only have data support at a unified time scale and forecast accuracy guaranteed by physical laws, but also have continuous and uninterrupted temporal performance. Furthermore, it can continuously optimize accuracy through closed-loop iteration, ultimately providing more reliable decision-making basis for scenarios such as meteorological disaster prevention and mitigation, agricultural production scheduling, transportation support, and public safety emergency response. This significantly enhances the practical value and application breadth of weather forecast services and effectively reduces the socio-economic losses caused by extreme weather. Attached Figure Description

[0051] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely exemplary. For those skilled in the art, other embodiments can be derived from the provided drawings without creative effort.

[0052] Figure 1 This is a diagram illustrating the steps of the method of the present invention;

[0053] Figure 2 This is a system structure diagram of the present invention; Detailed Implementation

[0054] The accompanying drawings are for illustrative purposes only and should not be construed as limiting the scope of this patent.

[0055] To better illustrate this embodiment, some parts in the accompanying drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions.

[0056] It will be understood by those skilled in the art that certain well-known structures and their descriptions may be omitted in the accompanying drawings.

[0057] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0058] Example 1

[0059] Please see Figure 1 A method for generating seamless weather forecast products based on the fusion of multiple time-scale elements includes the following steps:

[0060] Acquire multi-source meteorological observation data and numerical model data. The multi-source meteorological observation data includes radar echo data, satellite cloud image data and ground observation station data at different time resolutions. The numerical model data includes atmospheric physical field data at different forecast lead times.

[0061] The multi-source meteorological observation data and numerical model data are subjected to time-scale unification processing to obtain standardized time series data, which includes short-term high-frequency observation sequences, medium-term mid-frequency simulation sequences, and long-term low-frequency trend sequences.

[0062] The basic features are extracted from the standardized time series data to obtain a multi-scale feature map, which includes a short-term feature map, a medium-term feature map, and a long-term feature map.

[0063] Calculate the physical field parameters corresponding to the standardized time series data, including vertical integral liquid water content, divergence field, and pseudo-equivalent potential temperature; input the physical field parameters into a physical sensing coding network to generate a physical context vector;

[0064] The physical context vector is input into a dynamic gating network to generate a dynamic weight map corresponding to the spatial location of the multi-scale feature map. The dynamic weight map reflects the credibility of features at different time scales at each spatial grid point.

[0065] The multi-scale feature map is adaptively fused using the dynamic weight map to obtain a fused feature map; wherein, the weight ratio of short-term high-frequency observation sequences is increased in strong convection regions, and the weight ratio of long-term low-frequency trend sequences is increased in stable weather regions.

[0066] The fused feature map is input into the prediction network for prediction processing. During model training, a joint loss function is used for parameter optimization. The joint loss function includes a mean squared error loss term and a physical consistency constraint term.

[0067] Seamless weather forecast products are generated based on the forecast processing results. These seamless weather forecast products include continuous time-series forecast results of 0-12 hour short-term forecasts, 12-72 hour medium-term forecasts, and 72-168 hour long-term forecasts.

[0068] The process of unifying the time scale of the multi-source meteorological observation data and numerical model data to obtain standardized time series data includes:

[0069] A time interpolation algorithm is used to unify multi-source meteorological observation data with different time resolutions into an observation sequence with a preset time interval;

[0070] The numerical model data is time-matched to extract the simulation data that is time-aligned with the observation sequence, forming a simulation sequence.

[0071] The observed and simulated sequences were resampled using the sliding window method to generate short-term high-frequency sequences (time interval ≤ 1 hour), medium-term mid-frequency sequences (time interval = 3-6 hours), and long-term low-frequency sequences (time interval ≥ 12 hours), forming standardized time series data.

[0072] The extraction of basic features from the standardized time series data to obtain a multi-scale feature map includes:

[0073] Construct a multi-branch convolutional neural network, wherein the multi-branch convolutional neural network includes a short-term feature extraction branch, a medium-term feature extraction branch, and a long-term feature extraction branch;

[0074] Short-term high-frequency observation sequences are input into the short-term feature extraction branch, and refined evolution features of the convective system are extracted through multi-layer convolution and pooling operations to obtain a short-term feature map;

[0075] The mid-term intermediate frequency simulation sequence is input into the mid-term feature extraction branch to extract the movement and development features of the weather system, thus obtaining the mid-term feature map;

[0076] The long-term low-frequency trend sequence is input into the long-term feature extraction branch to extract atmospheric circulation background features and obtain a long-term feature map.

[0077] The step of inputting the physical context vector into a dynamic gating network to generate a dynamic weight map corresponding to the spatial location of the multi-scale feature map includes:

[0078] The dynamic gating network consists of a fully connected layer and a sigmoid activation function, receiving a physical context vector and outputting a weight matrix with the same size as the feature map.

[0079] The weight matrix includes a short-term weight map, a medium-term weight map, and a long-term weight map, and the sum of the weights in spatial locations is 1.

[0080] In strong convection regions where the vertical integral liquid water content is higher than a preset threshold, the weight value of the corresponding position in the short-term weighted map is ≥0.7; in stable regions where the absolute value of divergence is lower than a preset threshold, the weight value of the corresponding position in the long-term weighted map is ≥0.6.

[0081] The step of adaptively fusing the multi-scale feature map using the dynamic weight map to obtain a fused feature map includes:

[0082] The formula for calculating the fused feature map is: ;

[0083] in, This is a short-term weighted chart. For short-term feature maps, This is a mid-term weighted chart. For mid-term feature maps, For long-term weighted charts, For long-term feature maps;

[0084] Feature enhancement processing is performed on the fused feature map, and key physical signals are introduced into the original multi-scale feature map through residual connection to enhance the feature representation of the strong convection system.

[0085] The formula for calculating the joint loss function is as follows: ;

[0086] in, This is the mean squared error loss term, used to measure the deviation between the forecast result and the observed value; This is a physical consistency constraint term, used to constrain the statistical relationship between the forecast field and the physical quantity field; and These are the weighting coefficients, and + = 1;

[0087] The physical consistency constraints include: the correlation constraint between the precipitation field and the vertical velocity field, the total precipitation conservation constraint, and the thermodynamic relationship constraint between the temperature field and the pressure field.

[0088] After generating a seamless weather forecast product, the method further includes:

[0089] A time continuity verification algorithm is used to smooth the transitions between different time periods in the forecast product, eliminating abrupt changes in the forecast.

[0090] Error assessment of forecast results at different time scales is performed using a sliding validation window, calculating the root mean square error of short-term forecasts, the mean absolute error of medium-term forecasts, and the trend accuracy of long-term forecasts.

[0091] The weight generation strategy of the dynamic gating network is dynamically adjusted based on the error evaluation results to optimize the fusion ratio of features at different time scales.

[0092] In practice, multi-source meteorological observation data collection

[0093] Radar echo data: Radar echo data is acquired from multiple radar sites distributed across different geographical areas to ensure coverage of the entire forecast area. The temporal resolution of radar data varies; for example, some radars provide data every 5-10 minutes to capture rapidly changing weather systems, such as severe convective weather, while radars in more remote areas may provide data every 30 minutes. Radar echo data contains information such as reflectivity factors to identify precipitation areas and intensity.

[0094] Satellite cloud image data: Cloud image data is acquired from polar-orbiting satellites and geostationary satellites. Polar-orbiting satellites can provide high-resolution global coverage images, and their data update frequency is generally once every 3-6 hours. Geostationary satellites can provide continuous regional images with higher temporal resolution, and are generally updated every 15-30 minutes. Satellite cloud image data includes images in multiple bands such as visible light and infrared, and is used to analyze the type, distribution and evolution of clouds.

[0095] Ground observation station data: Data is collected from ground meteorological observation stations, which are distributed in different terrain areas such as cities, rural areas and mountains. The time resolution of the observation station data is usually once per hour, including basic meteorological elements such as temperature, air pressure, humidity, wind speed and wind direction.

[0096] Numerical model data collection: Data is acquired from multiple numerical weather prediction models, including global and regional models. Global models provide large-scale weather forecasts with longer lead times, typically 7-10 days or even longer. Regional models focus on high-resolution forecasts for smaller areas with relatively shorter lead times, typically 3-5 days. Numerical model data includes atmospheric physical field data, such as temperature, humidity, and wind fields. The temporal resolution of these data varies depending on the model, generally every 3-6 hours.

[0097] For multi-source meteorological observation data with different time resolutions, linear interpolation or spline interpolation algorithms are used. Taking radar echo data as an example, for radar data with low time resolution, assuming its time interval is 30 minutes, we need to unify it to a 10-minute time interval. Using the linear interpolation algorithm, based on the known data points at the 30-minute interval, the data points at the middle 10 minutes and 20 minutes are calculated. Specifically, if the radar measures reflectivity factors of 30 dBZ and 35 dBZ at 0:00 and 0:30 respectively, then through linear interpolation, the reflectivity factor at 0:10 is approximately 31.67 dBZ, and the reflectivity factor at 0:20 is approximately 33.33 dBZ. In this way, radar data with different time resolutions are unified into the preset 10-minute time interval observation sequence.

[0098] For satellite cloud image data, if the data update frequency of polar-orbiting satellites is once every 3 hours, and we need to unify it to a time interval of 1 hour, we can use the spline interpolation algorithm. Spline interpolation can better consider the smoothness of the data and avoid large fluctuations in the interpolation results. Through spline interpolation, based on the cloud image data of polar-orbiting satellites at 0:00, 3:00, 6:00, etc., we can calculate the cloud image data at 1:00, 2:00, 4:00, 5:00, etc., thereby generating an observation sequence with a unified time interval.

[0099] Time matching processing is performed on the numerical model data to extract simulated data that is time-aligned with the observation sequence. For example, if the time interval of the observation sequence is 1 hour and the time interval of the numerical model data is 3 hours, and the time points of the observation sequence are 0:00, 1:00, 2:00, 3:00, etc., and the time points of the numerical model data are 0:00, 3:00, 6:00, etc., then for 1:00 and 2:00 in the observation sequence, the data for 0:00 and 3:00 need to be extracted from the numerical model data. The simulated data corresponding to 1:00 and 2:00 are calculated by interpolation algorithm (such as linear interpolation) to form a simulated sequence that is time-aligned with the observation sequence.

[0100] The observed and simulated sequences are resampled to generate sequences at different time scales. For short-term high-frequency sequences (time interval ≤ 1 hour), a sliding window of 1 hour is used. Taking the observed sequence as an example, assuming the time interval is 10 minutes, starting from 0:00, the sliding window sequentially covers 0:00-1:00, 1:00-2:00, etc. Within each window, the average, maximum, and other statistics are calculated to generate the short-term high-frequency observed sequence. For medium-term intermediate-frequency sequences (time interval = 3-6 hours), a sliding window of 3 or 6 hours is used. For example, for the observed sequence, a sliding window of 3 hours is used, starting from 0:00, sequentially covering 0:00-3:00, 3:00-6:00, etc. The statistics within each window are calculated to generate the medium-term intermediate-frequency observed sequence. For long-term low-frequency trend sequences (time interval ≥ 12 hours), a sliding window of 12 hours or longer is used. Similarly, the statistics within each window are calculated to generate the long-term low-frequency observed sequence. In this way, the observed and simulated sequences are resampled into short-term high-frequency sequences, medium-term intermediate-frequency sequences, and long-term low-frequency sequences, respectively, to form standardized time series data.

[0101] Construct a multi-branch convolutional neural network containing short-term feature extraction, mid-term feature extraction, and long-term feature extraction branches. Each branch consists of multiple convolutional and pooling layers. The short-term feature extraction branch is mainly used to extract refined evolution features of convective systems, and its convolutional kernel size is relatively small, such as 3×3 or 5×5, in order to capture local and rapidly changing convective features. The mid-term feature extraction branch is used to extract the movement and development features of weather systems, and its convolutional kernel size can be appropriately increased, such as 7×7, to better capture the evolution of weather systems over a larger area. The long-term feature extraction branch is used to extract atmospheric circulation background features, and its convolutional kernel size is larger, such as 11×11, to extract more macroscopic circulation features.

[0102] Short-term high-frequency observation sequences are input into the short-term feature extraction branch. Taking radar echo data as an example, the short-term high-frequency observation sequence contains radar reflectivity factor data every 10 minutes or hour. After multiple convolution operations, each convolutional layer extracts features at different levels. For example, in the first convolutional layer, simple edge features, such as the boundary of precipitation areas, may be extracted. In subsequent convolutional layers, more complex convective features, such as the shape and intensity changes of convective cells, will be extracted. At the same time, through pooling operations, the spatial dimension of features can be reduced, the amount of computation can be reduced, and important feature information can be retained. After multiple convolution and pooling operations, the short-term feature map is finally obtained, which can reflect the refined evolution features of the convective system in the short term.

[0103] The mid-term mid-frequency simulation sequence is input into the mid-term feature extraction branch. Taking the wind field data provided by the numerical model as an example, the mid-term mid-frequency simulation sequence contains wind field data every 3-6 hours. In the mid-term feature extraction branch, the features of the wind field, such as the trend of wind direction change and the distribution of wind speed, are extracted through convolution operations. These features reflect the movement and development characteristics of the weather system on the mid-term time scale. For example, the center position of cyclones or anticyclones and their movement direction and speed can be identified through convolutional layers. After multiple convolution and pooling operations, a mid-term feature map is obtained, which can describe the evolution of the weather system on the mid-term time scale.

[0104] The long-term low-frequency trend sequence is input into the long-term feature extraction branch. Taking the temperature and pressure field data provided by the numerical model as an example, the long-term low-frequency trend sequence contains temperature and pressure field data every 12 hours or longer. In the long-term feature extraction branch, atmospheric circulation background features are extracted through convolution operations, such as the meridional and zonal distribution characteristics of atmospheric circulation and the location of high and low pressure belts. These features reflect the trend of atmospheric circulation on a long time scale. After multi-layer convolution and pooling operations, a long-term feature map is obtained, which can reflect the impact of atmospheric circulation background on weather.

[0105] To calculate the physical field parameters corresponding to standardized time series data, taking vertical integral liquid water content as an example, calculations can be performed using radar echo data and data such as temperature and humidity. First, based on the relationship between radar reflectivity factor and precipitation rate, the radar echo data is converted into precipitation rate. Then, combined with temperature and humidity data, the vertical integral liquid water content is calculated using physical formulas. For the divergence field, calculations can be performed using wind field data provided by numerical models. Based on the definition of divergence, the divergence value of each grid point is calculated using the finite difference method. For the pseudo-equivalent potential temperature, calculations can be performed using thermodynamic formulas based on data such as temperature, humidity, and air pressure.

[0106] The calculated physical field parameters are input into the physical sensing coding network, a deep neural network that encodes the physical field parameters into a physical context vector. The network structure can include multiple fully connected layers and activation functions. For example, the input physical field parameters first pass through a fully connected layer, mapping the input data to a high-dimensional space. Then, nonlinearity is introduced through activation functions (such as the ReLU function), enabling the network to learn complex physical feature relationships. After processing through multiple fully connected layers and activation functions, a physical context vector is finally generated. This vector contains comprehensive information about the physical field parameters and can provide physical background knowledge for subsequent dynamic weight generation.

[0107] The dynamic gating network consists of fully connected layers and a sigmoid activation function. The physical context vector is used as input. First, the features are transformed through the fully connected layers. The output of the fully connected layers is the original value of a weight matrix. Then, the sigmoid activation function maps these original values ​​to the (0, 1) interval to generate a weight matrix with the same size as the feature map. The weight matrix contains short-term weight map, medium-term weight map and long-term weight map. The sum of the weights of these three in spatial location is 1. That is, at each spatial grid point, the sum of the short-term weight value, medium-term weight value and long-term weight value is 1.

[0108] In strong convection regions where the vertical integral liquid water content exceeds a preset threshold, the weight value at the corresponding location in the short-term weighted map is ≥0.7. For example, assuming the preset threshold is 30 kg / m³. 2 When the vertical integral liquid water content at a certain spatial grid point is 40 kg / m³ 2 In such cases, the weight of this grid point in the short-term weighted map will be automatically adjusted to 0.7 or higher. This is because in such strong convective regions, short-term high-frequency observation data better reflects rapid weather changes, thus increasing the weight of short-term data. In stable regions where the absolute divergence value is below a preset threshold, the weight of the corresponding position in the long-term weighted map will be ≥0.6. For example, assuming the preset threshold is 0.1 × 10⁻⁶... -5 / s, when the absolute value of the divergence at a certain spatial grid point is 0.05×10 -5 When the value is / s, the weight of this grid point in the long-term weighted graph will be adjusted to 0.6 or higher. In the stable region, the long-term low-frequency trend sequence can better reflect the stable state of the weather, so the proportion of long-term weight is increased.

[0109] Adaptive fusion of multi-scale feature maps is performed using a dynamic weight map. The formula for calculating the fused feature map is as follows:

[0110]

[0111] in This is a short-term weighted chart. For short-term feature maps, This is a mid-term weighted chart. For mid-term feature maps, For long-term weighted charts, For long-term feature maps, in areas with strong convection, short-term feature maps play a more important role in the fused feature map due to their higher weight values; in areas with stable weather, long-term feature maps contribute more. This adaptive fusion method can reasonably fuse feature maps of different time scales according to different weather conditions.

[0112] Feature enhancement processing is performed on the fused feature map by incorporating key physical signals from the original multi-scale feature map through residual concatenation. Specifically, the fused feature map is residually concatenated with the original short-term, medium-term, and long-term feature maps. For example, the residual between the fused feature map and the original short-term feature map can be calculated and then added back to the fused feature map. This strengthens the representation of strong convective systems, as their characteristics may be more prominent in the original short-term feature map. In this way, the fused feature map can better preserve key physical signals and improve forecast accuracy.

[0113] The fused feature map is input into the forecasting network for prediction processing. The forecasting network can be a deep convolutional neural network that can predict future weather conditions based on the fused feature map. During model training, a joint loss function is used for parameter optimization. The joint loss function includes a mean squared error loss term and a physical consistency constraint term. The mean squared error loss term is used to measure the deviation between the forecast result and the observed value. For example, for temperature forecast, the mean squared error loss term calculates the sum of squares of the difference between the forecast temperature and the actual observed temperature. The physical consistency constraint term is used to constrain the statistical relationship between the forecast field and the physical quantity field. It includes the correlation constraint between the precipitation field and the vertical velocity field, the total precipitation conservation constraint, and the thermodynamic relationship constraint between the temperature field and the pressure field. For example, the correlation constraint between the precipitation field and the vertical velocity field requires that the forecast precipitation field and the vertical velocity field maintain a certain correlation because the actual precipitation process is closely related to vertical motion. Through the optimization of the joint loss function, the forecasting network can generate forecast results that are more consistent with physical laws and observation data.

[0114] Based on the forecast processing results, seamless weather forecast products are generated. These products include continuous time-series forecasts: 0-12 hour short-term forecasts, 12-72 hour medium-term forecasts, and 72-168 hour long-term forecasts. For example, the short-term forecast provides hourly forecasts of precipitation probability, temperature changes, wind speed and direction for the next 12 hours; the medium-term forecast provides weather system evolution information every 6 or 12 hours for the next 72 hours, such as the location and intensity changes of cyclones; and the long-term forecast provides atmospheric circulation patterns and trends of major weather systems every 24 hours for the next 168 hours. These forecast results provide continuous and seamless weather forecast services for users with different needs.

[0115] A time continuity verification algorithm is used to smooth the transitions between different time periods in the forecast products. For example, at the transition between short-term and medium-term forecasts (around 12 hours), the time continuity verification algorithm checks for abrupt changes in the forecast results. If an abrupt change is found, such as a sudden jump of 5°C in temperature around 12 hours, the forecast results are adjusted using interpolation or smoothing algorithms to make the temperature change smoother. Similarly, similar verification and adjustments are performed at the transition between medium-term and long-term forecasts (around 72 hours) to eliminate abrupt changes in the forecast and ensure the temporal continuity of the forecast products.

[0116] Error assessment of forecast results at different time scales is performed using a sliding validation window. For short-term forecasts, the root mean square error (RMSE) is calculated. For example, for temperature forecasts, the square of the difference between the forecast temperature and the actual observed temperature is calculated, then the average value is taken, and finally the square root is taken to obtain the RMSE. For medium-term forecasts, the mean absolute error (MAE) is calculated, which is the average of the absolute differences between the forecast value and the observed value. For long-term forecasts, the trend accuracy is calculated. For example, it is determined whether the forecast temperature trend (rising, falling, or stable) is consistent with the actual observed trend, and the proportion of consistency is used as the trend accuracy. The weight generation strategy of the dynamic gating network is dynamically adjusted based on the error assessment results. If a large forecast error is found at a certain time scale, such as a high RMSE in the short-term forecast, it may be due to an unreasonable weight allocation in the short-term weight map. In this case, the parameters of the dynamic gating network can be appropriately adjusted to optimize the fusion ratio of features at different time scales, thereby improving the forecast accuracy.

[0117] Example 2

[0118] Please see Figure 2 A seamless weather forecast product generation system based on multi-timescale element fusion includes:

[0119] The data acquisition module is used to acquire multi-source meteorological observation data and numerical model data. The multi-source meteorological observation data includes radar echo data, satellite cloud image data and ground observation station data at different time resolutions.

[0120] The time scale unification module is used to perform time scale unification processing on the multi-source meteorological observation data and numerical model data to obtain standardized time series data containing short-term high-frequency observation sequences, medium-term mid-frequency simulation sequences and long-term low-frequency trend sequences.

[0121] A multi-scale feature extraction module is used to extract basic features from the standardized time series data to obtain short-term feature maps, medium-term feature maps, and long-term feature maps.

[0122] The physical sensing dynamic gating module includes a physical quantity calculation unit, a physical encoding unit, and a dynamic weight generation unit; the physical quantity calculation unit is used to calculate physical quantity field parameters; the physical encoding unit is used to encode the physical quantity field parameters into a physical context vector; and the dynamic weight generation unit is used to generate a dynamic weight map corresponding to the spatial location of the feature map.

[0123] The feature fusion module is used to perform adaptive fusion processing on the multi-scale feature map through the dynamic weight map to obtain the fused feature map.

[0124] The forecast generation module is used to input the fused feature map into the forecast network for prediction processing. During the training process, a joint loss function including a physical consistency constraint term is used to generate seamless weather forecast products.

[0125] The product optimization module is used to perform time continuity verification and error assessment on forecast products, and dynamically optimize the fusion strategy.

[0126] The dynamic weight generation unit of the physical perception dynamic gating module adopts a gating network structure with an attention mechanism, which can dynamically adjust the weight distribution of features at different time scales according to real-time physical field parameters, and automatically enhance the weight ratio of short-term high-frequency observation data in strong convection regions.

[0127] The data acquisition module includes multiple data receiving terminals. For radar echo data, there is a dedicated radar data receiver that can connect to radar sites via a network to receive radar data in real time. For satellite cloud image data, there is a satellite data receiving antenna and decoding equipment that can receive signals from polar-orbiting and geostationary satellites and decode them into usable cloud image data. For ground observation station data, it connects to various observation stations through a meteorological data communication network to receive data uploaded by the observation stations.

[0128] The data acquisition module software has a data format conversion function. Since data from different sources may have different formats—for example, radar data may be in a proprietary binary format, satellite cloud image data may be in an image format, and ground observation station data may be in a text format—the software can convert these different data formats into a unified internal data format to facilitate subsequent processing. At the same time, the software also has a data quality check function, which can check the integrity and accuracy of the data. For example, it can check whether radar data has missing scanning angles, whether satellite cloud image data has image noise, and whether ground observation station data is within a reasonable physical range, such as whether the air temperature is within the normal atmospheric temperature range.

[0129] The timescale unification module mainly consists of servers, which are equipped with time processing software and have sufficient computing power to process large amounts of meteorological data.

[0130] The software can implement time interpolation algorithms. For radar echo data, the software has built-in linear interpolation and spline interpolation algorithms. Users can choose the appropriate interpolation algorithm according to the characteristics of the data. For example, for rapidly changing strong convective radar data, linear interpolation algorithm is preferred to maintain the dynamic change characteristics of the data. The software can also perform time matching processing on numerical model data, extracting simulated data that is time-aligned with the observation sequence through interpolation algorithms. At the same time, the software has a sliding window resampling function. Users can set different sliding window lengths and time intervals to generate short-term high-frequency sequences, medium-term mid-frequency sequences, and long-term low-frequency sequences. The software interface is user-friendly. Users can set parameters such as time interval and sliding window length through the graphical interface and can view the processed standardized time series data in real time.

[0131] The multi-scale feature extraction module consists of a high-performance graphics processing unit (GPU) server, which has powerful parallel computing capabilities that can accelerate the computation of convolutional neural networks.

[0132] The software implements a multi-branch convolutional neural network. Users can configure the network structure through the software interface, such as setting parameters like the number of convolutional layers and kernel size for short-term, medium-term, and long-term feature extraction branches. The software can automatically load standardized time-series data and input it into the corresponding branches for feature extraction. In the short-term feature extraction branch, the software can extract refined evolution features of convective systems; in the medium-term feature extraction branch, it can extract the movement and development features of weather systems; and in the long-term feature extraction branch, it can extract atmospheric circulation background features. The software also has feature visualization capabilities, allowing users to intuitively view short-term, medium-term, and long-term feature maps to understand the feature distribution at different time scales.

[0133] The physical sensing dynamic gating module is also composed of GPU servers to meet the computational needs of physical quantity calculation and neural network operation.

[0134] The physical quantity calculation unit can calculate physical quantity field parameters based on the input standardized time series data, such as vertical integral liquid water content, divergence field, and pseudo-equivalent potential temperature. The physical encoding unit encodes these physical quantity field parameters into physical context vectors, and the encoding process is implemented through a deep neural network. The dynamic weight generation unit generates a dynamic weight map based on the physical context vectors. The software can dynamically adjust the weight allocation of features at different time scales based on real-time physical quantity field parameters. For example, it can automatically increase the weight ratio of short-term high-frequency observation data in strong convection regions. The software also provides a visualization function for the weight map, allowing users to view the weight distribution of short-term, medium-term, and long-term weight maps at different spatial locations.

[0135] The feature fusion module consists of GPU servers and is used to perform feature fusion calculations efficiently.

[0136] The software can receive dynamic weight maps and multi-scale feature maps as inputs and perform adaptive fusion processing according to the above fusion formula. The software also has a feature enhancement function, which introduces key physical signals from the original multi-scale feature maps through residual connection. The software interface provides a visualization function for the fused feature maps, allowing users to view the comparison between the fused feature maps and the original feature maps and evaluate the fusion effect.

[0137] The forecast generation module consists of a high-performance GPU server used to run deep convolutional neural networks for forecasting.

[0138] The software can input fused feature maps into the forecast network for prediction processing. During model training, a joint loss function is used for parameter optimization. The joint loss function includes a mean squared error loss term and a physical consistency constraint term. The software can generate seamless weather forecast products based on the prediction processing results, including continuous time series forecasts of 0-12 hour short-term forecasts, 12-72 hour medium-term forecasts, and 72-168 hour long-term forecasts. The software also has a forecast product visualization function, allowing users to view forecast results at different time scales through a graphical interface, including information such as precipitation probability, temperature changes, wind speed and direction.

[0139] The product optimization module consists of servers used to run time continuity verification algorithms and error evaluation algorithms.

[0140] The software employs a time continuity verification algorithm to smooth the transitions between different time periods in the forecast products, eliminating abrupt changes in forecasts. It uses a sliding verification window to assess the errors in forecast results at different time scales, calculating the root mean square error of short-term forecasts, the mean absolute error of medium-term forecasts, and the trend accuracy of long-term forecasts. Based on the error assessment results, it dynamically adjusts the weight generation strategy of the dynamic gating network, optimizing the fusion ratio of features at different time scales. The software interface provides visualization of the error assessment results, allowing users to intuitively understand the accuracy of the forecast products and adjust optimization strategies as needed.

[0141] Example 3

[0142] A computer-readable storage medium storing a computer program, the computer program including program instructions that, when executed by a processor, cause the processor to perform the seamless weather forecast product generation method based on multi-timescale element fusion as described above.

[0143] This embodiment provides a computer-readable storage medium storing a computer program containing a series of program instructions. When these instructions are executed by a computer processor, a method for generating seamless weather forecast products based on the fusion of multiple time-scale elements can be implemented. Specifically, the method involves acquiring meteorological observation data and numerical model data from multiple sources, processing them uniformly on a time scale, extracting multi-scale features and calculating physical field parameters, generating a weight map through physical sensing dynamic gating, fusing the feature maps, and finally training a forecast network using a joint loss function containing physical consistency constraints to generate seamless weather forecast products covering short-term, medium-term, and long-term forecasts. The entire process is highly automated, accurate, and efficient. With the help of the program in this storage medium, meteorological forecasting agencies can quickly deploy and run this advanced method, significantly improving the accuracy and timeliness of weather forecasts and providing better meteorological services to the public and related industries.

[0144] The same or similar labels correspond to the same or similar parts;

[0145] The terms used to describe positional relationships in the accompanying drawings are for illustrative purposes only and should not be construed as limiting this patent.

[0146] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the implementation of the present invention. For those skilled in the art, other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all implementation methods here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the claims of the present invention.

Claims

1. A method for generating seamless weather forecast products based on multi-timescale element fusion, characterized in that, Includes the following steps: Acquire multi-source meteorological observation data and numerical model data. The multi-source meteorological observation data includes radar echo data, satellite cloud image data and ground observation station data at different time resolutions. The numerical model data includes atmospheric physical field data at different forecast lead times. The multi-source meteorological observation data and numerical model data are subjected to time-scale unification processing to obtain standardized time series data, which includes short-term high-frequency observation sequences, medium-term mid-frequency simulation sequences, and long-term low-frequency trend sequences. The basic features are extracted from the standardized time series data to obtain a multi-scale feature map, which includes a short-term feature map, a medium-term feature map, and a long-term feature map. Calculate the physical field parameters corresponding to the standardized time series data, including vertical integral liquid water content, divergence field, and pseudo-equivalent potential temperature; input the physical field parameters into a physical sensing coding network to generate a physical context vector; The physical context vector is input into a dynamic gating network to generate a dynamic weight map corresponding to the spatial location of the multi-scale feature map. The dynamic weight map reflects the credibility of features at different time scales at each spatial grid point. The multi-scale feature map is adaptively fused using the dynamic weight map to obtain a fused feature map; wherein, the weight ratio of short-term high-frequency observation sequences is increased in strong convection regions, and the weight ratio of long-term low-frequency trend sequences is increased in stable weather regions. The fused feature map is input into the prediction network for prediction processing. During model training, a joint loss function is used for parameter optimization. The joint loss function includes a mean squared error loss term and a physical consistency constraint term. Seamless weather forecast products are generated based on the forecast processing results. These seamless weather forecast products include continuous time-series forecast results of 0-12 hour short-term forecasts, 12-72 hour medium-term forecasts, and 72-168 hour long-term forecasts.

2. The method according to claim 1, characterized in that, The process of unifying the time scale of the multi-source meteorological observation data and numerical model data to obtain standardized time series data includes: A time interpolation algorithm is used to unify multi-source meteorological observation data with different time resolutions into an observation sequence with a preset time interval; The numerical model data is time-matched to extract the simulation data that is time-aligned with the observation sequence, forming a simulation sequence. The observed and simulated sequences are resampled using the sliding window method to generate short-term high-frequency sequences, medium-term mid-frequency sequences, and long-term low-frequency sequences, forming standardized time series data.

3. The method according to claim 1, characterized in that, The extraction of basic features from the standardized time series data to obtain a multi-scale feature map includes: Construct a multi-branch convolutional neural network, wherein the multi-branch convolutional neural network includes a short-term feature extraction branch, a medium-term feature extraction branch, and a long-term feature extraction branch; Short-term high-frequency observation sequences are input into the short-term feature extraction branch, and refined evolution features of the convective system are extracted through multi-layer convolution and pooling operations to obtain a short-term feature map; The mid-term intermediate frequency simulation sequence is input into the mid-term feature extraction branch to extract the movement and development features of the weather system, thus obtaining the mid-term feature map; The long-term low-frequency trend sequence is input into the long-term feature extraction branch to extract atmospheric circulation background features and obtain a long-term feature map.

4. The method according to claim 1, characterized in that, The step of inputting the physical context vector into a dynamic gating network to generate a dynamic weight map corresponding to the spatial location of the multi-scale feature map includes: The dynamic gating network consists of a fully connected layer and a sigmoid activation function, receiving a physical context vector and outputting a weight matrix with the same size as the feature map. The weight matrix includes a short-term weight map, a medium-term weight map, and a long-term weight map, and the sum of the weights in spatial locations is 1. In strong convection regions where the vertical integral liquid water content is higher than a preset threshold, the weight value of the corresponding position in the short-term weighted map is ≥0.7; in stable regions where the absolute value of divergence is lower than a preset threshold, the weight value of the corresponding position in the long-term weighted map is ≥0.

6.

5. The method according to claim 1, characterized in that, The step of adaptively fusing the multi-scale feature map using the dynamic weight map to obtain a fused feature map includes: The formula for calculating the fused feature map is: ; in, This is a short-term weighted chart. For short-term feature maps, This is a mid-term weighted chart. For mid-term feature maps, For long-term weighted charts, For long-term feature maps; Feature enhancement processing is performed on the fused feature map, and key physical signals are introduced into the original multi-scale feature map through residual connection to enhance the feature representation of the strong convection system.

6. The method according to claim 1, characterized in that, The formula for calculating the joint loss function is as follows: ; in, This is the mean squared error loss term, used to measure the deviation between the forecast result and the observed value; This is a physical consistency constraint term, used to constrain the statistical relationship between the forecast field and the physical quantity field; and These are the weighting coefficients, and + = 1; The physical consistency constraints include: the correlation constraint between the precipitation field and the vertical velocity field, the total precipitation conservation constraint, and the thermodynamic relationship constraint between the temperature field and the pressure field.

7. The method according to claim 1, characterized in that, After generating a seamless weather forecast product, the method further includes: A time continuity verification algorithm is used to smooth the transitions between different time periods in the forecast product, eliminating abrupt changes in the forecast. Error assessment of forecast results at different time scales is performed using a sliding validation window, calculating the root mean square error of short-term forecasts, the mean absolute error of medium-term forecasts, and the trend accuracy of long-term forecasts. The weight generation strategy of the dynamic gating network is dynamically adjusted based on the error evaluation results to optimize the fusion ratio of features at different time scales.

8. A seamless weather forecast product generation system based on multi-timescale element fusion, used to achieve this through the method described in any one of claims 1-7, characterized in that, include: The data acquisition module is used to acquire multi-source meteorological observation data and numerical model data. The multi-source meteorological observation data includes radar echo data, satellite cloud image data and ground observation station data at different time resolutions. The time scale unification module is used to perform time scale unification processing on the multi-source meteorological observation data and numerical model data to obtain standardized time series data containing short-term high-frequency observation sequences, medium-term mid-frequency simulation sequences and long-term low-frequency trend sequences. A multi-scale feature extraction module is used to extract basic features from the standardized time series data to obtain short-term feature maps, medium-term feature maps, and long-term feature maps. The physical sensing dynamic gating module includes a physical quantity calculation unit, a physical encoding unit, and a dynamic weight generation unit; the physical quantity calculation unit is used to calculate physical quantity field parameters; the physical encoding unit is used to encode the physical quantity field parameters into a physical context vector; and the dynamic weight generation unit is used to generate a dynamic weight map corresponding to the spatial location of the feature map. The feature fusion module is used to perform adaptive fusion processing on the multi-scale feature map through the dynamic weight map to obtain the fused feature map. The forecast generation module is used to input the fused feature map into the forecast network for prediction processing. During the training process, a joint loss function including a physical consistency constraint term is used to generate seamless weather forecast products. The product optimization module is used to perform time continuity verification and error assessment on forecast products, and dynamically optimize the fusion strategy.

9. The system according to claim 8, characterized in that, The dynamic weight generation unit of the physical perception dynamic gating module adopts a gating network structure with an attention mechanism, which can dynamically adjust the weight distribution of features at different time scales according to real-time physical field parameters, and automatically enhance the weight ratio of short-term high-frequency observation data in strong convection regions.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, the computer program including program instructions, which, when executed by a processor, cause the processor to perform the seamless weather forecast product generation method based on multi-timescale element fusion as described in any one of claims 1-7.

Images