A weather forecast method and system for severe convective weather
Through the construction of multi-source data fusion and deep learning models, the problems of high air reporting and underreporting rates of strong convective weather forecasts in the existing technology have been solved, and more accurate forecasts of short-term heavy precipitation and thunderstorms and strong winds have been achieved, which has improved the Forecast performance.
Patent Information
- Application Number
- CN202411245035.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-06
- Publication Date
- 2024-11-12
- Estimated Expiration
- 2044-09-06
AI Technical Summary
The prior art has high hollow and underreport rates for short-term heavy rainfall and thunderstorms and strong winds, and the forecast range is relatively large compared with the actual situation, making it difficult to effectively monitor and forecast strong convective weather.
Using a variety of deep learning models, including SEResUNet, U2Net, PhyDnet and StaRSU, through multi-source data fusion and feature extraction, a forecast model of 0-2 hours and 2-12 hours is constructed, and the DiceLoss loss function is used to optimize the model to improve forecast accuracy. .
It significantly improves the forecast accuracy of short-term heavy rainfall and thunderstorms and strong winds, reduces the air report rate and underreport rate, improves forecast performance, and provides more accurate 0-12-hour short-term objective forecast products.
Smart Images

Figure CN118759611B_ABST
Abstract
Description
Technical Field
[0001] The invention relates to a weather forecast method and system for severe convective weather, belonging to the technical field of weather forecasting. Background Art
[0002] For areas with complex terrain and frequent severe convective weather processes, short-term warning forecasts, as the last line of defense for meteorological services, undertake the important responsibility of severe convection monitoring and forecasting. In recent years, we have carried out research on objective meteorological forecasting technology and product development based on artificial intelligence, built a number of high-performance and GPU computing servers, and established a set of high-resolution training data sets by combining deep learning neural networks, Doppler radars, and mesoscale regional models. We mainly model AI forecast models for two types of severe convective weather, short-term heavy rainfall and thunderstorms. We have achieved objective forecast products covering areas with frequent severe convection and surrounding areas (24.5°N~30.25°N, 108.65°E-114.4°E), with a spatial resolution of 5km, a temporal resolution of 1 hour, and a forecast time of 0-12 hours.
[0003] The existing technical methods still have a high rate of false alarms and missed reports in forecasting short-term heavy rainfall and thunderstorms and gale forces, and the forecast range is often larger than the actual situation. Summary of the invention
[0004] The purpose of the present invention is to overcome the deficiencies in the prior art and provide a weather forecast method for severe convective weather, which can achieve objective forecasts of short-term heavy rainfall and thunderstorm gale within 0-12 hours.
[0005] To achieve the above object, the present invention is implemented by adopting the following technical solutions:
[0006] In a first aspect, the present invention provides a weather forecast method for severe convective weather, comprising:
[0007] Acquire real-time multi-source data, input it into pre-trained deep learning models, and output forecast results;
[0008] The training process of the multiple deep learning models that have been trained includes:
[0009] Acquiring historical multi-source data, wherein the historical multi-source data is meteorological-related data collected and integrated from multiple sources;
[0010] Preprocess historical multi-source data;
[0011] The pre-processed historical multi-source data is divided and respectively input into a plurality of pre-built deep learning models for training to obtain a plurality of trained deep learning models;
[0012] The methods for constructing the multiple deep learning models include:
[0013] Construct a 0-2 hour forecast model, which includes a thunderstorm gale forecast model based on the SEResUNet deep neural network and a U 2 Net deep neural network and PhyDnet deep neural network short-term heavy rainfall forecast model;
[0014] Constructing a 2-12 hour forecast model, wherein the 2-12 hour forecast model includes: a thunderstorm gale area forecast model and a short-term heavy rainfall area forecast model based on the StaRSU deep neural network;
[0015] The pre-processed historical multi-source data is divided and input into a plurality of pre-built deep learning models for training, including:
[0016] For the 0-2 hour forecast model, near-real-time data is extracted from the real-time observation data, and the 0-2 hour forecast model is trained by the near-real-time data;
[0017] For the 2-12 hour forecast model, the forecast data of the CMA-GD model, the Doppler weather radar and the real-time observation data of the regional automatic station are used as input data, and the 2-12 hour forecast model is trained through multi-source feature fusion.
[0018] Furthermore, the historical multi-source data includes forecast data of the CMA-GD model, Doppler weather radar and real-time observation data of regional automatic stations.
[0019] Furthermore, the U 2 Net deep neural network includes multiple layers of RSU modules with different depths, and the RSU modules are used to capture the contextual information of meteorological data at different scales.
[0020] Furthermore, the StaRSU deep neural network extracts multi-scale feature information through the RSU module, fuses the feature information through the fusion module, extracts key features of the fused feature information through the SE attention module, and predicts and outputs the key feature extraction results through the global pooling layer.
[0021] Furthermore, the PhyDnet deep neural network introduces the PhyCell recurrent physical unit based on the ConvLSTM convolutional long short-term memory network, and uses the convolutional network to simulate partial differential equations to model physical knowledge. Its position is parallel to that of ConvLSTM, and the input feature quantity passes through ConvLSTM and PhyCell at the same time to extract spatiotemporal information and physical information respectively, and the two are fused through the decoder for final prediction and output.
[0022] Furthermore, when the 2-12 hour forecast model is trained through multi-source feature fusion, the model encoder adopts a multi-branch structure to perform multi-scale feature extraction and fusion on the forecast data of the Doppler weather radar and CMA-GD mode respectively, and predicts and outputs the multi-scale feature extraction and fusion results through a pyramid structure network.
[0023] Furthermore, the method further includes: in the training phase, using DiceLoss as a loss function to optimize multiple deep learning models, and the calculation formula of the loss function is as follows:
[0024] ;
[0025] in, is the DiceLoss loss, N is the number of grid points, For the The predicted probability corresponding to the number of grid points is For the The true label corresponding to the number of grid points.
[0026] In a second aspect, the present invention provides a weather forecast system for severe convective weather, comprising:
[0027] The weather forecast module is used to obtain real-time multi-source data, input it into a variety of pre-trained deep learning models, and output forecast results;
[0028] The training process of the multiple deep learning models includes:
[0029] Acquiring historical multi-source data, wherein the historical multi-source data is meteorological-related data collected and integrated from multiple sources;
[0030] Preprocess historical multi-source data;
[0031] The pre-processed historical multi-source data is divided and respectively input into a plurality of pre-built deep learning models for training to obtain a plurality of trained deep learning models;
[0032] The methods for constructing the multiple deep learning models include:
[0033] Construct a 0-2 hour forecast model, which includes a thunderstorm gale forecast model based on the SEResUNet deep neural network and a U 2 Net deep neural network and PhyDnet deep neural network short-term heavy rainfall forecast model;
[0034] Constructing a 2-12 hour forecast model, wherein the 2-12 hour forecast model includes: a thunderstorm gale area forecast model and a short-term heavy rainfall area forecast model based on the StaRSU deep neural network;
[0035] The pre-processed historical multi-source data is divided and input into a plurality of pre-built deep learning models for training, including:
[0036] For the 0-2 hour forecast model, near-real-time data is extracted from the real-time observation data, and the 0-2 hour forecast model is trained by the near-real-time data;
[0037] For the 2-12 hour forecast model, the forecast data of the CMA-GD model, the Doppler weather radar and the real-time observation data of the regional automatic station are used as input data, and the 2-12 hour forecast model is trained through multi-source feature fusion.
[0038] Compared with the prior art, the present invention has the following beneficial effects:
[0039] The present invention provides a weather forecasting method and system for severe convective weather, proposes a new technical deep learning model for short-term heavy precipitation and thunderstorm gale forecasting method, adopts a multi-branch structure to extract features for multiple types of variables separately, and performs feature fusion through the SE attention module. The method not only improves the feature expression ability of the network, but also effectively reduces the number of parameters and improves the generalization ability of the network. The use of DiceLoss for 2-12 hour forecasts can effectively reduce the impact of data imbalance, and the forecast effect is significantly improved, which can provide a technical reference for the short-term objective forecast of short-term heavy precipitation and thunderstorm gale. Its forecasting performance is better than the forecast products based on the traditional technical methods of batching and fuzzy logic methods to a certain extent. Through the inspection and evaluation of the severe convective forecast in Hunan since 2023, the TS score, hit rate and other indicators of the objective forecast products of short-term heavy precipitation and thunderstorm gale are better than the existing objective forecast products. The system has the technical characteristics of stable operation, fast calculation, strong portability, etc., and has high application value in the field of meteorological forecasting. BRIEF DESCRIPTION OF THE DRAWINGS
[0040] Figure 1 is a flow chart of a meteorological forecast method for severe convective weather provided by an embodiment of the present invention;
[0041] Figure 2 It is a schematic diagram of TS scores of the Central Station guidance report on short-term heavy rainfall from April to September 2023 and the objective forecast product of Hunan provided by an embodiment of the present invention;
[0042] Figure 3It is a schematic diagram of the hit rate of the Central Station guidance report on short-term heavy rainfall from April to September 2023 and the objective forecast product of Hunan provided by an embodiment of the present invention;
[0043] Figure 4 It is a schematic diagram of the false alarm rate of the Central Station's guidance report on short-term heavy rainfall from April to September 2023 and the objective forecast product of Hunan Province provided by an embodiment of the present invention;
[0044] Figure 5 It is a schematic diagram of the underreporting rate of the Central Station guidance report on short-term heavy rainfall from April to September 2023 and the objective forecast product of Hunan provided by an embodiment of the present invention;
[0045] Figure 6 It is a schematic diagram of TS scores of the Central Station's guidance report on thunderstorms and gale from April to September 2023 and Hunan's objective forecast products provided by the embodiment of the invention;
[0046] Figure 7 It is a schematic diagram of the hit rate of the central station guidance report on thunderstorm and gale from April to September 2023 and the objective forecast product of Hunan provided by the embodiment of the invention;
[0047] Figure 8 It is a schematic diagram of the false alarm rate of the central station guidance report of thunderstorm and gale and Hunan objective forecast products from April to September 2023 provided by the embodiment of the invention;
[0048] Fig. 9 It is a schematic diagram of the underreporting rate of the central station's guidance report on thunderstorms and gale winds and Hunan's objective forecast products from April to September 2023 provided by an embodiment of the invention. DETAILED DESCRIPTION
[0049] The present invention will be further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and cannot be used to limit the protection scope of the present invention.
[0050] Embodiment 1: This embodiment introduces a weather forecast method for severe convective weather, including:
[0051] Acquire real-time multi-source data, input it into pre-trained deep learning models, and output forecast results;
[0052] The training process of the multiple deep learning models that have been trained includes:
[0053] Acquiring historical multi-source data, wherein the historical multi-source data is meteorological-related data collected and integrated from multiple sources;
[0054] Preprocess historical multi-source data;
[0055] The pre-processed historical multi-source data is divided and respectively input into a plurality of pre-built deep learning models for training to obtain a plurality of trained deep learning models;
[0056] The methods for constructing the multiple deep learning models include:
[0057] Construct a 0-2 hour forecast model, which includes a thunderstorm gale forecast model based on the SEResUNet deep neural network and a U 2 Net deep neural network and PhyDnet deep neural network short-term heavy rainfall forecast model;
[0058] Constructing a 2-12 hour forecast model, wherein the 2-12 hour forecast model includes: a thunderstorm gale area forecast model and a short-term heavy rainfall area forecast model based on the StaRSU deep neural network;
[0059] The pre-processed historical multi-source data is divided and input into a plurality of pre-built deep learning models for training, including:
[0060] For the 0-2 hour forecast model, near-real-time data is extracted from the real-time observation data, and the 0-2 hour forecast model is trained by the near-real-time data;
[0061] For the 2-12 hour forecast model, the forecast data of the CMA-GD model, the Doppler weather radar and the real-time observation data of the regional automatic station are used as input data, and the 2-12 hour forecast model is trained through multi-source feature fusion.
[0062] like Figure 1 As shown, the weather forecast method for severe convective weather provided in this embodiment specifically involves the following steps:
[0063] This technology adopts a deep learning severe convection forecasting method based on multi-source data fusion. The model framework consists of four parts: data layer, function layer, algorithm layer and output layer. The data layer mainly includes the forecast data of the CMA-GD model (South China medium-scale numerical forecast model) and the real-time observation data of Doppler weather radar and regional automatic stations. The functional layer mainly performs data preprocessing. According to various factors such as data integrity and historical test results, the data is preprocessed through regional clipping, spatial downscaling, standardization and other methods to construct a standard data set. In the algorithm layer, two different technical solutions are adopted for 0-2 hour nowcasting and 2-12 hour short-term forecasting, and finally the objective forecast of 0-12 hour short-term heavy precipitation and thunderstorm gale is achieved in the output layer.
[0064] (1) 0-2 hour severe convection nowcast
[0065] The 0~2h forecast is greatly affected by the near-real-time data, so the near-real-time data are used for extrapolated forecasting, with Doppler weather radar and minute-level regional automatic station real-time precipitation observation data as multi-source data sets.
[0066] The SEResUNet (a deep learning model that combines SENet and UNet for image segmentation tasks, aiming to improve the performance and accuracy of image segmentation) deep neural network is used to build a 0~2h thunderstorm gale forecast model. The Squeeze-and-Excitation (SE) (attention mechanism) module in the SEResUNet model can effectively capture the complex features of multi-source data and model the interdependence between channel features. In order to better fuse the features of multi-source data of different scales, the encoder part is improved, and a multi-branch structure is used to extract features of multiple types of variables separately, and then the features are fused through the SE attention module, which not only improves the feature expression ability of the network, but also effectively reduces the number of parameters and improves the generalization ability of the network.
[0067] Using U 2 UNet (an improved model based on UNet architecture, designed to improve the performance of image segmentation) is combined with PhyDnet to build a 0~2h short-term heavy rainfall forecast model. 2 The Net model consists of multiple layers of RSU (ReSidual U-block) (a residual module commonly used in deep learning models. It contains multiple residual connections and enhances the model's feature extraction capabilities by stacking these residual connections) modules of different depths. It introduces more pooling operations and receptive fields of different sizes, which can extract rich multi-scale context features and significantly improve the prediction performance of the model. PhyDNet (a physical dynamic network, the main idea is to try to use a deep network to build a physical constraint model, the method is to use convolution to simulate partial derivatives, use moment loss for supervision, learn physical information, and thus supplement the information of the existing network) introduces the PhyCell cyclic physical unit, simulates PDE through convolution in the physical predictor part, and introduces momentum loss for restriction, so as to better perform physical dynamic modeling. Inspired by data assimilation, real-time forecast correction is introduced. Based on PhyCell (a neural network structure based on physical units, used to process spatiotemporal series data, which can effectively capture the structure and dynamic information in spatiotemporal data) and ConvLSTM (a deep learning model that combines convolutional neural networks and long short-term memory networks, which can simultaneously consider spatiotemporal information and time series information, so as to better capture the structure and dynamic characteristics in the data), a dual-branch disentanglement structure is constructed to learn physical dynamics and other unknown information required for forecasting. 2Net has a more accurate forecast for precipitation areas, while PhyDNet has a better forecast for the motion trend of precipitation development. 2 Net and PhyDNet predictions can effectively improve the generalization ability and stability of the model.
[0068] (2) 2-12 hour severe convection forecast
[0069] The 2-12h forecast time is relatively long, and the use of real-time data extrapolation will result in serious attenuation. Therefore, the numerical model forecast field is added for constraint, and the forecast data of the CMA-GD model, Doppler weather radar, and real-time observation data of hourly regional automatic stations are used as multi-source data sets.
[0070] The StaRSU deep neural network is used to build a 2-12 hour severe convection classification area forecast model. In order to effectively integrate multi-source data with different temporal and spatial resolutions for training, the StaRSU network encoder uses a dual-branch structure to extract features from model forecast data and radar data respectively. The encoder feature extraction is constructed using the RSU module. The feature information extracted by the dual branches is spliced along the channel and then extracted through the convolution fusion module. The site prediction output is then realized through the pyramid structure prediction network.
[0071] Since severe convective event data has extremely high data imbalance, in addition to balancing measures such as resampling the data during the training phase, a suitable loss function needs to be used. According to relevant research by team members, DiceLoss (a loss function commonly used in image segmentation tasks, mainly used to measure the similarity between the predicted results and the true label) can effectively reduce the impact of data imbalance. For each category, Diceloss only calculates the impact of positive events in that category, while ignoring the impact of a large number of negative events. The loss of each category has the same magnitude of contribution, so there is no need to give each category a corresponding weight for balancing. Its training for category imbalanced data is significantly better than MAE (mean absolute error), MSE (mean square error), Weighted Cross Entropy (a loss function commonly used in deep learning, usually used to deal with category imbalance) and other loss functions, while improving the hit rate and effectively suppressing the increase in the false alarm rate. The calculation formula is as follows:
[0072] ;
[0073] Where N is the number of grid points, For the The predicted probability corresponding to the number of grid points is For the The true label corresponding to the number of grid points.
[0074] (3) Inspection and evaluation
[0075] The TS score, hit rate, false alarm rate, missed alarm rate and other indicators are used to test the hourly short-term heavy precipitation and thunderstorm gale classification forecasts from 0 to 12 hours. The SAL spatial test is used to analyze individual cases:
[0076] (31) Hourly short-term heavy rainfall forecast from 0 to 12 hours
[0077] For short-term heavy rainfall from April to September 2023, the statistical scoring results show that Figure 2 As shown, the comprehensive TS score of Hunan's objective forecast products in 2023 is 0.12, and the hit rate, false alarm rate and missed alarm rate are 27.4%, 82.5% and 72.6% respectively. All indicators have obvious advantages over the 2022 objective forecast products and the CCTV guidance reports.
[0078] like Figure 3-Figure 5 As shown in the figure, from the results of hit rate, false alarm rate and missed alarm rate, it can be seen that the comprehensive hit rate of Hunan's objective forecast products in 2023 was 27.4%, 19.7% in 2022, and 19.0% in the guidance report of the Central Station; the comprehensive false alarm rate of Hunan in 2023 was 82.5%, 82.8% in 2022, and 89.2% in the guidance report of the Central Station; the comprehensive missed alarm rate of Hunan in 2023 was 72.6%, 80.3% in 2022, and 81.0% in the guidance report of the Central Station. It can be seen that the hit rate of Hunan's objective forecast products in 2023 is higher than that of the 22-year products and the guidance report of the Central Station, while the false alarm rate and missed alarm rate are lower than those of the 22-year products and the guidance report of the Central Station. The advantages are mainly reflected in the first 6 hours, especially the first 2 hours, which is consistent with the results shown by the TS score.
[0079] The short-term heavy rainfall forecast reported from 02:00 on June 30, 2023 was tested after target matching with the actual situation. The results show that the first forecast time period has relatively accurate forecast area and location; in the first 6 forecast time periods, the forecast range is basically consistent with the actual situation, and the main error of the location forecast is the dispersion of the actual situation; as the forecast time period continues to extend, the overall forecast position is biased to the north and more dispersed.
[0080] (32) Hourly thunderstorm and gale forecast from 0 to 12 hours
[0081] For the 0-12 hour thunderstorm gale from April to September 2023, the statistical scoring results are as follows: Figure 6As shown, the Hunan Comprehensive TS score (a statistical indicator used to evaluate the forecast accuracy of numerical weather forecast models) in 2023 was 0.041, 0.027 in 2022, and 0.007 in the Central Station guidance report. It can be seen that the Hunan Comprehensive TS score in 2023 is higher than the 22-year product and the Central Station guidance report product, and the overall performance is the best, which is 0.014 and 0.033 higher than the Hunan objective forecast product and the Central Station guidance report in 2022, respectively.
[0082] like Figure 7-Figure 9 As shown in the figure, from the results of hit rate, false alarm rate and missed alarm rate, it can be seen that the hit rate of Hunan in 2023 was 11.0%, 6.6% in 2022, and 6.4% in the guidance report of the Central Station; the false alarm rate of Hunan in 2023 was 93.7%, the objective forecast product in 2022 was 95.7%, and the guidance report of the Central Station was 99.2%; the missed alarm rate of Hunan in 2023 was 89.0%, the objective forecast product in 2022 was 93.4%, and the guidance report of the Central Station was 93.6%. It can be seen that the hit rate of Hunan's objective forecast products in 2023 is higher than that of the 22-year products and the guidance report of the Central Station, and the false alarm rate and missed alarm rate are lower than those of the 22-year products and the guidance report of the Central Station. Its advantage is reflected in 3-12 hours, and compared with the guidance report of the Central Station, the objective forecast products of Hunan in 2023 have certain advantages over all time periods.
[0083] The target matching test was conducted on the thunderstorm gale forecast reported from 02:00 on April 18, 2023, and the actual situation. The results show that the forecast range is more dispersed than the actual situation, and the position deviation of the first 6 forecast time periods is larger. The main error is that the individual forecast is more dispersed. The longer forecast time period is manifested as a decrease in position deviation, and the overall fall area is biased to the north.
[0084] Embodiment 2: This embodiment provides a weather forecast system for severe convective weather, including:
[0085] The weather forecast module is used to obtain real-time multi-source data, input it into a variety of pre-trained deep learning models, and output forecast results;
[0086] The training process of the multiple deep learning models includes:
[0087] Acquiring historical multi-source data, wherein the historical multi-source data is meteorological-related data collected and integrated from multiple sources;
[0088] Preprocess historical multi-source data;
[0089] The pre-processed historical multi-source data is divided and respectively input into a plurality of pre-built deep learning models for training to obtain a plurality of trained deep learning models;
[0090] The methods for constructing the multiple deep learning models include:
[0091] Construct a 0-2 hour forecast model, which includes a thunderstorm gale forecast model based on the SEResUNet deep neural network and a U 2 Net deep neural network and PhyDnet deep neural network short-term heavy rainfall forecast model;
[0092] Constructing a 2-12 hour forecast model, wherein the 2-12 hour forecast model includes: a thunderstorm gale area forecast model and a short-term heavy rainfall area forecast model based on the StaRSU deep neural network;
[0093] The pre-processed historical multi-source data is divided and input into a plurality of pre-built deep learning models for training, including:
[0094] For the 0-2 hour forecast model, near-real-time data is extracted from the real-time observation data, and the 0-2 hour forecast model is trained by the near-real-time data;
[0095] For the 2-12 hour forecast model, the forecast data of the CMA-GD model, the Doppler weather radar and the real-time observation data of the regional automatic station are used as input data, and the 2-12 hour forecast model is trained through multi-source feature fusion.
[0096] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.
[0097] It should be understood by those skilled in the art that the embodiments of the present disclosure may be provided as methods, systems or computer program products. Therefore, the present disclosure may take the form of a complete hardware embodiment, a complete software embodiment or an embodiment combining software and hardware aspects. Moreover, the present disclosure may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0098] The present disclosure is described with reference to flowcharts and / or block diagrams of methods, devices (systems) and computer program products according to embodiments of the present disclosure. It should be understood that each process and / or block in the flowchart and / or block diagram and the combination of processes and / or blocks in the flowchart and / or block diagram can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.
[0099] These computer program instructions may also be stored in a computer readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture including an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.
[0100] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process in the computer or other programmable device. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.
[0101] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present disclosure rather than to limit its protection scope. Although the present disclosure has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that after reading the present disclosure, those skilled in the art can still make various changes, modifications or equivalent substitutions to the specific implementation methods of the invention, but these changes, modifications or equivalent substitutions are all within the protection scope of the disclosed claims to be approved.
Claims
1. A weather forecast method for severe convective weather, characterized in that: include: Acquire real-time multi-source data, input it into pre-trained deep learning models, and output forecast results; The training process of the multiple deep learning models that have been trained includes: Acquiring historical multi-source data, wherein the historical multi-source data is meteorological-related data collected and integrated from multiple sources; Preprocess historical multi-source data; The pre-processed historical multi-source data is divided and respectively input into a plurality of pre-built deep learning models for training to obtain a plurality of trained deep learning models; The methods for constructing the multiple deep learning models include: Construct a 0-2 hour forecast model, which includes a thunderstorm gale forecast model based on the SEResUNet deep neural network and a U 2 Net deep neural network and PhyDnet deep neural network short-term heavy rainfall forecast model; Constructing a 2-12 hour forecast model, wherein the 2-12 hour forecast model includes: a thunderstorm gale area forecast model and a short-term heavy rainfall area forecast model based on the StaRSU deep neural network; The pre-processed historical multi-source data is divided and input into a plurality of pre-built deep learning models for training, including: For the 0-2 hour forecast model, near-real-time data is extracted from the real-time observation data, and the 0-2 hour forecast model is trained by the near-real-time data; For the 2-12 hour forecast model, the forecast data of the CMA-GD model, the Doppler weather radar and the real-time observation data of the regional automatic station are used as input data, and the 2-12 hour forecast model is trained through multi-source feature fusion.
2. The weather forecast method for severe convective weather according to claim 1, characterized in that: The historical multi-source data include forecast data of CMA-GD model, Doppler weather radar and real-time observation data of regional automatic stations.
3. The weather forecast method for severe convective weather according to claim 1, characterized in that: The U 2 Net deep neural network includes multiple layers of RSU modules with different depths, and the RSU modules are used to capture the contextual information of meteorological data at different scales.
4. The weather forecast method for severe convective weather according to claim 3, characterized in that: The StaRSU deep neural network extracts multi-scale feature information through the RSU module, fuses the feature information through the fusion module, extracts key features from the fused feature information through the SE attention module, and predicts and outputs the key feature extraction results through the global pooling layer.
5. The weather forecast method for severe convective weather according to claim 1, characterized in that: The PhyDnet deep neural network introduces the PhyCell recurrent physical unit on the basis of the ConvLSTM convolutional long short-term memory network, and uses the convolutional network to simulate partial differential equations to model physical knowledge. Its position is parallel to that of the ConvLSTM. The input feature quantity passes through the ConvLSTM and the PhyCell recurrent physical unit at the same time to extract spatiotemporal information and physical information respectively, and the two are fused through the decoder for final prediction and output.
6. The weather forecast method for severe convective weather according to claim 1, characterized in that: When the 2-12 hour forecast model is trained by multi-source feature fusion, the model encoder adopts a multi-branch structure to extract and fuse multi-scale features of the forecast data of the Doppler weather radar and CMA-GD mode respectively, and predicts and outputs the multi-scale feature extraction and fusion results through a pyramid structure network.
7. The weather forecast method for severe convective weather according to claim 1, characterized in that: The method further includes: in the training phase, using DiceLoss as a loss function to optimize multiple deep learning models, and the calculation formula of the loss function is as follows: ; in, is the DiceLoss loss, N is the number of grid points, For the The predicted probability corresponding to the number of grid points is For the The true label corresponding to the number of grid points.
8. A weather forecast system for severe convective weather, characterized in that: include: The weather forecast module is used to obtain real-time multi-source data, input it into a variety of pre-trained deep learning models, and output forecast results; The training process of the multiple deep learning models includes: Acquiring historical multi-source data, wherein the historical multi-source data is meteorological-related data collected and integrated from multiple sources; Preprocess historical multi-source data; The pre-processed historical multi-source data is divided and respectively input into a plurality of pre-built deep learning models for training to obtain a plurality of trained deep learning models; The methods for constructing the multiple deep learning models include: Construct a 0-2 hour forecast model, which includes a thunderstorm gale forecast model based on the SEResUNet deep neural network and a U 2 Net deep neural network and PhyDnet deep neural network short-term heavy rainfall forecast model; Constructing a 2-12 hour forecast model, wherein the 2-12 hour forecast model includes: a thunderstorm gale area forecast model and a short-term heavy rainfall area forecast model based on the StaRSU deep neural network; The pre-processed historical multi-source data is divided and input into a plurality of pre-built deep learning models for training, including: For the 0-2 hour forecast model, near-real-time data is extracted from the real-time observation data, and the 0-2 hour forecast model is trained by the near-real-time data; For the 2-12 hour forecast model, the forecast data of the CMA-GD model, the Doppler weather radar and the real-time observation data of the regional automatic station are used as input data, and the 2-12 hour forecast model is trained through multi-source feature fusion.
