Forecasting method and forecasting system for thunderstorm and gale

Abstract

The invention provides a weather forecasting method and system for thunderstorm and gale. The method comprises the following steps: acquiring multi-source weather observation data in a past preset time period; inputting the multi-source meteorological observation data into a pre-trained progressive wind speed forecasting model to extract and fuse multi-scale spatial-temporal characteristics in the multi-source meteorological observation data, and sequentially generating average wind speed forecasting results covering a plurality of future time periods; inputting the average wind speed forecast result into a pre-trained gust mapping model, and performing nonlinear mapping to obtain gust wind speed forecast results corresponding to a plurality of future time periods; and based on the gust speed forecast result, whether thunderstorm and gale risks exist in the future time period is judged. In the mode, the multi-source meteorological observation data is acquired and prediction is performed in combination with the progressive wind speed prediction model and the gust mapping model, so that the accuracy of wind speed and gust prediction can be improved, the thunderstorm and gale risk can be identified in advance, and the disaster early warning and disaster prevention and reduction capabilities are further improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of weather forecasting, in particular to a forecasting method and system for thunderstorm gale. BACKGROUND

[0002] Thunderstorm gale is an important disastrous weather that causes public safety and property loss due to its strong suddenness and wide influence range, so it is of great significance to make accurate short-term forecast for 0-2 hours in the future. Existing forecasting methods are generally divided into two categories: one is potential analysis based on environmental parameters, which is suitable for 6 hours or more of medium and long-term trend judgment, but it is difficult to capture the minute-level wind speed mutation; the other is near-term forecast based on radar echo pattern, which is suitable for 0-2 hours of short-term warning, but it relies on experience extrapolation and is difficult to reflect the storm explosive enhancement process. In recent years, although deep learning methods have been introduced, there are still deficiencies in the fusion of multi-source vertical observation data and the modeling of multi-period wind speed evolution. Therefore, the existing forecasting system has obvious limitations in accuracy and timeliness. SUMMARY

[0003] Therefore, the purpose of the present application is to provide a forecasting method and system for thunderstorm gale, which can improve the accuracy of short-term forecast for thunderstorm gale and the timeliness of effective forecast.

[0004] In a first aspect, the present application provides a weather forecasting method for thunderstorm gale, comprising: acquiring multi-source meteorological observation data in a past preset time period; the multi-source meteorological observation data at least includes parameters for characterizing vertical movement and thermal instability state; inputting the multi-source meteorological observation data into a pre-trained progressive wind speed forecasting model to extract and fuse multi-scale spatio-temporal features in the multi-source meteorological observation data, and sequentially generating average wind speed forecasting results covering multiple future time periods; the progressive wind speed forecasting model is generated based on an encoder-decoder hybrid architecture; inputting the average wind speed forecasting results into a pre-trained gust mapping model for nonlinear mapping to obtain gust wind speed forecasting results corresponding to the multiple future time periods; and judging whether there is a thunderstorm gale risk in the future time period based on the gust wind speed forecasting results.

[0005] In a second aspect, the present application provides a thunderstorm gale prediction system, comprising: a data acquisition module configured to acquire multi-source meteorological observation data in a past preset time period; the multi-source meteorological observation data at least comprises parameters for representing vertical movement and thermal instability state; an average wind speed prediction module configured to input the multi-source meteorological observation data into a pre-trained progressive wind speed prediction model to extract and fuse multi-scale spatio-temporal features in the multi-source meteorological observation data, and sequentially generate average wind speed prediction results covering multiple future time periods; the progressive wind speed prediction model is generated based on an encoder-decoder hybrid architecture; a gust wind speed prediction module configured to input the average wind speed prediction results into a pre-trained gust mapping model to perform nonlinear mapping, and obtain gust wind speed prediction results corresponding to the multiple future time periods; and a thunderstorm gale prediction module configured to determine whether there is a thunderstorm gale risk in the future time period based on the gust wind speed prediction results.

[0006] The present application provides a thunderstorm gale prediction method and system, which can effectively extract and fuse the evolution features of thunderstorm systems in multi-scale space-time by fusing multi-source observation data representing atmospheric vertical movement and thermal instability state, and using a progressive prediction model based on Swin Transformer and a nonlinear gust mapping model, thereby overcoming the defects of traditional methods that are difficult to analyze the explosive enhancement process of storms due to insufficient data resolution and model architecture limitations, and further significantly improving the accuracy and effective prediction time of thunderstorm gale short-term prediction.

[0007] Other features and advantages of the present application will be set forth in the following description, and in part will become apparent to those skilled in the art from the description, or can be learned by practice of the present application.

[0008] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the following preferred embodiments are described in detail below, and the accompanying drawings are described as follows. BRIEF DESCRIPTION OF DRAWINGS

[0009] In order to more clearly illustrate the specific embodiments of the present application or the technical solutions in the prior art, the following will briefly introduce the drawings needed to be used in the description of specific embodiments or prior art. Obviously, the drawings described below are some embodiments of the present application, and those skilled in the art can also obtain other drawings according to these drawings without creative labor.

[0010] Figure 1 The meteorological prediction method flowchart for thunderstorm gale provided by the embodiments of the present application; Figure 2 The multi-source meteorological observation data acquisition flowchart provided by the embodiments of the present application; Figure 3An architecture diagram of the G-net1H model provided for the embodiments of the present application; Figure 4 A progressive wind speed prediction model schematic diagram provided for the embodiments of the present application; Figure 5 A G-net1H model average wind speed prediction performance comparison result schematic diagram provided for the embodiments of the present application; Figure 6 A G-net2H model average wind speed prediction performance comparison result schematic diagram provided for the embodiments of the present application; Figure 7 A G-net1H model prediction error comparison result schematic diagram provided for the embodiments of the present application; Figure 8 An evaluation index comparison result schematic diagram in the process of mapping the average wind speed to the gust wind speed provided for the embodiments of the present application; Figure 9 A prediction result schematic diagram of nine thunderstorm gale cases provided for the embodiments of the present application; Figure 10 An observation and model prediction comparison result schematic diagram at Beijing time 13:00 provided for the embodiments of the present application; Figure 11 An observation and model prediction comparison result schematic diagram at Beijing time 14:00 provided for the embodiments of the present application; Figure 12 An one-hour and two-hour prediction evaluation index comparison result schematic diagram at Beijing time 12:00 initial time provided for the embodiments of the present application; Figure 13 An observation and model prediction comparison result schematic diagram at Beijing time 15:00 provided for the embodiments of the present application; Figure 14 An observation and model prediction comparison result schematic diagram at Beijing time 16:00 provided for the embodiments of the present application; Figure 15 An one-hour and two-hour prediction evaluation index comparison result schematic diagram at Beijing time 14:00 initial time provided for the embodiments of the present application; Figure 16 A prediction system schematic diagram for thunderstorm gale provided for the embodiments of the present application.

[0011] Icon: 1-data acquisition module; 2-average wind speed prediction module; 3-gust wind speed prediction module; 4-thunderstorm gale prediction module. DETAILED DESCRIPTION

[0012] To make the purposes, technical solutions, and advantages of the embodiments of the present application clearer, the technical solutions of the present application will be described below in connection with the drawings in a clear and complete manner. Obviously, the described embodiments are only some of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of protection of the present application.

[0013] To facilitate the understanding of the present embodiment, the application scenarios and design ideas of the present application will be briefly introduced below.

[0014] The current short-term forecast system of thunderstorm gale mainly follows the two-stage paradigm of "potential analysis - nowcasting". Potential forecast assesses the potential possibility of thunderstorm gale disaster by comprehensively analyzing the key environmental parameters such as the activity characteristics of dry air under weather background, the temperature lapse rate of the lower troposphere, K index, and convective available potential energy (CAPE). This prediction method can effectively identify the environmental conditions conducive to convective development 6-12 hours in advance, providing an important reference for subsequent nowcasting. Nowcasting relies on the morphological characteristics of radar products for prediction, among which the radar echo features such as mid-level radial convergence and bow echo have important indicative effects. In recent years, with the breakthrough progress of deep learning technology, this technology has provided an innovative solution to the key technical bottlenecks in the field of meteorological prediction, especially in the short-term forecast of thunderstorm gale. This kind of deep learning method adopts a data-driven paradigm and can automatically extract the nonlinear evolution characteristics of atmospheric systems from massive historical observation data through deep neural network architecture, effectively modeling the spatiotemporal dynamic evolution process of thunderstorm gale, and significantly improving the prediction robustness of traditional extrapolation methods when facing structural mutations of convective systems. It is worth noting that compared with traditional numerical weather prediction models, deep learning methods also have significant advantages in computational efficiency, and the improvement of inference speed greatly improves real-time business forecasting, providing more timely technical support for disaster prevention and mitigation decision-making.

[0015] Although the potential prediction method can effectively identify the environment background conducive to the development of convection 6-12 hours in advance, it is limited by the low temporal and spatial resolution of existing observation data, and it usually lacks the physical mechanism of the convection occurrence, development and evolution process, so there is still a lot of uncertainty in the prediction results. Among them, the prediction error of spatial position is particularly prominent. The actual prediction of the convection triggering area often deviates significantly from the real occurrence position, and with the extension of the prediction time, the spatial deviation shows a trend of continuous accumulation and intensification. In addition, the potential prediction mainly relies on the static judgment of environmental field indicators, and it is difficult to dynamically reflect the rapid evolution characteristics of the convection system. In contrast, the nowcasting method focuses more on using high temporal and spatial resolution real-time observation data such as radar echoes to make short-term prediction (usually no more than 6 hours) based on the current convection system structure and movement trend. This method has high prediction accuracy and practical value when the convection system structure is relatively stable and the observation data is continuous. However, the effective prediction time of the nowcasting method is short, and it is obviously insufficient for the prediction of rapidly changing convection system structure or newly born convection system in the future for a long period of time, and it is easy to miss or deviate significantly. In recent years, artificial intelligence technology represented by deep learning has made breakthroughs in computer vision, natural language processing and other fields, providing a new solution to the key technical bottlenecks in weather prediction, especially in the short-term prediction of thunderstorm gale. Although the current weather prediction method based on deep learning has made significant progress, there are still some limitations. First, some existing methods are mainly based on the modeling idea of static environmental parameter classification, that is, relying on historical meteorological observation data (such as radar reflectivity factor) to discretely classify and predict the future weather state (for example, whether a specific meteorological disaster will occur). This method is essentially a data-driven pattern recognition process, which pays more attention to the statistical mapping relationship between input and output. However, this method has two core deficiencies: one is the lack of explicit modeling of the nonlinear, time-varying and multi-scale coupled dynamics of the atmospheric system, which fails to effectively capture the continuous evolution characteristics of the weather system; the other is the lack of physical mechanism representation of weather phenomena, which lacks the explanatory power of dynamics and thermodynamics. These defects directly lead to poor performance in the continuity of time series prediction, the reliability of spatial extrapolation, and the generalization ability of extreme weather event prediction, especially when facing new or strong convection weather outside the distribution of training samples, the prediction accuracy often decreases significantly. Second, most of the existing research still relies too much on a single observation data source (such as radar reflectivity, ground wind speed or lightning observation data), ignoring the fusion potential of multi-source observation information, and has not yet established a dynamic interpretation framework that can effectively combine observation data and physical constraints. In addition, the short-term prediction based on the spatiotemporal extrapolation method still has obvious room for improvement in prediction timeliness and stability at this stage.These problems greatly limit the practicality and application promotion of deep learning methods in the short-term prediction of severe convective weather such as thunderstorm gale.

[0016] In general, the current research still faces two core challenges in the field of thunderstorm gale short-term prediction: 1) The existing methods lack the fusion of key parameters such as vertical velocity field and thermal dynamic trigger factor. In addition, thunderstorm gale is a typical mesoscale severe convective process with short life history (tens of minutes to hours), strong spatial locality and significant wind speed jump. Observation data shows that the disastrous wind speed increase (> 15m / s) of thunderstorm gale in the Beijing-Tianjin-Hebei region is mostly completed within 10-30 minutes, for example, in the "7·27" extreme gale event in Beijing in 2021, the wind speed increased from 17m / s to 32m / s within 12 minutes, and hourly resolution data is difficult to analyze the minute-level wind speed mutation characteristics. 2) The traditional model architecture has serious shortcomings in capturing the minute-level evolution characteristics. Previous research found that the time feature decay rate of the recurrent convolutional network at the 10-minute scale was as high as 40%, making it difficult for the system to capture the explosive enhancement process of thunderstorm cells.

[0017] Therefore, with reference to Figure 1 , the present application proposes a prediction method and system for thunderstorm gale, which can ensure that the input information has sufficient representativeness for the formation and development of thunderstorm gale by acquiring multi-source meteorological observation data including vertical motion and thermal instability parameters, thereby improving the basic accuracy of the prediction; the progressive wind speed prediction model based on the encoder-decoder hybrid architecture can extract and fuse multi-scale spatio-temporal features, and can generate average wind speed prediction results covering multiple future time periods in sequence, thereby realizing the gradual depiction of the wind speed evolution process; by inputting the average wind speed prediction results into the trained gust mapping model and establishing a nonlinear mapping relationship, the gust wind speed prediction results for multiple future time periods can be obtained, thereby improving the ability to capture instantaneous extreme wind speed; by judging whether there is a risk of thunderstorm gale based on the gust wind speed prediction results and combining a preset threshold, the early identification of thunderstorm gale can be realized, thereby improving the disaster warning capability of meteorological departments, and providing strong support for disaster prevention and mitigation and public safety.

[0018] After introducing the application scenario and design idea of the present application, the technical solutions provided by the present application will be described in detail.

[0019] The embodiment of the present application provides a prediction method for thunderstorm gale, with reference to Figure 1 , the steps of the prediction method for thunderstorm gale, comprising: Step S101, acquiring multi-source meteorological observation data in a past preset time period; the multi-source meteorological observation data at least includes parameters for representing vertical motion and thermal instability state.

[0020] Here, a timing task scheduling mechanism is adopted to realize the 24-hour x 10-minute interval automatic circulation prediction process through crontab (a schedule table). That is, the multi-source meteorological observation data is automatically acquired every 10 minutes to predict thunderstorm gales.

[0021] In order to comprehensively and accurately capture the three-dimensional atmospheric environmental characteristics before the occurrence of thunderstorm gales, data from different observation sources need to be fused.

[0022] In a preferred embodiment, the past preset time period can be 1 hour before the current time, and is sampled at a time resolution of 10 minutes to capture the dynamic evolution process of the weather. The multi-source meteorological observation data at least includes parameters for representing vertical movement and thermal instability. Specifically, the variable field with three-dimensional vertical structure provided by the Rapid Refresh Seamless Analysis and Prediction System (RMAPS- NOW) (such as wind speed, temperature, and humidity at different height layers), the ground observation information provided by the Rapid Refresh Seamless Analysis and Prediction System Real-Time Analysis and Prediction System (RMAPS-RISE), and radar combined reflectivity factor data can be fused. From these data, a series of physical quantity parameters directly or indirectly representing the atmospheric vertical movement and thermal instability state can be extracted, such as the vertical velocity field (Wnow), the convective available potential energy (CAPEave), and the temperature difference between the 850 hPa and 500 hPa height layers (FTEMPdiff).

[0023] In other feasible embodiments, the multi-source meteorological observation data can further fuse satellite observation data (such as cloud top brightness temperature for judging the degree of convective development), lightning positioning data (for identifying the convective core area), wind profile radar data or ground-based GPS remote sensing data (for obtaining high vertical resolution water vapor and wind field information), and the like. Similarly, the parameters for representing vertical movement and thermal instability state can also be other physical quantities familiar to those skilled in the art, such as K index, lifting index (LI), vertical wind shear (DWMAG), or storm relative helicity (SRH), and the like. The multi-source meteorological observation data needs to fuse as much observation information as possible that can reflect the three-dimensional dynamic, thermal, and water vapor conditions of the convective system.

[0024] In an embodiment, the step S101 comprises steps S201-S203.

[0025] Step S201, collecting meteorological related data from at least one data source.

[0026] Here, referring to Figure 2 , the data source includes: RMAPS-NOW (Rapidly-Updated Seamless Fusion and Integrated Prediction System - Nowcasting Analysis System): can provide multiple key meteorological variables with complete three-dimensional vertical structure.

[0027] RMAPS-RISE (Rapidly-Updated Seamless Fusion and Integrated Prediction System - Real-Time Surface Analysis System): can provide high-resolution two-dimensional surface meteorological variables.

[0028] RADAR (Radar): can provide radar combined reflectivity factor for identifying convective cores.

[0029] The meteorological related data obtained from at least one data source is taken as raw data. When a specific variable at time t has data missing, the observation record of the variable at time point t-10 minutes is automatically retrieved as a replacement value.

[0030] Step S202, extracting preset meteorological elements in meteorological related data; the preset meteorological elements are parameters related to the generation and development of thunderstorm gales.

[0031] Here, the preset meteorological elements are key physical quantity parameters closely related to the generation, development and evolution process of thunderstorm gales.

[0032] In a preferred embodiment, 16 meteorological characteristic elements are selected as preset meteorological elements. The preset meteorological element parameters are shown in Table 1 below.

[0033] Table 1: Preset meteorological element parameter table

[0034] Among them, the RMAPS-NOW data has a three-dimensional structure, and each variable corresponds to a specific vertical level. The RMAPS-RISE data and the RADAR data are two-dimensional structures.

[0035] Step S203, preprocessing the meteorological elements to obtain multi-source meteorological observation data; the preprocessing includes at least one of interpolation processing, normalization processing and cutting processing.

[0036] Here, the multi-source meteorological observation data is obtained by cleaning, standardizing and formatting the raw data conforming to the meteorological elements.

[0037] Referring to Figure 2 , the preprocessing includes: Interpolation processing (Linear Interpolation): the original RMAPS-NOW data with a spatial resolution of 3 kilometers can be down-scaled to 500 meters by linear interpolation method.

[0038] Data Quality Control: Quality inspection of data, elimination of outliers and missing values.

[0039] Min-Max Normalization: To eliminate the differences in the order of magnitude of different meteorological variables in the input data, the maximum and minimum normalization method is used for standardization. Among them, the maximum and minimum normalization is shown in formulas (1) and (2): (1) (2) Where represents the normalized value of the meteorological variable sample in the time step, represents the original value of the meteorological variable sample in the time step, represents the minimum value of the meteorological variable sample in the time step, represents the maximum value of the meteorological variable sample in the time step, represents the normalized entire sample.

[0040] Data Chunking: To reduce the cost of calculation and adapt to model input, the original large-scale sample (such as 1521x1221) is cut into multiple smaller sample blocks (such as 128x128) for model training. In actual prediction, the real-time large-scale data is also cut into small blocks, and the model is used to make individual predictions for each small block. Finally, all the prediction results of the small blocks are spliced together (inference construction) to form a complete prediction map covering the entire area.

[0041] Data Partitioning: All prepared data samples are divided into several independent subsets for model training and evaluation.

[0042] Through the above preprocessing methods, the Thunderstorm Gust Dataset, which is directly used for training and evaluation of thunderstorm gust data, can be obtained, that is, multi-source meteorological observation data.

[0043] Step S102, input the multi-source meteorological observation data into the pre-trained progressive wind speed prediction model to extract and fuse the multi-scale spatio-temporal features in the multi-source meteorological observation data, and sequentially generate average wind speed prediction results covering multiple future time periods; the progressive wind speed prediction model is generated based on an encoder-decoder hybrid architecture.

[0044] In a preferred embodiment, the progressive wind speed prediction model is based on an encoder-decoder hybrid architecture. Specifically, a U-Net type architecture (G-net) with Swin Transformer network as the backbone network can be adopted. The encoder part extracts hierarchical features of different spatial scales from the input multi-variable meteorological field layer by layer through the window self-attention mechanism and patch merging operation in the Swin Transformer. The decoder part gradually restores the spatial resolution through upsampling and skip connection, and finally generates a prediction map consistent with the input size.

[0045] Wherein, referring to Figure 3 and Figure 4 , in the progressive wind speed prediction model, referring to Figure 3 (b) the encoder with Swin Transformer network as the backbone network (Swin Transformer Block) is used to learn and extract multi-scale hierarchical features from the multi-source meteorological observation data; the decoder is used to generate average wind speed prediction results based on the hierarchical features extracted by the encoder. Referring to Figure 3 (c), the output end of the decoder includes a coordinate attention module, which is used to capture the spatial dependence of direction perception and position perception.

[0046] The progressive wind speed prediction model can be divided into a short-term wind speed prediction model (such as G-net1H, responsible for predicting 0-1 hours) and at least one extended wind speed prediction model (such as G-net2H, responsible for predicting 1-2 hours). When making a second hour prediction, the extended wind speed prediction model not only receives the original multi-source meteorological observation data, but also receives the first hour prediction result generated by the short-term wind speed prediction model as additional input through a feature bridging layer, thereby realizing cross-period information transmission and error correction, and effectively suppressing error accumulation.

[0047] wherein the G-net1H model takes as input reanalysis meteorological data obtained at 10-minute intervals in the past 1 hour, including 16 meteorological elements, some of which have a three-dimensional structure, and the total number of channels at a single time is 35. The dimension of the overall input data of a single sample is 6x35x128x128, wherein the time dimension and the variable dimension have been fused into the channel dimension before inputting the model, so the input channel number of the G-net1H model is 210, and the output data is 6x128x128, representing the average wind speed field of the future 1 hour at 10-minute intervals.

[0048] In other feasible embodiments, the encoder-decoder hybrid architecture can also be based on other advanced backbone networks, such as Vision Transformer (ViT), ResNet (Residual Network) or DenseNet (Densely Connected Convolutional Network) in CNN (Convolutional Neural Network), etc. The implementation of progressive prediction is not limited to the cascade of two independent models, but can be completed within a single model containing a loop or iterative mechanism. When predicting a later time step, the model can automatically use its previous time step prediction result as an internal state or additional input for iterative optimization. In addition, the progressive framework can also cascade more than two models, such as decomposing into 0-30 minutes, 30-60 minutes, 60-120 minutes, etc. for more detailed prediction stages to adapt to different business needs. The core is to decompose the long-time prediction task into multiple short-time tasks through a spatiotemporal decoupling and information feedback mechanism to improve the overall prediction performance.

[0049] Specifically, the G-net1H model, the G-net2H model and the G-netG model all adopt Swin Transformer as the backbone network for feature extraction, and an end-to-end prediction framework based on the encoder-decoder structure is constructed, similar to the structure of U-Net. Considering that the formation process of thunderstorm gale involves the nonlinear interaction of multiple meteorological variables (such as wind speed, air pressure, humidity, and convective available potential energy) at different spatial and temporal scales, the Patch Merging mechanism of Swin Transformer first divides the input meteorological field into multiple local windows, and then constructs a hierarchical feature representation by gradually merging adjacent windows, which is equivalent to the downsampling operation in the convolutional neural network. The encoder part of the model extracts spatial hierarchical features of multiple variables layer by layer through this kind of inverted pyramid structure, improving the representation ability of different scale weather systems. Within the Swin Transformer module, a local window-based self-attention mechanism is used to capture the significance of the features, which not only maintains efficient computational performance, but also models the complex spatial dependency between meteorological variables. Since there are nonlinear interactions between different variables in the thunderstorm gale system (such as temperature driving wind speed changes), and the spatial correlation between variables decays with distance, Swin Transformer efficiently extracts the dependency between local variables by calculating attention weights within local windows, and realizes information exchange across windows through window offset mechanism, thereby capturing long-distance dependent features. The encoder part contains 3 Swin Transformer layers that process features at different resolutions, each layer consists of two serial Swin Transformer modules, and down-sampling is achieved through Patch Merging between modules. The window size of each Swin Transformer module is set to 16, and experimental results show that this parameter setting achieves a good balance between prediction accuracy and computational efficiency. At the bottom of the encoder, 6 serial Swin Transformer modules are additionally introduced to further enhance the model's fitting ability for nonlinear relationships between meteorological variables. The decoder part gradually restores the spatial resolution through up-sampling operations, and finally realizes the output prediction map consistent with the input size. To further enhance the model's ability to extract key information, a Coordinate Attention (CA) mechanism is introduced at the output end of the Swin Transformer U-Net backbone structure. The coordinate attention mechanism can preserve accurate position information while introducing direction-aware channel attention, which helps the model focus more effectively on the regional features related to the target prediction in the spatial dimension, thereby improving the perception of local structures and edge details. The introduction of the coordinate attention module not only enhances the model's feature expression ability, but also improves the accuracy and robustness of the prediction results.

[0050] In an embodiment, the step of step S102 comprises steps S301-S302.

[0051] Step S301: inputting the multi-source meteorological observation data into a short-term wind speed prediction model to output a first average wind speed prediction result of a first future time period.

[0052] Here, the first future time period can be one hour in the future of the current time, for example, if the current time is T, then the first future time period is 0-1 hour.

[0053] Referring to Figure 4 , the short-term wind speed prediction model, namely the G-net1H model, takes multi-source meteorological observation data of a past period of time (for example, T-50 minutes to T minutes) as input to generate an average wind speed prediction of a first future time period (for example, T+10 minutes to T+60 minutes).

[0054] For the G-net1H model, the input is reanalysis meteorological data obtained at 10-minute intervals within the past 1 hour, which includes 16 meteorological elements, some of which have a three-dimensional structure, and the total number of channels at a single time is 35. The dimension of the overall input data of a single sample is 6x35x128x128, where the time dimension and the variable dimension are fused into the channel dimension before inputting the model, so the input channel number of G-net1H is 210, and the output data is 6x128x128, which respectively represents the average wind speed prediction field of the future 1 hour at 10-minute intervals. The specific architecture of the G-net1H model is shown in (a) of Figure 3 , where N represents the number of input data channels, and M represents the number of output data channels. In the G-net1H model, N=210 and M=6. In the G-net2H model, N=216 and M=6. In the G-netG model, N=1 and M=1.

[0055] Step S302: inputting the multi-source meteorological observation data and the first average wind speed prediction result into an extended wind speed prediction model to generate average wind speed prediction results of one or more subsequent future time periods after the first future time period; wherein for prediction of any subsequent future time period, one or more average wind speed prediction results before the subsequent future time period are taken as additional input, and are used together with the multi-source meteorological observation data to generate the average wind speed prediction result of the subsequent future time period.

[0056] Here, the input received by the extended wind speed forecasting model includes not only the original multi-source meteorological observation data, but also the relatively accurate forecasting result of the first future time period generated by the short-term wind speed forecasting model. Through one intermediate truncation and correction of the error propagation path, the extended wind speed forecasting model can extrapolate the prediction of a longer time period (e.g. 1-2 hours) based on the previously obtained average wind speed forecasting result.

[0057] In a preferred embodiment, with reference to Figure 4 , the extended wind speed forecasting model, i.e. the G-net2H model, realizes information fusion through a feature bridging layer (FBL). In addition to containing the same meteorological variables as G-net1H, the input of the G-net2H model additionally introduces the average wind speed information output by G-net1H, so the number of input channels is increased. The G-net2H model outputs the wind speed field every ten minutes in the second hour (e.g. T+70 minutes to T+120 minutes) in the future.

[0058] In the architecture, the G-net2H model introduces a feature bridging layer, in addition to containing the same meteorological variables as G-net1H, the input of the G-net2H model additionally introduces the average wind speed information output by G-net1H. Therefore, the number of input channels is increased to 216, and the size of the output data is 6x128x128, i.e. the wind speed field every ten minutes in the second hour in the future. Finally, the function of the G-netG model is to realize the nonlinear mapping between the average wind speed and the gust wind speed, and the input is only single-channel wind speed data, i.e. the number of input channels is 1, the size of the input data is 1x128x128, and the size of the output data is 1x128x128.

[0059] In another feasible embodiment, the progressive framework can include multiple extended wind speed forecasting models to realize longer time-extended forecasting. For example, a second extended wind speed forecasting model can be set up to forecast 2-3 hours. The input will include the original multi-source meteorological observation data, the average wind speed forecasting result of the first hour (from the short-term wind speed forecasting model), and the average wind speed forecasting result of the second hour (from the first extended wind speed forecasting model), and in this way, the forecasting time is further extended in a layer-by-layer progressive manner.

[0060] In an embodiment, the short-term wind speed forecasting model is trained in the following way: In step S401, historical meteorological observation data is taken as the first training input data, and the measured average wind speed of the first future time period corresponding to the historical meteorological observation data is taken as the first training label.

[0061] Here, the first training input data is multi-source meteorological observation data in the past 1 hour, every 10 minutes, in the historical data, as shown in Table 1.

[0062] The first training label is the measured average wind speed field in the future first time period (i.e. in the future 1 hour) corresponding to the first training input data.

[0063] Step S402, input the first training input data into the initial short-time wind speed prediction model, and output the first prediction result corresponding to the first training input data.

[0064] Here, the first training input data constructed is input into the G-net1H model to be trained, and a corresponding first prediction result (i.e. predicted future 1 hour average wind speed field) is generated.

[0065] Step S403, based on the preset loss function, the first prediction result and the first training label, the initial short-time wind speed prediction model is optimized until the preset training condition is met, and a trained short-time wind speed prediction model is obtained.

[0066] Here, the preset loss function is a weighted absolute error loss function (Weighted Absolute Error Loss Function, WAE).

[0067] The loss function in the average wind prediction is shown in formula (3), and the loss function in the gust prediction is shown in formula (4). By adjusting the sample weight distribution, the loss value of the high wind speed area is significantly enhanced, and the model pays more attention to the extraction and modeling of extreme wind speed characteristics in the optimization process. In the weight distribution process, the distribution characteristics of the average wind speed and the gust wind speed under different threshold conditions are considered to ensure that the distribution result of the loss weight and the actual proportion of each wind speed interval remain balanced and stable. The weight distribution is shown in Table 2.

[0068] (3) (4) Wherein represents the number of samples, represents the number of Y-axis grid points, represents the number of X-axis grid points, represents the weight allocated to different average wind speed and gust wind speed, represents the predicted average wind speed value of the (i, j) grid point, represents the observed average wind speed value of the (i, j) grid point, represents the predicted gust wind speed value of the (i, j) grid point, represents the observed gust wind speed value of the (i, j) grid point, represents the predicted average wind speed value of the (i, j) grid point, represents the observed average wind speed value of the (i, j) grid point, represents the predicted gust wind speed value of the (i, j) grid point, observed gust wind speed value of the grid.

[0069] Table 2 Weight distribution table of different wind speed intervals

[0070] Finally, the model parameters are iteratively optimized according to the calculated errors by an optimization algorithm (such as Adam (Adaptive Moment Estimation)) until the preset training conditions (for example, the validation set loss does not decrease for a plurality of consecutive periods or reaches the maximum number of training periods) are met, and a trained short-term wind speed prediction model is obtained.

[0071] Specifically, the spatial resolution of the input and output data is set to 128x128 to match the high-resolution characteristics of the meteorological field. To fully utilize the GPU memory capacity and improve the calculation efficiency, the batch size is optimized to 12. Adam is adopted, the initial learning rate is set to 0.001, and a dynamic adaptive adjustment mechanism is implemented based on the validation set performance: if the validation set loss does not decrease for 5 consecutive training periods, the learning rate decay strategy is triggered (the decay coefficient is 0.5). To prevent overfitting, an early stopping mechanism (Patience=10 epoch) is introduced during the training process, and the training is terminated in advance when the validation set loss does not improve for 10 consecutive periods. Through pre-experiment analysis, the upper limit of the time required for model convergence is set to 100 training periods. This parameter combination ensures the stability of the training while achieving a balance between computational resources and model performance.

[0072] In an embodiment, the extended wind speed prediction model is trained in the following manner: In step S501, the historical meteorological observation data and the average wind speed prediction result of the first future time period generated by the short-term wind speed prediction model are fused to obtain second training input data; the measured average wind speed of the second future time period corresponding to the second training input data is taken as the second training label; wherein the second future time period is later than the first future time period.

[0073] Here, the historical meteorological observation data and the average wind speed prediction result of the first future time period generated by the trained G-net1H model for the above historical meteorological observation data are fused to obtain the second training input data.

[0074] Referring to Figure 4, G-net2H model introduces a Feature Bridging Layer (FBL) in structure, which inputs include the same meteorological variables as G-net1H, and additionally introduces the average wind speed information output by G-net1H model. Therefore, the number of input channels increases to 216, and the size of output data is 6x128x128, i.e. the wind speed field every ten minutes in the second hour in the future.

[0075] Step S502, input the second training input data into the initial extended wind speed prediction model, and output the second prediction result corresponding to the second training input data.

[0076] Here, inputting the constructed second training input data into the G-net2H to be trained will generate a corresponding second prediction result.

[0077] Step S503, based on the preset loss function, the second prediction result and the second training label, the initial extended wind speed prediction model is optimized until the preset training condition is met, and the trained extended wind speed prediction model is obtained.

[0078] Here, similar to the training process of the short-term wind speed prediction model, a weighted absolute error loss function is also used to calculate the error between the second prediction result and the second training label.

[0079] Finally, the model is iteratively optimized by the optimization algorithm until the preset training condition is met, and the trained extended wind speed prediction model is obtained.

[0080] Step S103, input the average wind speed prediction result into the pre-trained gust mapping model, perform nonlinear mapping, and obtain the gust wind speed prediction result corresponding to multiple future time periods.

[0081] In a preferred embodiment, the gust mapping model can also use a deep learning network architecture similar to the progressive wind speed prediction model (such as G-netG). The gust mapping model learns the nonlinear mapping relationship between the average wind speed observation data at the historical whole point time and the gust wind speed observation data at the same time by using the average wind speed observation data at the historical whole point time as the training input and using the gust wind speed observation data at the same time as the training label.

[0082] In other possible embodiments, the gust mapping model is not limited to a specific network architecture. Any deep learning model that can effectively solve the image-to-image translation or regression problem can be applicable, such as a generative adversarial network (GAN), e.g. Pix2Pix-GAN, or other types of convolutional neural networks. In addition, the mapping process is not limited to a single variable mapping of average wind speed to gust wind speed. In a more complex embodiment, the gust mapping model can also receive other related meteorological forecast fields as auxiliary input, such as near-surface temperature field, humidity field or atmospheric stability parameter, etc., so as to establish a multivariate nonlinear mapping.

[0083] In an embodiment, the gust mapping model is trained by the following way: Step S601, the historical average wind speed is taken as the third training input data, and the measured gust wind speed corresponding to the historical average wind speed is taken as the third training label.

[0084] Here, since high spatiotemporal resolution gust observation data is difficult to obtain, the average wind speed observation value at the whole point time in the historical data is selected as the third training input data.

[0085] The measured gust wind speed observation value at the same time as the third training input data is selected as the third training label. The measured gust wind speed is derived from the RISE system (Rapid-refresh Integrated Ensemble System), which represents the maximum gust wind speed within 1 hour.

[0086] Step S602, a nonlinear mapping relationship between the historical average wind speed and the measured gust wind speed is established by the deep learning model to obtain an initial gust mapping model.

[0087] Here, the third training input data constructed is input into the G-netG model to be trained, which will generate the corresponding third prediction result (i.e. the predicted gust wind speed field). The mapping relationship based on the G-netG model is shown in formula (5).

[0088] (5) Wherein represents the average wind speed field at the whole point time, 1 represents the number of channels, H and W represent the length and width of the data respectively, represents a 1x1 convolution kernel, which aims to change the number of channels of the data. and represent the CA module and the Swin Transformer U-Net structure respectively, represents the final output gust wind speed field.

[0089] Step S603, input the third training input data into the initial gust mapping model, and output the third prediction result corresponding to the third training input data.

[0090] Step S604, based on the preset loss function, the third prediction result and the third training label, the initial gust mapping model is optimized until the preset training condition is met, and a trained gust mapping model is obtained.

[0091] Here, the error between the third prediction result and the third training label is calculated by the preset loss function. The preset loss function is a weighted absolute error loss function.

[0092] Finally, the model parameters are iteratively optimized according to the calculated error by the optimization algorithm until the preset training condition is met, and the trained gust mapping model is finally obtained.

[0093] Step S104, based on the gust wind speed prediction result, it is judged whether there is a thunderstorm gale risk in the future time period.

[0094] In a preferred embodiment, it can be realized by comparing the gust wind speed prediction result output by the gust mapping model with one or more preset gust wind speed thresholds. For example, according to the meteorological service specification, the area greater than 17.2 m / s can be judged as existing thunderstorm gale risk.

[0095] In other feasible embodiments, the preset gust wind speed threshold can not be a fixed value, but a threshold dynamically adjusted according to different geographical regions, seasons or underlying surface types (such as city, mountain, water body). In addition, the thunderstorm gale risk judgment can also be a comprehensive evaluation, not only considering the wind speed peak value, but also considering the area exceeding the threshold, the duration and other factors.

[0096] In an embodiment, to evaluate the performance of the proposed model in the short-term nowcasting of thunderstorm gale and the effectiveness of the progressive spatiotemporal decoupling framework constructed by the institute, three model variants are analyzed in a horizontal comparison. The performance of the model is evaluated by using multiple evaluation indexes commonly used in meteorological tests, including Critical Success Index (CSI), Probability of Detection (POD), False Alarm Ratio (FAR), Equitable Threat Score (ETS), Heidke Skill Score (HSS), and Dice Coefficient (DICE). The evaluation indexes are shown in formulas (6)-(14). The closer the CSI, POD, ETS, HSS, and DICE are to 1, the better the performance of the model, and the closer the FAR is to 0, the better the performance of the model. The above indexes are calculated based on the confusion matrix, and the calculation method of the confusion matrix is shown in Table 3. In addition, to further verify the consistency of the prediction results and the actual wind field, the Root Mean Square Error (RMSE) and the Mean Absolute Error (MAE) are introduced to quantitatively analyze the prediction error.

[0097] Firstly, the G-net1H is compared with a typical benchmark, Convolutional Long Short-Term Memory (ConvLSTM), in terms of 1-hour prediction results, and the related experimental results are shown in FIG. 2. Figure 5 Figure 5 The 0-1 hour average wind speed prediction performance of the G-net1H model and the ConvLSTM model under the wind speed threshold of 8.0 m / s, 10.8 m / s, 13.9 m / s, and 17.2 m / s is shown in FIG. 2. In the figure, the red solid line represents the G-net1H model, and the green solid line represents the ConvLSTM model.

[0098] The results show that the G-net1H model has significantly better prediction performance than the ConvLSTM in multiple evaluation indexes, especially in the Critical Success Index (CSI), the scores of all thresholds corresponding to each time step are significantly higher than those of the benchmark model. This result shows that the G-net1H model has a significant advantage in capturing the evolution of minute-level features, and can effectively make up for the lack of existing models in high-time-resolution wind speed prediction.

[0099] ​Further, the G-net2H (With FBL) is compared with the G-net2H (Without FBL) and the ConvLSTM, aiming to verify the effectiveness of the proposed progressive spatio-temporal coupling architecture in improving the extrapolation ability of the model. The experimental results are shown in Figure 6 . Among them, Figure 6 is the comparison of the 1-2 hour average wind speed prediction performance of the G-net2H model, the G-net2H (remove FBL) model and the ConvLSTM model under the wind speed threshold of 8.0 m / s, 10.8 m / s, 13.9 m / s and 17.2 m / s. The red solid line in the figure represents the G-net2H model, the blue solid line represents the G-net2H (remove FBL) model, and the green solid line represents the ConvLSTM model.

[0100] The results show that after introducing the FBL, the model has achieved significant improvement in various evaluation indicators. In addition, based on the error comparison analysis in Figure 7 , it can be further found that the progressive spatio-temporal coupling architecture effectively alleviates the trend of error accumulation in the prediction process, thereby significantly enhancing the extrapolation ability and timeliness of the model. Among them, Figure 7 shows the prediction error comparison of the G-net1H model and the ConvLSTM model in the 0-2 hour average wind speed prediction. The error indicators include MAE and RMSE. The purple solid line in the figure represents the G-net1H model, the dark blue solid line represents the G-net2H model without the feature bridge layer (G-net2H Without FBL), the red solid line represents the G-net2H model containing the feature bridge layer (G-net2H With FBL), and the green solid line represents the ConvLSTM model. The mutation phenomenon appearing between 60 minutes and 70 minutes in the figure corresponds to the introduction time of the feature bridge layer.

[0101] Finally, the mapping performance from the surface average wind speed to the gust wind speed is compared and analyzed to verify the effectiveness of the deep learning-based gust mapping model G-netG. In the evaluation process, the currently commonly used gust coefficient model (a traditional mapping method based on linear regression) and two end-to-end deep learning models are selected as the comparison baseline to carry out horizontal performance comparison. The experimental results are shown in Figure 8 . Among them, Figure 8 shows the evaluation index comparison of the four models in the process of mapping the surface average wind speed to the gust wind speed. The column chart uses four different colors to represent each model, from left to right, G-netG, U-Net, Pixel2Pixel-GAN and the traditional gust coefficient model.

[0102] The experimental results show that the G-netG model achieves greater improvement at multiple wind speed thresholds, indicating that the gust mapping model based on deep learning can effectively extract the nonlinear complex relationship between the gust wind speed and the average ground wind speed, further improve the mapping accuracy, and based on this, the embodiment of the present application can effectively solve the problem that the thunderstorm gale cannot be accurately predicted due to the lack of time resolution of observation data.

[0103] To verify the performance of the embodiment of the present application, the Rapid Refresh Seamless Ensemble and Integrated Forecasting System (RISE) is selected as a comparative baseline for horizontal comparison analysis. In terms of evaluation indicators, the critical success index (CSI) and bias score (BIAS) are used to quantitatively test the final thunderstorm gale prediction effect, and data from May to August 2024 are selected as the final test set for evaluation, and the experimental results are shown in Figure 9 , and two typical thunderstorm gale cases on May 30, 2024 are also selected for qualitative and quantitative analysis, and the experimental results are shown in Figures 10-15 . Figure 9 The light blue column represents the prediction results of the model provided by the embodiment of the present application, and the orange column represents the prediction results of the GustFC model. As can be seen from the performance evaluation indicators of the test set shown in Figure 9 , the model in the embodiment of the present application has achieved significant improvement in the critical success index (CSI) compared with the RISE benchmark model, indicating that the model has high accuracy in the spatio-temporal positioning of thunderstorm gale prediction. For the two typical thunderstorm gale cases on May 30, 2024, the evaluation results show that the embodiment of the present application performs particularly outstanding in the capture ability of mesoscale convective systems. Compared with the RISE system, the present application can accurately identify a small convective storm system initially in Zhangjiakou City, Hebei Province, and effectively predict the subsequent moving path and position information of the system, while the RISE system has obvious false negatives. The above results fully prove the real-time and accuracy advantages of the embodiment of the present application in the short-term nowcasting of thunderstorm gale.

[0104] wherein, Figure 10 is the observation and model prediction comparison of the thunderstorm gale in the Beijing-Tianjin-Hebei region at 13:00 on May 30, 2024. Figure 10 (a) in the figure is the real-time wind field at 13:00 Beijing time. Figure 10 (b-c) in the figure is the one-hour prediction wind field with 12:00 as the initial time: (b) is the one-hour prediction result of the embodiment of the present application; (c) is the one-hour prediction result of the GustFC model. The deep red area in the figure represents the initial position of the thunderstorm gale observation and prediction.

[0105] Figure 11Comparison of observations and model forecasts for thunderstorms and strong winds in the Beijing-Tianjin-Hebei region at 14:00 Beijing time on May 30, 2024. Figure 11 (a) The actual wind field at 14:00 Beijing time; Figure 10 (b–c) shows the two-hour forecast wind field with 12:00 as the initial time: (b) is the two-hour forecast result of the embodiment of this application; (c) is the two-hour forecast result of the GustFC model. The dark red area in the figure represents the observed and forecasted impact range of thunderstorm winds. The observation results show that the thunderstorm wind process is gradually intensifying, and its impact range is further expanding towards the urban area of ​​Beijing.

[0106] Figure 12 A comparison of the evaluation indicators of one-hour and two-hour forecasts for different models at the initial time of 12:00 Beijing time on May 30, 2024.

[0107] Figure 13 Comparison of observations and model forecasts for thunderstorms and strong winds in the Beijing-Tianjin-Hebei region at 15:00 Beijing time on May 30, 2024. Figure 13 (a) in the image shows the actual wind field at 15:00 Beijing time; Figure 13 (b–c) shows the one-hour wind field forecast with 14:00 as the initial time: (b) is the forecast result of the embodiment of this application; (c) is the forecast result of GustFC in the RISE system. The dark red box in the figure represents the area affected by the strong storm system at 15:00, which has now fully entered the urban area of ​​Beijing and continues to move in a southeast direction.

[0108] Figure 14 Comparison of observations and model forecasts for thunderstorms and strong winds in the Beijing-Tianjin-Hebei region at 16:00 Beijing time on May 30, 2024. Figure 14 (a) in the image shows the actual wind field at 16:00 Beijing time; Figure 14 The two-hour forecast wind field (b–c) with 14:00 as the initial time is: Figure 14 (b) in the figure shows the forecast results of the embodiment of this application; (c) shows the forecast results of GustFC in the RISE system. The dark red box in the figure marks the area where the strong wind system affected the northwest of Tianjin at 16:00, forming a southwest-northeast trending gust line.

[0109] Figure 15 The evaluation indicators of the 1-hour and 2-hour forecasts of each model are compared at 14:00 Beijing time on May 30, 2024.

[0110] (6) wherein, CSI (Critical Success Index) measures the proportion of all events in which thunderstorm gale is forecasted or observed, which are correctly forecasted. This index penalizes both false negatives (FN) and false positives (FP), so the higher the score, the stronger the comprehensive forecasting ability of the model. TP (True Positive) represents the number of hits or correct forecasts. That is, the forecast result is thunderstorm gale, and the ground truth also actually occurs thunderstorm gale. FP (False Positive) represents the number of false alarms or false forecasts. That is, the forecast result is thunderstorm gale, but the ground truth does not occur thunderstorm gale. FN (False Negative) represents the number of false negatives. That is, the forecast result is no thunderstorm gale, but the ground truth actually occurs thunderstorm gale.

[0111] (7) wherein, POD (Probability of Detection) measures the proportion of all actual thunderstorm gale events that the model successfully forecasts. This index only focuses on false negatives (FN), and a high POD value means that the model rarely misses actual events.

[0112] (8) wherein, FAR (False Alarm Ratio) measures the proportion of all events that the model forecasts as thunderstorm gale, which are false (i.e. false alarms). This index only focuses on false positives (FP), and a low FAR value means that the model's forecast is highly reliable.

[0113] (9) wherein, TN (True Negative) represents the number of correct no forecasts. That is, the forecast result is no thunderstorm gale, and the ground truth also does not actually occur thunderstorm gale.

[0114] (10) wherein, ETS (Equitable Threat Score) is an improvement on CSI. Based on CSI, an expected number of hits E generated by random forecasting is subtracted, so as to more fairly measure the forecasting skill of the model relative to pure random guessing. E represents the number of times that may be hit by chance in the case of random forecasting, in order to eliminate the skill brought by random forecasting from the forecast score, so as to more fairly evaluate the real ability of the forecasting model.

[0115] (11) where HSS (Heidke Skill Score) measures the improvement of the prediction accuracy of the model relative to a random prediction. The score range is usually between -1 and 1, 1 represents perfect prediction, 0 represents the same level as random prediction, and negative value indicates worse than random prediction.

[0116] (12) where BIAS (Bias Score) measures the systematic bias of the prediction system. By calculating the ratio of the total number of times the model predicts thunderstorm gale to the total number of times the actual thunderstorm gale occurs, it can be determined whether the model has a tendency to overestimate or underestimate the frequency of event occurrence.

[0117] (13) where MAE (Mean Absolute Error) calculates the average of the absolute value of the difference between the predicted value and the true value at all forecast grid points. N is the total number of samples. i and j represent the grid index of Y-axis and X-axis in the two-dimensional forecast grid, respectively. represents the average wind speed value predicted at (i, j) grid point in the nth sample. represents the average wind speed value observed at (i, j) grid point in the nth sample.

[0118] (14) where RMSE (Root Mean Square Error) calculates the square root of the average of the square of the difference between the predicted value and the true value. Compared with MAE, RMSE gives higher weight to larger errors (outliers), and it can better reflect whether there is extreme deviation in the prediction result. represents the gust wind speed value predicted at (i, j) grid point in the nth sample. represents the gust wind speed value observed at (i, j) grid point in the nth sample.

[0119] Table 3 Confusion matrix

[0120] The meteorological forecasting method for thunderstorm gale provided by the embodiments of the present application can guarantee that the input information can fully represent the formation and development conditions of thunderstorm gale, thereby improving the representativeness and reliability of the input data, by acquiring and preprocessing multi-source meteorological observation data containing vertical movement and thermal instability parameters; the gradual wind speed forecasting model based on the encoder-decoder architecture can be used in combination with the step-by-step iteration mechanism of the short-term and extended wind speed forecasting models, so as to continuously generate average wind speed forecasting results in multiple future time periods, thereby significantly improving the accuracy of the description of the wind speed evolution process; the nonlinear mapping relationship between the average wind speed and the gust wind speed can be established by introducing the gust mapping model, so as to effectively capture the instantaneous extreme wind speed characteristics, thereby improving the identification accuracy of the thunderstorm gale risk; the multi-scale feature extraction network and the coordinate attention module can be used in the model, so as to further enhance the capturing ability of the spatial dependence relationship, thereby improving the spatial resolution and overall accuracy of the prediction results; further, the present application can realize the early identification and accurate early warning of the thunderstorm gale risk, thereby significantly enhancing the prevention and response capability of meteorological disasters.

[0121] The embodiments of the present application provide a human-computer interface device login system, referring to Figure 16 , the forecasting system for thunderstorm gale, comprising: The data acquisition module 1 is used for acquiring multi-source meteorological observation data in a past preset time period; the multi-source meteorological observation data at least includes parameters for representing the vertical movement and the thermal instability state.

[0122] The average wind speed forecasting module 2 is used for inputting the multi-source meteorological observation data into a pre-trained gradual wind speed forecasting model, so as to extract and fuse the multi-scale spatiotemporal features in the multi-source meteorological observation data, and sequentially generate average wind speed forecasting results covering multiple future time periods; the gradual wind speed forecasting model is generated based on the encoder-decoder hybrid architecture.

[0123] The gust wind speed forecasting module 3 is used for inputting the average wind speed forecasting result into a pre-trained gust mapping model, performing nonlinear mapping, and obtaining gust wind speed forecasting results corresponding to multiple future time periods.

[0124] The thunderstorm gale forecasting module 4 is used for judging whether there is a thunderstorm gale risk in the future time period based on the gust wind speed forecasting result.

[0125] In an optional implementation, the gradual wind speed forecasting model includes a short-term wind speed forecasting model and at least one extended wind speed forecasting model. The average wind speed forecasting module 2 is further used for: inputting the multi-source meteorological observation data into the short-term wind speed forecasting model, and outputting a first average wind speed forecasting result of a first future time period.

[0126] inputting the multi-source meteorological observation data and the first average wind speed prediction result into the extended wind speed prediction model to generate average wind speed prediction results for one or more subsequent future time periods after the first future time period; wherein for any subsequent future time period, one or more average wind speed prediction results before the subsequent future time period are taken as additional inputs, together with the multi-source meteorological observation data, to generate the average wind speed prediction result for the subsequent future time period.

[0127] In an optional embodiment, the data acquisition module 1 is further configured to: acquire meteorological related data from at least one data source.

[0128] extract preset meteorological elements from the meteorological related data; the preset meteorological elements are parameters related to the generation and development of thunderstorm gales.

[0129] perform preprocessing on the meteorological elements to obtain the multi-source meteorological observation data; the preprocessing includes at least one of interpolation processing, normalization processing and cutting processing.

[0130] In an optional embodiment, the progressive wind speed prediction model includes a short-time wind speed prediction model. The average wind speed prediction module 2 is further configured to: take historical meteorological observation data as first training input data, and take measured average wind speed of a first future time period corresponding to the historical meteorological observation data as a first training label.

[0131] input the first training input data into an initial short-time wind speed prediction model to output a first prediction result corresponding to the first training input data.

[0132] optimize the initial short-time wind speed prediction model based on a preset loss function, the first prediction result and the first training label until a preset training condition is met, to obtain a trained short-time wind speed prediction model.

[0133] In an optional embodiment, the progressive wind speed prediction model includes a short-time wind speed prediction model and an extended wind speed prediction model. The average wind speed prediction module 2 is further configured to: fuse the historical meteorological observation data and the average wind speed prediction result of the first future time period generated by the short-time wind speed prediction model to obtain second training input data; take measured average wind speed of a second future time period corresponding to the second training input data as a second training label; wherein the second future time period is later than the first future time period.

[0134] input the second training input data into an initial extended wind speed prediction model to output a second prediction result corresponding to the second training input data.

[0135] The initial extended wind speed prediction model is optimized based on a preset loss function, the second prediction result and the second training label until a preset training condition is met, and a trained extended wind speed prediction model is obtained.

[0136] In an optional embodiment, the gust wind speed prediction module 3 is further configured to: The historical average wind speed is taken as the third training input data, and the measured gust wind speed corresponding to the historical average wind speed is taken as the third training label.

[0137] A nonlinear mapping relationship between the historical average wind speed data and the measured gust wind speed is established by using the deep learning model, and an initial gust mapping model is obtained.

[0138] The third training input data is input into the initial gust mapping model, and a third prediction result corresponding to the third training input data is output.

[0139] The initial gust mapping model is optimized based on a preset loss function, the third prediction result and the third training label until a preset training condition is met, and a trained gust mapping model is obtained.

[0140] In an optional embodiment, the preset loss function is a weighted absolute error loss function.

[0141] In an optional embodiment, in the progressive wind speed prediction model, the encoder takes a Swin Transformer network as a backbone network, and is configured to learn and extract multi-scale hierarchical features from the multi-source meteorological observation data; and the decoder is configured to generate the average wind speed prediction result based on the hierarchical features extracted by the encoder.

[0142] In an optional embodiment, the output end of the decoder includes a coordinate attention module, and the coordinate attention module is configured to capture spatial dependency of direction perception and position perception.

[0143] The human-computer interface device login system provided by the embodiments of the present application can improve the accuracy and real-time performance of thunderstorm gale prediction through multi-source data acquisition, wind speed and gust joint prediction and risk judgment, so as to realize early identification of disaster risk and further enhance the ability of disaster prevention and reduction.

[0144] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the system and device described above can refer to the corresponding process in the foregoing method embodiments, which will not be described herein.

[0145] In addition, in the description of the embodiments of the present application, unless otherwise explicitly specified and limited, the terms "mounting", "connection", "connecting" should be understood in a broad sense, for example, can be fixedly connected, or can be detachably connected, or integrally connected; can be mechanically connected, or can be electrically connected; can be directly connected, or can be indirectly connected through an intermediate medium; can be the communication inside two elements. For those skilled in the art, the specific meanings of the above terms in the present application can be understood according to the specific circumstances.

[0146] In the description of the present application, it should be noted that the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer" and the like indicate the orientation or positional relationship shown in the drawings, and are only for the purpose of facilitating the description of the present application and simplifying the description, and do not indicate or imply that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the present application. In addition, the terms "first", "second", "third" are only for the purpose of description, and cannot be understood as indicating or implying relative importance.

[0147] Finally, it should be noted that: the above-described embodiments are only specific embodiments of the present application, used to illustrate the technical solutions of the present application, and are not limited thereto, the protection scope of the present application is not limited thereto, although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any skilled person familiar with the technical field can modify or easily think of changes to the technical solutions recorded in the foregoing embodiments within the technical range disclosed by the present application, or make equivalent replacement to part of the technical features; and these modifications, changes or replacements do not make the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be covered within the protection scope of the present application.

Claims

1. A weather forecast method for thunderstorm gales, characterized in that, The method comprises: obtaining multi-source meteorological observation data in a preset time period in the past; the multi-source meteorological observation data at least includes parameters for characterizing vertical movement and thermal instability; inputting the multi-source meteorological observation data into a pre-trained progressive wind speed prediction model to extract and fuse multi-scale spatio-temporal features in the multi-source meteorological observation data, and sequentially generating average wind speed prediction results covering multiple future time periods; the progressive wind speed prediction model is generated based on an encoder-decoder hybrid architecture; inputting the average wind speed prediction results into a pre-trained gust mapping model for nonlinear mapping to obtain gust wind speed prediction results corresponding to the multiple future time periods; based on the gust wind speed prediction results, it is determined whether there is a thunderstorm gale risk in the future time period.

2. The weather forecast method for thunderstorm gale according to claim 1, characterized by, The progressive wind speed prediction model comprises a short-term wind speed prediction model and at least one extended wind speed prediction model; the step of inputting the multi-source meteorological observation data into a pre-trained progressive wind speed prediction model to extract and fuse multi-scale spatio-temporal features in the multi-source meteorological observation data, and sequentially generating average wind speed prediction results covering multiple future time periods, comprises: inputting the multi-source meteorological observation data into the short-term wind speed prediction model to output the first average wind speed prediction result of the first future time period; inputting the multi-source meteorological observation data and the first average wind speed prediction result into the extended wind speed prediction model to generate average wind speed prediction results of one or more subsequent future time periods after the first future time period; wherein for any subsequent future time period, one or more average wind speed prediction results before the subsequent future time period are taken as additional input, which are used together with the multi-source meteorological observation data to generate the average wind speed prediction result of the subsequent future time period.

3. The weather forecast method for thunderstorm gale according to claim 1, characterized by, The step of obtaining multi-source meteorological observation data in a preset time period in the past comprises: collecting meteorological data from at least one data source; extracting preset meteorological elements from the meteorological data; the preset meteorological elements are parameters related to the generation and development of thunderstorm gales; preprocessing the meteorological elements to obtain the multi-source meteorological observation data; the preprocessing includes at least one of interpolation processing, normalization processing and cutting processing.

4. The weather forecast method for thunderstorm gale according to claim 1, characterized by, The progressive wind speed prediction model comprises a short-term wind speed prediction model; the short-term wind speed prediction model is trained by: taking historical meteorological observation data as first training input data, and taking the measured average wind speed of the first future time period corresponding to the historical meteorological observation data as the first training label; inputting the first training input data into an initial short-term wind speed prediction model to output the first prediction result corresponding to the first training input data; based on a preset loss function, the first prediction result and the first training label, the initial short-term wind speed prediction model is optimized until a preset training condition is met, and the trained short-term wind speed prediction model is obtained.

5. The weather forecast method for thunderstorm gale according to claim 1, characterized in that, The progressive wind speed prediction model comprises a short-term wind speed prediction model and an extended wind speed prediction model; the extended wind speed prediction model is trained by: The historical meteorological observation data and the average wind speed prediction result of a first future time period generated by the short-time wind speed prediction model are fused to obtain second training input data; a measured average wind speed of a second future time period corresponding to the second training input data is taken as a second training label; wherein the second future time period is later than the first future time period; The second training input data is input into an initial extended wind speed prediction model to output a second prediction result corresponding to the second training input data; The initial extended wind speed prediction model is optimized based on a preset loss function, the second prediction result and the second training label until a preset training condition is met, so as to obtain the trained extended wind speed prediction model.

6. The weather forecast method for thunderstorm gale according to claim 1, characterized by, The gust mapping model is trained in the following manner: The historical average wind speed is taken as third training input data, and a measured gust wind speed corresponding to the historical average wind speed is taken as a third training label; A nonlinear mapping relationship between the historical average wind speed data and the measured gust wind speed is established by a deep learning model to obtain an initial gust mapping model; The third training input data is input into the initial gust mapping model to output a third prediction result corresponding to the third training input data; The initial gust mapping model is optimized based on a preset loss function, the third prediction result and the third training label until a preset training condition is met, so as to obtain the trained gust mapping model.

7. The meteorological method for predicting thunderstorm gales according to any of claims 4 or 5 or 6, characterized in that, The preset loss function is a weighted absolute error loss function.

8. The weather forecast method for thunderstorm gale according to claim 1, characterized by, In the progressive wind speed prediction model, an encoder takes a Swin Transformer network as a backbone network, and is used to learn and extract multi-scale hierarchical features from the multi-source meteorological observation data; A decoder is used to generate the average wind speed prediction result based on the hierarchical features extracted by the encoder.

9. The weather forecast method for thunderstorm gale according to claim 8, characterized in that, An output end of the decoder includes a coordinate attention module, which is used to capture spatial dependence of direction perception and position perception.

10. A forecasting system for thunderstorm gales, characterized in that, It comprises: A data acquisition module is configured to acquire multi-source meteorological observation data in a past preset time period; The multi-source meteorological observation data at least includes parameters for representing vertical motion and thermal instability state; An average wind speed prediction module is configured to input the multi-source meteorological observation data into a pre-trained progressive wind speed prediction model to extract and fuse multi-scale spatio-temporal features in the multi-source meteorological observation data, and sequentially generate average wind speed prediction results covering multiple future time periods; The progressive wind speed prediction model is generated based on an encoder-decoder hybrid architecture; A gust wind speed prediction module is configured to input the average wind speed prediction result into a pre-trained gust mapping model to perform nonlinear mapping, so as to obtain gust wind speed prediction results corresponding to the multiple future time periods; A thunderstorm gale prediction module is configured to determine whether there is a thunderstorm gale risk in a future time period based on the gust wind speed prediction result.

Images