Short-term and temporary disaster weather forecast pushing method, system, equipment and medium

Abstract

The invention relates to a short-term and temporary disaster weather forecast pushing method, system and device and a medium. The method comprises the steps that meteorological radar echo intensity data and wind and rain data of a ground observation station are collected, the radar echo intensity data are grouped through a Marching Square algorithm, and longitude and latitude contour lines are generated; after the groups are effectively filtered, wind and rain signals of the attribution are calculated, a wind and rain signal time sequence is constructed, strong weather starting and ending time is judged, and strong weather marking signals are generated; based on the strong weather mark signal, the radar echo intensity data and the effective group set, a fusion forecast result graph is obtained through AI prediction, optical flow extrapolation prediction and Kalman filtering fusion processing; and based on the graph and the latitude and longitude contour lines, calculating a rainfall rate and a maximum accumulated rainfall value through rainfall intensity conversion processing, and generating an early warning instruction in combination with a preset grading threshold value. According to the method, by means of AI prediction, optical flow extrapolation processing and the like, the accuracy and timeliness of short-term disaster weather forecast are improved.

Description

Technical Field

[0001] This invention belongs to the field of computer technology, and in particular relates to a method, system, device and medium for pushing short-term disaster weather forecasts. Background Technology

[0002] Currently, short-term extreme weather forecasts mainly rely on real-time data from weather radar and ground observation stations. However, both radar and automatic observation station data often suffer from missing data and anomalies. Furthermore, relying on a single set of data for extreme weather forecasting makes it difficult to achieve automated and accurate identification of severe weather. At the same time, existing short-term forecasting technologies each have their advantages and disadvantages. For example, numerical models in physical short-term forecasting methods are computationally time-consuming and their results are not accurate enough within the 0-1 hour forecast interval. Short-term forecasting methods based on optical flow extrapolation technology struggle to effectively reproduce the evolution characteristics of storm cells. Moreover, the forecasting effectiveness of AI-based short-term forecasting methods is highly dependent on the quality of training data and generally suffers from intensity decay and a decline in forecast accuracy over time. Furthermore, current early warning and forecasting services mostly require forecasters to manually issue warnings based on algorithm results and professional knowledge. However, extreme weather is characterized by strong locality and rapid occurrence and development. Different locations within the same county or district are affected differently by the same extreme weather. Manual forecasting makes it difficult to provide localized and customized services. Moreover, due to manpower limitations, severe weather prevention services can only cover the entire city or county level, which cannot meet the needs of specific sectors such as agriculture, fisheries, and tourism for precise and customized meteorological services. Summary of the Invention

[0003] Therefore, it is necessary to provide a method, system, equipment, and medium for short-term hazardous weather forecasting to address the aforementioned technical issues, aiming to improve the automated and accurate identification of extreme weather and enhance the accuracy and timeliness of forecasts.

[0004] Firstly, this application provides a method for delivering short-term hazardous weather forecasts, including:

[0005] Meteorological radar echo intensity data and wind and rain data from ground observation stations are collected. The radar echo intensity data is grouped using the Marching Square algorithm to obtain multiple groups, and the latitude and longitude contour lines corresponding to each group are generated. The validity of each group is filtered to generate a set of valid groups.

[0006] Based on the effective group set, calculate the local wind and rain signal, construct the wind and rain signal time series, determine the start and end time of severe weather according to the wind and rain signal time series, and generate severe weather marker signal;

[0007] Based on strong weather marker signals and radar echo intensity data, weather forecasting is performed to obtain AI forecast results; optical flow extrapolation is performed based on the effective group set to obtain optical flow forecast results; the AI ​​forecast results and optical flow forecast results are combined and fused through Kalman filtering to obtain a fused forecast result map.

[0008] Based on the fused forecast results map and latitude and longitude contour lines, the precipitation rate is obtained by converting the precipitation intensity. The maximum cumulative precipitation value is calculated based on the precipitation rate, and a warning instruction is generated by combining the preset graded thresholds.

[0009] In one embodiment, the radar echo intensity data is grouped using the Marching Square algorithm to obtain multiple groups, and latitude and longitude contour lines corresponding to each group are generated; each group is then filtered for validity to generate a set of valid groups, including:

[0010] The radar echo intensity data is grouped according to a preset echo intensity threshold sequence using the Marching Square algorithm to obtain multiple groups;

[0011] Each group is processed for contour extraction to generate the latitude and longitude contour lines corresponding to each group.

[0012] Extract the unit-time precipitation data of each group of associated ground stations from the wind and rain data of the ground observation stations;

[0013] Each group is filtered for validity, and groups with maximum echo intensity greater than the first preset echo threshold and precipitation data per unit time greater than the first preset precipitation threshold are retained. A set of valid groups is generated, which includes the contour coordinate set, internal observation station list, maximum echo intensity and maximum precipitation per unit time information of each valid group.

[0014] In one embodiment, a local wind and rain signal is calculated based on the effective group set, a wind and rain signal time series is constructed, the start and end times of severe weather are determined based on the wind and rain signal time series, and a severe weather marker signal is generated, including:

[0015] Based on the effective group set, the maximum precipitation information per unit time of each effective group in the target area is extracted for each collection time. The maximum value of the maximum precipitation information per unit time of each effective group in the target area is used as the local wind and rain signal at the corresponding collection time.

[0016] The local wind and rain signals corresponding to multiple acquisition times are arranged in chronological order and processed to construct a time series, resulting in a wind and rain signal time series with a preset time resolution.

[0017] The severe weather initiation determination process is performed based on the wind and rain signal time series. When the wind and rain signal in the first continuous preset duration wind and rain signal time series is greater than or equal to the second preset precipitation threshold, a severe weather initiation marker is generated.

[0018] The severe weather termination determination process is performed based on the wind and rain signal time series. When the wind and rain signal in the second consecutive preset duration wind and rain signal time series is equal to the third preset precipitation threshold, a severe weather termination mark is generated.

[0019] The severe weather start marker and severe weather end marker are integrated and processed to generate a severe weather marker signal that includes whether the current situation is a severe weather process, the start time, and the end time.

[0020] In one embodiment, weather forecasting is performed based on strong weather marker signals and radar echo intensity data to obtain AI prediction results, including:

[0021] The severe weather marker signal is processed for state recognition. When the severe weather marker signal is in an active state, the preset frame data of radar echo intensity data is acquired.

[0022] Using preset frame data as input data, the Swim-Transformer is used to perform multi-scale spatial feature extraction processing on the input data to obtain multi-scale spatial features;

[0023] Semantic features are extracted from multi-scale spatial features using a feature extraction network based on the U-Net architecture to obtain high-dimensional feature maps.

[0024] A dual-branch output model is constructed based on meteorological physical equations. The meteorological physical equations consist of local variation terms and advection-induced intensity variation terms. The local variation terms are used to reflect the local changes in radar echo intensity over time, while the advection-induced intensity variation terms are used to reflect the changes in radar echo intensity caused by horizontal motion.

[0025] The high-dimensional feature map is processed by the first and second branches of the dual-branch output model to obtain the corresponding advection term learning results and local term learning results.

[0026] The MOE weighted architecture is used to fuse the learning results of the advection term and the local term to obtain the model fusion features;

[0027] By using a generative adversarial network, the dual-branch output model is optimized based on radar echo intensity data to obtain an optimized AI model. The generative adversarial network includes a discriminator and a generator. The discriminator is used to discriminate and enhance high-frequency signals in the radar echo intensity data according to the wavelet transform algorithm, and the generator is used to focus on the strong echo region in the radar echo intensity data using an attention mechanism based on the echo intensity variance distribution.

[0028] The AI ​​model is optimized by inputting fused features into the model and then used for weather prediction, outputting AI prediction results for a preset forecast duration.

[0029] In one embodiment, optical flow extrapolation prediction is performed based on the effective group set to obtain optical flow prediction results. The AI ​​prediction results and optical flow prediction results are then fused using Kalman filtering to obtain a fused prediction result map, including:

[0030] Based on the contour coordinate set and the corresponding maximum echo intensity in the effective group set, the optical flow extrapolation algorithm is used to perform extrapolation prediction processing to obtain the optical flow prediction result;

[0031] Acquire AI prediction results and collect real-time observation data; real-time observation data includes real-time observation data of meteorological radar echoes and real-time wind and rain observation data from ground observation stations;

[0032] The AI ​​prediction error is obtained by calculating the average absolute error of the AI ​​prediction results over a preset number of time periods.

[0033] The average absolute error of the optical flow prediction results over a preset number of time steps is calculated to obtain the optical flow prediction error.

[0034] Based on the AI ​​prediction error and the optical flow prediction error, the AI ​​prediction fusion weight and the optical flow prediction fusion weight are determined by an inverse error weighting method.

[0035] Using Kalman filtering, the AI ​​prediction results and optical flow prediction results are fused according to the AI ​​prediction fusion weight and the optical flow prediction fusion weight to obtain a preliminary fusion result.

[0036] The preliminary fusion results are corrected by combining real-time observation data, and the corrected results are then visualized to obtain a fusion forecast result map with a preset fusion forecast duration.

[0037] In one embodiment, based on the fused forecast result map and latitude / longitude contour lines, the precipitation rate is obtained by converting precipitation intensity. The maximum cumulative precipitation value is calculated based on the precipitation rate, and a warning instruction is generated by combining it with preset tiered thresholds, including:

[0038] Based on the fused forecast results map and the latitude and longitude contour lines corresponding to each effective group in the effective group set, the maximum echo intensity corresponding to each effective group in the effective group set is converted into precipitation intensity using a preset ZR relationship to obtain the precipitation rate of the region corresponding to each effective group.

[0039] Based on the precipitation rate of the region corresponding to each effective group, the cumulative precipitation value within different preset durations is calculated. The maximum value of the cumulative precipitation value within each preset duration is then filtered to obtain the maximum cumulative precipitation value for each region corresponding to each effective group within different preset durations.

[0040] The maximum cumulative precipitation value of the area corresponding to each effective group is compared with the preset classification threshold to generate the classification result of the area corresponding to each effective group.

[0041] Disaster characteristics are extracted based on the fused forecast results map, and the disaster characteristics are judged according to the preset disaster conditions to obtain the disaster judgment results. The disaster characteristics include thunderstorm and strong wind characteristics and hail characteristics.

[0042] Based on the latitude and longitude contour lines corresponding to each effective group in the effective group set and the disaster prediction range in the fused forecast result map, the impact range information of the region corresponding to each effective group is determined;

[0043] The judgment results, disaster assessment results, and impact range information are integrated to generate early warning instructions that include disaster type, intensity level, affected area, and effective time period.

[0044] In one embodiment, the mathematical expression for the preset ZR relation is as follows:

[0045]

[0046] in, For a moment coordinate Radar reflectivity factor at that location For the corresponding time coordinate Precipitation rate at the location For a moment coordinate The vertical thickness h of the echo, and All are dynamically adjusted coefficients, and , , and The basic ZR coefficient.

[0047] Secondly, this application also provides a short-term disaster weather forecasting and delivery system, including:

[0048] The data acquisition and grouping generation module is used to collect meteorological radar echo intensity data and wind and rain data from ground observation stations. It uses the Marching Square algorithm to group the radar echo intensity data, resulting in multiple groups and generating the latitude and longitude contour lines corresponding to each group. It then performs validity filtering on each group to generate a set of valid groups.

[0049] The time series construction and severe weather labeling module is used to calculate the local wind and rain signals based on the effective group set, construct the wind and rain signal time series, determine the start and end times of severe weather based on the wind and rain signal time series, and generate severe weather labeling signals.

[0050] The dual-model fusion prediction module is used to perform weather prediction processing based on strong weather marker signals and radar echo intensity data to obtain AI prediction results; it also performs optical flow extrapolation prediction processing based on the effective group set to obtain optical flow prediction results; and combines the AI ​​prediction results and optical flow prediction results through Kalman filtering to obtain the fused forecast result map.

[0051] The graded early warning command generation module is used to generate early warning commands based on the fused forecast result map and latitude and longitude contour lines, by converting precipitation intensity to obtain the precipitation rate, calculating the maximum cumulative precipitation value based on the precipitation rate, and combining it with preset graded thresholds.

[0052] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the first aspect.

[0053] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the first aspect.

[0054] The aforementioned method, system, equipment, and medium for short-term severe weather forecasting first collects dual-source meteorological data and uses the Marching Square algorithm to group and filter it to generate a valid set of categories, thus solving the identification bias problem caused by missing or abnormal data from a single data source. Secondly, it determines the start and end times of severe weather by constructing wind and rain signal time series, achieving automated identification of severe weather processes and eliminating reliance on subjective human judgment. Furthermore, it employs a dual-model scheme combining AI prediction and optical flow extrapolation with Kalman filtering, overcoming the limitations of single prediction technologies and improving the reliability of short-term forecasts. Finally, it generates warning instructions based on precipitation rate classification, addressing the lack of specificity in traditional warnings and enhancing the accuracy of localized customized services. Attached Figure Description

[0055] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0056] Figure 1A flowchart of a short-term disaster weather forecasting and delivery method is provided as an exemplary embodiment of the present invention;

[0057] Figure 2 A flowchart of a method for generating strong weather marker signals is provided as an exemplary embodiment of the present invention;

[0058] Figure 3 This is a schematic diagram of a short-term disaster weather forecasting and delivery system provided as an exemplary embodiment of the present invention. Detailed Implementation

[0059] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0060] In one embodiment, such as Figure 1 As shown, a method for pushing short-term hazardous weather forecasts is provided. This embodiment illustrates the application of this method to a terminal. It is understood that this method can also be applied to a server, or to a system including both a terminal and a server, and implemented through interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0061] S101: Collect meteorological radar echo intensity data and wind and rain data from ground observation stations. Use the Marching Square algorithm to group the radar echo intensity data to obtain multiple groups and generate the latitude and longitude contour lines corresponding to each group. Perform validity filtering on each group to generate a set of valid groups.

[0062] Specifically, meteorological radar data can provide information on precipitation echo intensity over a wide area, while ground observation station data can provide specific wind and rain conditions. Combining the two can compensate for the shortcomings of a single data source. Furthermore, since radar and observation station data may contain missing data or anomalies, preprocessing and validity filtering can be performed. Specifically, the Marching Square algorithm can be used to group radar echo intensity data. The Marching Square algorithm is a contour extraction algorithm that divides radar echo intensity data into multiple groups according to a preset threshold and generates a latitude and longitude contour line for each group. Through grouping, the spatial distribution range of echoes of different intensities can be clearly defined, providing a geospatial information foundation for subsequent analysis. Subsequently, validity filtering can be performed on each group to generate a set of valid groups, retaining groups that meet preset grouping conditions and removing noise and invalid information from the data to ensure the accuracy of subsequent analysis.

[0063] S102: Calculate the local wind and rain signal based on the effective group set, construct the wind and rain signal time series, determine the start and end time of severe weather based on the wind and rain signal time series, and generate severe weather marker signal.

[0064] Specifically, based on the effective group set, wind and rain data, such as maximum wind and rain intensity values, can be extracted from eligible groups within the target area using ground-based observation station data. This data serves as the local wind and rain signal for the current moment and is saved as a time series. Subsequently, the start and end times of severe weather can be determined based on the wind and rain signal time series. For example, considering the short lifespan of severe storms, a maximum wind and rain intensity value greater than 1 mm or equal to 0 mm for 20 consecutive minutes can be used as the criterion for determining the start and end times of severe weather impact. Finally, a severe weather marker signal can be generated as an important input for subsequent weather forecasting. This severe weather marker signal not only identifies the time of occurrence of severe weather but also provides spatial location information, offering crucial spatiotemporal characteristics for subsequent forecasting models.

[0065] S103: Based on strong weather marker signals and radar echo intensity data, weather forecasting is performed to obtain AI forecast results; optical flow extrapolation forecasting is performed based on the effective group set to obtain optical flow forecast results; the AI ​​forecast results and optical flow forecast results are combined and fused through Kalman filtering to obtain the fused forecast result map.

[0066] Specifically, AI prediction focuses on capturing the complex spatial characteristics and evolutionary patterns of severe weather, while optical flow extrapolation prediction is used to fill the forecast gap for severe weather in advection precipitation. Illustratively, based on severe weather marker signals and radar echo intensity data, predictions can be made using an AI model based on the Swim-Transformer and Unet architecture. The optical flow extrapolation method, on the other hand, can predict the future location and intensity changes of the weather system based on its current direction and speed of movement. Furthermore, Kalman filtering, as a dynamic system estimation method, can effectively integrate multi-source prediction results, improving forecast accuracy and stability. By combining AI and optical flow prediction results and fusing them using Kalman filtering, a fused forecast map can be obtained. This fused forecast map retains the AI ​​model's accurate capture of severe weather characteristics while incorporating the advantages of optical flow extrapolation in characterizing advection motion, significantly improving the reliability and robustness of short-term forecasts.

[0067] S104: Based on the fused forecast results map and latitude and longitude contour lines, the precipitation rate is obtained by converting the precipitation intensity. The maximum cumulative precipitation value is calculated based on the precipitation rate, and a warning instruction is generated by combining the preset graded thresholds.

[0068] Specifically, based on the fused forecast results map and latitude / longitude contour lines, a preset ZR relationship can be used to convert the maximum echo intensity corresponding to each effective group into precipitation intensity, obtaining the precipitation rate for each region. Subsequently, based on the precipitation rate of each region, the cumulative precipitation values ​​within different preset durations such as 1 hour and 3 hours can be calculated. The maximum value of the cumulative precipitation value for each preset duration is then filtered to obtain the maximum cumulative precipitation value for each region at different durations. This value can intuitively reflect the degree of precipitation impact that severe weather may cause. Combined with preset grading thresholds such as 10mm, 30mm, 50mm, and 80mm, the maximum cumulative precipitation value can be compared with the thresholds to preliminarily determine the precipitation warning level for each region. Furthermore, thunderstorm, strong wind, and hail disaster characteristics can be extracted from the fused forecast results map to determine whether such disaster risks exist. Based on the latitude / longitude contour lines and the disaster prediction range, the specific impact range and effective time period for each region can be clarified. Finally, by integrating the results of precipitation warning level determination, disaster assessment results, and impact range information, a warning instruction containing disaster type, intensity level, affected area, and effective time period can be generated. This instruction can be accurately pushed to specific users in the corresponding area, such as agriculture, fishery, and tourism, to achieve tiered and personalized warning services, effectively making up for the shortcomings of traditional manual warnings, such as limited coverage and insufficient targeting.

[0069] The aforementioned method first groups radar echo intensity data using the Marching Square algorithm and generates latitude and longitude contour lines, resolving the issues of data grouping and regional division, and providing clear geospatial information for subsequent analysis. Secondly, based on the effective group set, it calculates local wind and rain signals, constructs a time series of wind and rain signals, accurately determines the start and end times of severe weather, and generates severe weather marker signals, improving the accuracy and timeliness of severe weather identification. Furthermore, it combines AI prediction and optical flow extrapolation prediction, and uses Kalman filtering for fusion processing to obtain a fused forecast result map, effectively integrating the advantages of multiple prediction technologies and enhancing the accuracy and reliability of the forecast. Finally, based on the fused forecast result map and latitude and longitude contour lines, it calculates the maximum cumulative precipitation value and generates early warning instructions by converting precipitation intensity, achieving efficient conversion from forecast results to early warning decisions, significantly improving the scientific rigor and practicality of the early warning system.

[0070] In one embodiment, the radar echo intensity data is grouped using the Marching Square algorithm to obtain multiple groups, and latitude and longitude contour lines corresponding to each group are generated; each group is then filtered for validity to generate a set of valid groups, including:

[0071] The radar echo intensity data is grouped according to a preset echo intensity threshold sequence using the Marching Square algorithm to obtain multiple groups;

[0072] Each group is processed for contour extraction to generate the latitude and longitude contour lines corresponding to each group.

[0073] Extract the unit-time precipitation data of each group of associated ground stations from the wind and rain data of the ground observation stations;

[0074] Each group is filtered for validity, and groups with maximum echo intensity greater than the first preset echo threshold and precipitation data per unit time greater than the first preset precipitation threshold are retained. A set of valid groups is generated, which includes the contour coordinate set, internal observation station list, maximum echo intensity and maximum precipitation per unit time information of each valid group.

[0075] Specifically, radar echo intensity data is essentially continuous spatial field data, making it difficult to distinguish regions corresponding to different intensities of precipitation directly. The Marching Square algorithm, however, has the advantage of discretizing and partitioning continuous data fields, allowing for precise division of echo regions with specific intensities based on preset thresholds. For example, in this embodiment, a preset echo intensity threshold sequence of 15, 25, 35, 45, 55, and 65 dBz can be set according to meteorological classification standards for precipitation intensity. 15 dBz is the lowest threshold at which radar can effectively identify precipitation echoes (echoes below this value are mostly noise or non-precipitating clouds), 25-35 dBz corresponds to light to moderate rain, 45-55 dBz corresponds to heavy to torrential rain, and 65 dBz and above may be accompanied by severe convection, hail, and other extreme weather. This sequence can achieve full coverage of weak precipitation, moderate to heavy precipitation, and extreme precipitation. By gridding radar echo intensity data, such as dividing the grid with a latitude and longitude resolution of 0.01°×0.01°, and then using the MarchingSquare algorithm to determine the threshold interval to which the echo intensity of each grid point belongs, adjacent grid points within the same threshold interval can be grouped into the same group, ultimately forming multiple discrete groups corresponding to different intensity echoes, thereby achieving a clear definition of precipitation areas of different intensities.

[0076] Specifically, during group processing, the latitude and longitude contour lines corresponding to each group can be generated simultaneously. These contour lines are crucial for clearly defining the spatial boundaries of the echo area, associating discrete groups with actual geographical regions and providing spatial coordinates for subsequent location of the impact range of severe weather. Illustratively, the Marching Square algorithm, when dividing into groups, can calculate the precise latitude and longitude coordinates of the boundary grid points of each group through linear interpolation. That is, for each group, the algorithm first identifies its edge grid points (i.e., grid points adjacent to other groups or non-echo areas), and then, based on the latitude and longitude information of these edge grid points, fits and generates continuous closed contour lines, ensuring that each contour line completely encompasses all grid points of its corresponding group. For example, if a group covers the area surrounding a township in Taizhou, its generated latitude and longitude contour lines will accurately fit the outer boundaries of all grid points with echoes of the same intensity within that area, intuitively reflecting the actual geographical coverage of the echo area and avoiding subsequent misjudgments of the impact range due to blurred spatial boundaries. Furthermore, while radar echo intensity data can reflect the spatial distribution of precipitation clouds, it is indirect observation data and is susceptible to deviations due to factors such as topography and electromagnetic interference. In contrast, the unit-time precipitation data from ground observation stations is a direct record of near-surface precipitation, possessing higher accuracy and reliability. Therefore, based on the latitude and longitude contours of each group, ground observation stations within the contour lines and within a 5-kilometer radius around them can be identified (the 5-kilometer range is set based on the effective observation radius of the ground observation stations to ensure coverage of all possible precipitation areas corresponding to each group). Then, the unit-time precipitation data of these associated stations can be extracted from the real-time monitoring data of the ground observation stations, where the unit time can be set to 1 hour, conforming to the conventional time scale for precipitation intensity statistics in meteorological operations. Ultimately, a mapping relationship can be established between each group and the precipitation data of its corresponding associated stations.

[0077] Furthermore, the validity screening process for each group can eliminate false groups or weak echo groups that have no actual precipitation significance, reducing interference in subsequent data processing and ensuring that only valid data related to real severe weather events are retained. Since 15 dBz is the minimum threshold for radar to identify precipitation echoes, noise echo groups with no precipitation significance can be eliminated. 0.1 mm is the critical value for effective precipitation in meteorology, eliminating groups with weak echoes but no actual precipitation (such as weak echoes caused by fog or light fog). Therefore, the screening criteria set in this embodiment can be "the maximum echo intensity within the group is greater than the first preset echo threshold (15 dBz) and the precipitation data per unit time is greater than the first preset precipitation threshold (0.1 mm)". Subsequently, for each group, the echo intensity data of all its grid points can be traversed, and the maximum value can be selected as the maximum echo intensity of that group. For the unit time precipitation data of the associated ground stations for each group, either the average value or the maximum value can be taken, but the maximum value better reflects extreme precipitation conditions. Therefore, in this embodiment, the maximum value is taken as the unit time precipitation data for that group. By determining whether the above two indicators simultaneously meet the screening criteria, a group is retained if it does, and removed if it does not. Finally, all retained groups are integrated into a valid group set, and key information for each valid group is recorded simultaneously, such as the contour line coordinate set, the list of internal observation stations, maximum echo intensity, and maximum precipitation per unit time. The contour line coordinate set is used for subsequent location of the affected area, the list of internal observation stations is used for subsequent real-time comparison, the maximum echo intensity is used for subsequent precipitation intensity conversion, and the maximum precipitation per unit time information can be used for subsequent severe weather intensity determination. This screening process ensures that each group in the valid group set corresponds to a real precipitation event, providing a data foundation for subsequent severe weather identification and forecasting.

[0078] In one embodiment, such as Figure 2 As shown, based on the effective group set, the local wind and rain signal is calculated, a wind and rain signal time series is constructed, the start and end times of severe weather are determined according to the wind and rain signal time series, and a severe weather marker signal is generated, including:

[0079] S201: Based on the effective group set, extract the maximum unit time precipitation information of each effective group in the target area for each collection time, and take the maximum value of the maximum unit time precipitation information of each effective group in the target area as the local wind and rain signal at the corresponding collection time;

[0080] S202: Arrange the local wind and rain signals corresponding to multiple acquisition times in chronological order, perform time series construction processing, and obtain a wind and rain signal time series with a preset time resolution;

[0081] S203: Perform severe weather initiation determination processing based on the wind and rain signal time series. When the wind and rain signal in the first continuous preset duration wind and rain signal time series is greater than or equal to the second preset precipitation threshold, a severe weather initiation marker is generated.

[0082] S204: Perform severe weather termination determination processing based on the wind and rain signal time series. When the wind and rain signal in the second consecutive preset duration wind and rain signal time series is equal to the third preset precipitation threshold, a severe weather termination mark is generated.

[0083] S205: Integrate the severe weather start marker and severe weather end marker to generate a severe weather marker signal that includes whether the current situation is a severe weather process, the start time, and the end time.

[0084] Specifically, the set of valid groups generated in the above embodiments already contains the maximum precipitation information per unit time for each valid group, and invalid data has been removed. Therefore, based on this set, the target area range can be defined first, and for each data collection time, all valid groups within the target area can be traversed. Furthermore, since the impact of severe weather on a region is determined by extreme meteorological factors, the precipitation intensity of different valid groups within the same target area varies. Precipitation data from a single group cannot reflect the overall extreme conditions of the region. Therefore, the maximum precipitation information per unit time for each valid group within the target area can be extracted, and the maximum value can be used as the local wind and rain signal for the target area at the corresponding collection time. This allows for accurate capture of the strongest precipitation state in the region at that time, avoiding the masking of the potential impact of severe weather by local weak precipitation data. For example, if a target area contains three valid groups at time t1, with maximum precipitation per unit time of 0.8mm, 1.2mm, and 1.1mm respectively, then the local wind and rain signal at that time is 1.2mm, directly reflecting the most significant precipitation intensity in the region at present, providing a reliable single-time indicator for subsequent time-series analysis.

[0085] Specifically, severe weather is a continuously changing meteorological process, and a single-moment local wind and rain signal cannot determine whether it is in a state of continuous impact. Therefore, the local wind and rain signals corresponding to each collection moment can be arranged sequentially according to the chronological order of data collection to form a wind and rain signal time series with a preset time resolution. The time resolution is consistent with the collection interval, such as 5 minutes per data point, which ensures that the sequence has no time gaps and can continuously cover the weather change process. For example, if the collection interval is 5 minutes, 12 local wind and rain signals can be generated within 1 hour. Arranging them in chronological order yields a wind and rain signal time series. This series can visually present the fluctuation trend of regional precipitation intensity within 1 hour in the form of a curve, providing continuous data support for subsequent determination of the start and end of severe weather and avoiding judgment bias caused by data dispersion. Furthermore, due to the short life cycle and rapid development of severe weather, the first preset continuous duration can be set to 20 minutes, and the second preset precipitation threshold can be set to 1 mm. Subsequently, by traversing the wind and rain signal time series, it is checked segment by segment whether there is a continuous 20-minute signal segment (e.g., with a time resolution of 5 minutes, corresponding to 4 consecutive data points), and the local wind and rain signal at each moment within the signal segment is greater than or equal to 1 mm. If a signal segment that meets the conditions is detected, the first acquisition time of that signal segment is marked as the start time of severe weather, and a severe weather start marker is generated. For example, if t20: 1.1mm, t25: 1.3mm, t30: 1.2mm, and t35: 1.4mm (for 20 consecutive minutes) are all ≥1mm, then t20 is determined to be the start time of severe weather, and a start marker is generated.

[0086] Specifically, based on the meteorological pattern of severe weather dissipation—that is, precipitation gradually weakens to no effective precipitation, and the lack of precipitation can be considered the termination of the process—a second continuous preset duration (20 minutes) can be set to be consistent with the first continuous preset duration to ensure the consistency of the start and end determination logic. The third preset precipitation threshold is 0 mm. Then, by traversing the time series of wind and rain signals, it checks for a continuous 20-minute signal segment where the local wind and rain signal at each moment within that segment is equal to 0 mm. If such a signal segment is detected, the last acquisition time of that segment is marked as the severe weather termination time, generating a severe weather termination marker. For example, if t80:0mm, t85:0mm, t90:0mm, and t95:0mm (continuously 20 minutes) are all 0 mm, then t95 is determined as the severe weather termination time, generating a termination marker. By comparing the start time corresponding to the severe weather start marker, the end time corresponding to the termination marker, and the current acquisition time, it is determined whether a severe weather process is currently underway. If the current time is after the start time and before the end time, then the process is in a severe weather active state. If the current time is after the termination time, it is in an inactive state. The current state, start time, and termination time are then integrated to obtain a severe weather marker signal. For example, when no termination marker is detected, the signal is "Currently in a severe weather process, start time t20, termination time undetermined." After a termination marker is detected, the signal is updated to "Currently not in a severe weather process, start time t20, termination time t95." This marker signal provides clear state information for subsequent multi-model predictions, allowing AI prediction and optical flow extrapolation to be initiated only during active severe weather periods, reducing invalid computations during non-severe weather periods and significantly improving algorithm efficiency.

[0087] In one embodiment, weather forecasting is performed based on strong weather marker signals and radar echo intensity data to obtain AI forecast results, including:

[0088] The severe weather marker signal is processed for state recognition. When the severe weather marker signal is in an active state, the preset frame data of radar echo intensity data is acquired.

[0089] Using preset frame data as input data, the Swim-Transformer is used to perform multi-scale spatial feature extraction processing on the input data to obtain multi-scale spatial features;

[0090] Semantic features are extracted from multi-scale spatial features using a feature extraction network based on the U-Net architecture to obtain high-dimensional feature maps.

[0091] A dual-branch output model is constructed based on meteorological physical equations. The meteorological physical equations consist of local variation terms and advection-induced intensity variation terms. The local variation terms are used to reflect the local changes in radar echo intensity over time, while the advection-induced intensity variation terms are used to reflect the changes in radar echo intensity caused by horizontal motion.

[0092] The high-dimensional feature map is processed by the first and second branches of the dual-branch output model to obtain the corresponding advection term learning results and local term learning results.

[0093] The MOE weighted architecture is used to fuse the learning results of the advection term and the local term to obtain the model fusion features;

[0094] By using a generative adversarial network, the dual-branch output model is optimized based on radar echo intensity data to obtain an optimized AI model. The generative adversarial network includes a discriminator and a generator. The discriminator is used to discriminate and enhance high-frequency signals in the radar echo intensity data according to the wavelet transform algorithm, and the generator is used to focus on the strong echo region in the radar echo intensity data using an attention mechanism based on the echo intensity variance distribution.

[0095] The AI ​​model is optimized by inputting fused features into the model and then used for weather prediction, outputting AI prediction results for a preset forecast duration.

[0096] Specifically, since the lifecycle of severe weather events is mostly within 0-2 hours, the AI ​​prediction process can be initiated only when the severe weather marker signal indicates "currently active," meaning a severe weather initiation marker has been detected but no termination marker has been detected, thus reducing computational consumption. Furthermore, since radar echo data is typically acquired at 5-minute intervals per frame, the first 10 frames of radar echo data can be obtained when the severe weather marker signal indicates "currently active," covering a 50-minute period of recent weather evolution. This comprehensively captures the dynamic characteristics of severe weather from its initial development to its current state without causing computational delays due to excessive data volume, meeting the real-time requirements of short-term forecasting. Using this data as input, multi-scale spatial feature extraction can be performed through Swim-Transformer, accurately capturing weather system features at different scales within the radar echoes. Swim-Transformer can employ a window attention + cross-window connection architecture. For example, each frame of radar echo data (with a resolution of 256×256 pixels, corresponding to an actual geographical area of ​​approximately 50km×50km) can be divided into a 16×16 basic window. The first scale level calculates attention on the pixels within the basic window to capture small-scale convective cell features. The second scale level merges four adjacent basic windows into a large 32×32 window to extract mesoscale cloud cluster features. The subsequent two scale levels continue to merge windows until regional-scale precipitation system features are captured. Through this process, spatial feature tensors at four scale levels can be obtained, with dimensions of [10,64,64,256], [10,32,32,512], [10,16,16,1024], and [10,8,8,2048], respectively, where 10 represents the frame number, and the last three dimensions represent the spatial resolution and the number of feature channels, respectively. This addresses the problem of insufficient small-scale feature capture by traditional Transformers and lays a multi-dimensional foundation for subsequent semantic feature extraction.

[0097] Specifically, semantic features can then be extracted from multi-scale spatial features using a feature extraction network based on the U-Net architecture, transforming discrete spatial features into structured features with meteorological significance. The advantage of the U-Net architecture lies in its symmetrical encoder-decoder structure. The encoder part can use convolutional and pooling layers to reduce the dimensionality of the multi-scale features output by the Swim-Transformer, gradually compressing the spatial dimension and improving feature abstraction. For example, features of [10,64,64,256] can be compressed to [10,8,8,2048], focusing on extracting semantic category features such as strong echo core areas, echo edge areas, and no-echo areas. The decoder part can use transposed convolutional layers and skip connections to fuse highly abstract features with shallow features of the corresponding scale from the encoder, restoring the spatial resolution to [10,256,256,256], ultimately obtaining a high-dimensional feature map. Furthermore, a dual-branch output model can be constructed based on meteorological physical equations to inject physical constraints into the AI ​​model, addressing the problem of purely data-driven models lacking interpretability and easily deviating from actual weather patterns. Illustratively, the meteorological physical equations can be:

[0098]

[0099] in, The radar echo intensity (unit: dBz) reflects the intensity of precipitation clouds. Indicates time, This represents the total rate of change of radar echo intensity. This is a local variation term, which reflects the local variation of echo intensity at a fixed spatial point over time. , These are the wind speed components in the horizontal direction (e.g., east-west, north-south). , These are latitude and longitude coordinates, The term representing the intensity change caused by advection (referred to as the advection term) reflects the intensity change of the echo system caused by the movement of the horizontal airflow.

[0100] Specifically, based on this equation, the first branch (advection term branch) of the dual-branch output model can employ a 3-layer convolutional neural network, with inputs being features related to the spatial gradient in a high-dimensional feature map, such as... , The mapping characteristics are obtained by learning the weight parameters to fit the advection term pair. The contribution of the local term can output the learning results of the advection term. The second branch (local term branch) can use a convolutional neural network with the same structure. The input is the time-varying features in the high-dimensional feature map, such as the difference features of the echo intensity between adjacent frames, and it fits the local term pair. The model outputs the local component's learning results. By using a dual-branch output model, the two core physical mechanisms of echo intensity change can be modeled separately, avoiding the confusion between the advection and local variations in a single-branch model, thus improving the model's understanding of severe weather evolution. Since the advection and local components have different magnitudes of influence on the total echo intensity change, and their proportions vary in different scenarios (e.g., periphery precipitation of typhoons is dominated by the advection component, while local convection is dominated by the local component), fixed-weight fusion can lead to prediction biases in some scenarios. Therefore, a MOE (Mixture of Experts) weighted architecture can be used to fuse the two learning results to dynamically balance the contribution weights of the advection and local components in different weather scenarios. For example, two expert networks (corresponding to the weight predictions of the advection and local components, respectively) and one gating network can be set up. The gating network takes the echo movement velocity and echo intensity variance from the high-dimensional feature map as input. The movement velocity reflects the importance of the advection component, and the variance reflects the importance of the local component. The weight coefficients of the two expert networks can then be output using a softmax function. The final model fusion features are obtained by weighting and summing the learning results of the advection term and the local term according to the weight coefficients.

[0101] Specifically, optimizing the dual-branch output model using Generative Adversarial Networks (GANs) can solve the problem of intensity attenuation in strong weather forecasts using traditional AI models (i.e., predicted echo intensity is generally lower than the actual value). The GAN consists of a discriminator and a generator. The generator's input is the model's fused features, and its output is the predicted radar echo intensity map. The discriminator's input is the real radar echo intensity map (i.e., historical data of radar echo intensity) and the predicted radar echo intensity map, and its output is the true / predicted probability. Furthermore, to enhance the prediction accuracy for strong echo regions, a wavelet transform algorithm can be introduced into the discriminator. This involves performing two-dimensional wavelet decomposition (using the db4 wavelet basis) on the input echo intensity map to obtain low-frequency components (reflecting the overall cloud structure) and high-frequency components (reflecting strong echo details, such as sharp changes in the core area of ​​strong echoes). Subsequently, the discriminator is trained only on the high-frequency components, and gradient backpropagation forces the generator to improve the accuracy of generating high-frequency details. Since the extreme characteristics of severe weather (such as hail and short-duration heavy rain) are mainly reflected in high-frequency components, the above processing can enhance high-frequency discrimination and thus effectively alleviate intensity attenuation. Simultaneously, an attention mechanism based on the echo intensity variance distribution can be added to the generator, i.e., calculating the echo intensity variance of the generator's intermediate feature maps. The attention weight is calculated using the following formula. :

[0102]

[0103] in For feature maps Echo intensity at location To maximize the global variance, the attention weights are multiplied element-wise with the intermediate feature maps, allowing the generator to focus on strong echo regions with high variance, such as those greater than 20 dBz², preventing these regions from being diluted by weak echo features. Through iterative training of the GAN, such as setting the number of iterations to 100 rounds and the batch size to 8, an optimized AI model can be obtained. The fused features are then input into the optimized AI model for weather prediction, outputting AI prediction results for a preset forecast duration. For example, since the lifecycle of severe weather is 0-2 hours, the preset forecast duration can be set to 2 hours, corresponding to 20 frames of prediction data (one frame every 5 minutes, consistent with the input frame interval). The AI ​​prediction result can be presented as a spatial distribution map of radar echo intensity every 5 minutes for the next 2 hours, with a resolution of 256×256 pixels, and the corresponding latitude and longitude range can be adjusted according to the target area. The results not only contain the spatial distribution of echo intensity, but also imply information such as the direction of movement and the range of impact of severe weather, laying the foundation for subsequent fusion with optical flow prediction results, while meeting the requirements of short-term forecasts for high temporal resolution and accurate spatial details.

[0104] In one embodiment, optical flow extrapolation prediction is performed based on the effective group set to obtain optical flow prediction results. The AI ​​prediction results and optical flow prediction results are then fused using Kalman filtering to obtain a fused prediction result map, including:

[0105] Based on the contour coordinate set and the corresponding maximum echo intensity in the effective group set, the optical flow extrapolation algorithm is used to perform extrapolation prediction processing to obtain the optical flow prediction result;

[0106] Acquire AI prediction results and collect real-time observation data; real-time observation data includes real-time observation data of meteorological radar echoes and real-time wind and rain observation data from ground observation stations;

[0107] The AI ​​prediction error is obtained by calculating the average absolute error of the AI ​​prediction results over a preset number of time periods.

[0108] The average absolute error of the optical flow prediction results over a preset number of time steps is calculated to obtain the optical flow prediction error.

[0109] Based on the AI ​​prediction error and the optical flow prediction error, the AI ​​prediction fusion weight and the optical flow prediction fusion weight are determined by an inverse error weighting method.

[0110] Using Kalman filtering, the AI ​​prediction results and optical flow prediction results are fused according to the AI ​​prediction fusion weight and the optical flow prediction fusion weight to obtain a preliminary fusion result.

[0111] The preliminary fusion results are corrected by combining real-time observation data, and the corrected results are then visualized to obtain a fusion forecast result map with a preset fusion forecast duration.

[0112] Specifically, the advantages of optical flow algorithms in motion trajectory prediction can be leveraged to further enhance the forecasting capability for advection-dominated precipitation. The effective group set clearly defines the spatial boundaries (contour coordinate set) and core intensity characteristics (maximum echo intensity) of each strong echo region. Therefore, these can be used as input to ensure targeted extrapolation. The principle of optical flow extrapolation is to assume that the motion pattern of the echo region remains stable over a short period, such as 0-2 hours, and to predict the echo position and range at future times by calculating the motion vectors of the echo contour lines at adjacent times. Illustratively, the Lucas-Kanade sparse optical flow algorithm can be used. First, key feature points, such as inflection points and endpoints of the contour lines, are extracted from the contour coordinate set of each effective group. 20-30 feature points are extracted for each contour line to balance accuracy and efficiency. Then, the horizontal motion vector u and vertical motion vector v of the feature points in two adjacent frames (with a 5-minute acquisition interval) are calculated. The motion trend of the feature points is fitted using the least squares method to ensure the stability of the vector calculation. Based on the fitted motion vectors, the echo contour lines for each frame (20 frames in total, 5 minutes / frame) within the next 2 hours can be extrapolated. Furthermore, since the optical flow algorithm has a weak ability to characterize intensity changes, the maximum echo intensity of each group can be kept constant. The final optical flow prediction result is the set of echo region latitude and longitude contour lines plus the maximum echo intensity for each frame within the next 2 hours. This accurately reflects the movement trajectory of the echoes and is suitable for precipitation dominated by advection systems (such as frontal precipitation), thus compensating for the shortcomings of AI prediction in forecasting this type of precipitation.

[0113] Specifically, the AI ​​prediction result is the spatial distribution map of radar echo intensity over the next 2 hours (20 frames) output in the above embodiment, which already has the ability to characterize the evolution of strong convective cell intensity. By collecting real-time observation data, the latest situation can be obtained to calibrate prediction errors. This real-time observation data includes real-time observation data of meteorological radar echoes (reflecting the current cloud system situation) and real-time observation data of wind and rain from ground observation stations (reflecting the near-surface meteorological elements). The collection interval between these two data sets and the prediction data is consistent (5 minutes / time) to ensure time alignment. Subsequently, the mean absolute error (MAE) of the AI ​​prediction result and the optical flow prediction result over a preset time period can be calculated to quantify the forecast accuracy of the two models and provide data support for subsequent dynamic weight allocation. The preset time periods can be set as three core time periods: 15 minutes, 30 minutes, and 60 minutes, corresponding to the early warning needs of immediate preparation, emergency response activation, and continuous prevention, respectively. Taking AI prediction error as an example, for a preset time period, such as 30 minutes, the echo intensity map of that time period in the AI ​​prediction result can be extracted and spatially matched (aligned by latitude and longitude coordinates) with the real-time radar echo observation map at the corresponding time. Within the effective echo region (i.e., the region enclosed by the contour line in the optical flow extrapolation), the predicted echo intensity is calculated pixel by pixel. Compared with actual echo intensity The absolute difference is then taken, and the arithmetic mean of the absolute differences of all pixels is taken to obtain the AI ​​prediction error for that time. The calculation formula is:

[0114]

[0115] in The set of pixels in the effective echo region. for The total number of pixels within. And the optical flow prediction error. The calculation logic is completely consistent with the above formula, only replacing the predicted value with the echo intensity extrapolated from optical flow (the maximum echo intensity remains unchanged in optical flow extrapolation, so its predicted intensity map is obtained by filling the contour area with the maximum echo intensity of each group). Through this calculation, the accuracy difference between the two prediction processes at different times can be intuitively distinguished. Subsequently, the fusion weight is determined by using an inverse error weighting method based on the AI ​​prediction error and the optical flow prediction error, which can ensure that the fusion result fully utilizes the advantageous time periods of each model. For example, let the AI ​​prediction fusion weight be... The optical flow prediction fusion weight is The corresponding weight is then calculated using the following formula:

[0116]

[0117] And satisfy This is to ensure the rationality of weight allocation.

[0118] Specifically, based on Kalman filtering, a prediction-update loop can be used to dynamically correct the fusion result by combining the model's predicted values ​​with error statistics, avoiding the problem that simple weighting cannot handle sudden error changes. For example, first define the state vector. To predict the echo intensity of each pixel within the echo region at the predicted time, prior estimates can be calculated based on the aforementioned weights during the prediction phase. ,in , Let A and B be the predicted intensity vectors of optical flow at time k, respectively. The prior error covariance matrix can also be calculated:

[0119]

[0120] These are the prediction error covariances of AI and optical flow, respectively, obtained from historical error statistics. During the update phase, the intensity vector of real-time observed data can be used. Calculate the Kalman gain for the observed values. ( The observation error covariance (determined by the noise characteristics of real-time data) is then processed through... The posterior estimate, i.e. the preliminary fusion result, is obtained.

[0121] Finally, by combining real-time observation data to correct and visualize the preliminary fusion results, the latest real-time data can be used to further calibrate prediction biases and transform abstract numerical results into intuitive geographic forecast products. For example, the intensity residual between the preliminary fusion results and the corresponding real-time observation data can be calculated first. Assuming that the residual changes linearly over a short period (e.g., 15 minutes), the corresponding linear formula can be used to correct the preliminary fusion results of subsequent forecast frames, ensuring that the corrected results better match the latest weather evolution trends. After correction, visualization transformation can be performed, that is, the corrected echo intensity vector of each frame is projected onto a geographic base map according to latitude and longitude coordinates, and the administrative boundaries of the target area, such as the scope of towns, farms, and fishing grounds, are superimposed. Color-coded mapping is used, such as light blue for 15-25dBz, dark blue for 25-35dBz, yellow for 35-45dBz, orange for 45-55dBz, and red for above 55dBz, thus intuitively distinguishing the echo intensity levels and finally generating a fusion forecast result map for a preset fusion forecast duration (2 hours). This map clearly shows the movement path, impact range, and intensity changes of the strong echo over the next two hours, providing direct spatial basis for generating warning instructions.

[0122] In one embodiment, based on the fused forecast result map and latitude / longitude contour lines, the precipitation rate is obtained by converting precipitation intensity; the maximum cumulative precipitation value is calculated based on the precipitation rate; and a warning instruction is generated by combining the preset grading thresholds, including:

[0123] Based on the fused forecast results map and the latitude and longitude contour lines corresponding to each effective group in the effective group set, the maximum echo intensity corresponding to each effective group in the effective group set is converted into precipitation intensity using a preset ZR relationship to obtain the precipitation rate of the region corresponding to each effective group.

[0124] Based on the precipitation rate of the region corresponding to each effective group, the cumulative precipitation value within different preset durations is calculated. The maximum value of the cumulative precipitation value within each preset duration is then filtered to obtain the maximum cumulative precipitation value for each region corresponding to each effective group within different preset durations.

[0125] The maximum cumulative precipitation value of the area corresponding to each effective group is compared with the preset classification threshold to generate the classification result of the area corresponding to each effective group.

[0126] Disaster characteristics are extracted based on the fused forecast results map, and the disaster characteristics are judged according to the preset disaster conditions to obtain the disaster judgment results. The disaster characteristics include thunderstorm and strong wind characteristics and hail characteristics.

[0127] Based on the latitude and longitude contour lines corresponding to each effective group in the effective group set and the disaster prediction range in the fused forecast result map, the impact range information of the region corresponding to each effective group is determined;

[0128] The judgment results, disaster assessment results, and impact range information are integrated to generate early warning instructions that include disaster type, intensity level, affected area, and effective time period.

[0129] Specifically, the preset ZR relation can be:

[0130]

[0131] in, For a moment coordinate Radar reflectivity factor at that location For the corresponding time coordinate Precipitation rate at the location For a moment coordinate The vertical thickness h of the echo, and All are dynamically adjusted coefficients, and , , and The basic ZR coefficient.

[0132] The aforementioned preset ZR relationship, through dynamic adjustment of echo vertical thickness, can improve the conversion accuracy for different cloud types (convective clouds, stratiform clouds). Illustratively, based on the above formula, the maximum echo intensity of each effective group can be extracted from the fused forecast result map. Then, the vertical thickness of the echo is obtained through radar vertical profile data or empirical models. Substitute into the above formula to calculate and Finally, the precipitation rate can be solved by deforming the above-mentioned preset ZR relationship. .

[0133] Specifically, based on the precipitation rate of the corresponding region for each effective group. (t) (a time-varying sequence) can calculate the cumulative precipitation value within different preset durations (e.g., 1 hour, 3 hours) and filter out the maximum cumulative precipitation value. By comparing the maximum cumulative precipitation value of each valid group with preset classification thresholds, a precipitation intensity level determination result can be generated. The preset classification thresholds must conform to meteorological warning standards and can be adjusted according to regional disaster prevention needs. For example, a blue warning can be set for 1-hour cumulative precipitation ≥20mm, yellow for ≥30mm, orange for ≥50mm, and red for ≥80mm. Similarly, a blue warning can be set for 3-hour cumulative precipitation ≥50mm, yellow for ≥70mm, orange for ≥100mm, and red for ≥150mm. If the maximum 1-hour cumulative precipitation in a certain area is 35mm, it can be classified as a yellow warning; if the maximum 3-hour cumulative precipitation is 90mm, it can be classified as an orange warning. This allows for standardized classification of precipitation intensity, giving warnings clear intensity levels and facilitating tiered responses from disaster prevention departments. Furthermore, it can extract disaster characteristics such as thunderstorms, strong winds, and hail from the fused forecast results map to compensate for the limitations of single precipitation warnings. The core characteristics of thunderstorms and strong winds are large and rapidly moving echo intensity gradients, i.e., echo intensity gradient ≥ 10 dBz / km and moving speed ≥ 20 m / s. The core characteristics of hail are strong echoes and thick vertical development, i.e., echo intensity ≥ 50 dBz and vertical thickness ≥ 8 km. Therefore, the spatial gradient of echo intensity (the ratio of the difference in echo intensity between adjacent pixels to the spatial distance) and the movement vector (the movement trend predicted by optical flow extrapolation or AI) can be extracted from the fused forecast results map. If a region simultaneously meets the criteria of "echo intensity gradient ≥ 10 dBz / km" and "moving speed ≥ 20 m / s," it can be identified as having thunderstorm and strong wind characteristics; if the echo intensity is ≥ 50 dBz and the vertical thickness is ≥ 8 km, it can be identified as having hail characteristics. Subsequently, by performing logical judgments on the preset disaster conditions, disaster judgment results such as thunderstorms, strong winds, hail, or none can be obtained, thereby achieving comprehensive identification of multiple disaster types.

[0134] Furthermore, latitude and longitude contour lines define the spatial boundaries of echoes. It is necessary to spatially overlay the disaster prediction range (such as the impact path of thunderstorms and strong winds) with these contour lines, and combine this with the forecast time series to clarify the affected area (such as specific towns and villages) and the effective time period (such as the next 0-2 hours). For example, if the latitude and longitude contour line of a thunderstorm and strong wind echo covers the western part of town A, and the fused forecast shows that it will move eastward at a speed of 25 m / s, covering the entire area of ​​town A in the next hour, then the affected area information is "town A, next 0-1 hour." This refines the warning from the regional level to the local level, meeting the needs of agriculture, fisheries, and other fields for precise spatial services. Finally, by integrating the precipitation level determination results, disaster assessment results, and affected area information, a warning instruction containing the disaster type, intensity level, affected area, and effective time period can be generated, providing precise and multi-dimensional action guidance for localized disaster prevention.

[0135] Based on the same inventive concept, such as Figure 3 As shown in the figure, this application also provides a short-term disaster weather forecast push system 300 for implementing the short-term disaster weather forecast push method described above. The solution provided by this system is similar to the solution described in the above method; therefore, the specific limitations of one or more embodiments of the short-term disaster weather forecast push system provided below can be found in the limitations of the method embodiments above, and will not be repeated here. The system includes:

[0136] The data acquisition and grouping generation module 301 is used to acquire meteorological radar echo intensity data and wind and rain data from ground observation stations. It uses the Marching Square algorithm to group the radar echo intensity data to obtain multiple groups and generates the latitude and longitude contour lines corresponding to each group. It also performs validity filtering on each group to generate a set of valid groups.

[0137] The time series construction and severe weather marking module 302 is used to calculate the local wind and rain signals based on the effective group set, construct the wind and rain signal time series, determine the start and end times of severe weather based on the wind and rain signal time series, and generate severe weather marking signals.

[0138] The dual-model fusion prediction module 303 is used to perform weather prediction processing based on strong weather marker signals and radar echo intensity data to obtain AI prediction results; to perform optical flow extrapolation prediction processing based on the effective group set to obtain optical flow prediction results; and to combine the AI ​​prediction results and optical flow prediction results through Kalman filtering fusion processing to obtain the fused forecast result map.

[0139] The graded early warning instruction generation module 304 is used to generate early warning instructions based on the fused forecast result map and latitude and longitude contour lines, by converting precipitation intensity to obtain the precipitation rate, calculating the maximum cumulative precipitation value based on the precipitation rate, and combining it with preset graded thresholds.

[0140] In one exemplary embodiment, the present invention also provides a computer device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the short-term disaster weather forecast push method of this application. A multi-core processor is preferred to improve the parallel processing capability of the system. The memory provides sufficient temporary storage space to support program execution and data processing. The memory capacity should be large enough to accommodate large amounts of data and computational tasks.

[0141] In one exemplary embodiment, the present invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of a short-term hazardous weather forecast push method of the present application. The computer-readable storage medium may include: a read-only memory, a random access memory (RAM), a solid-state drive (SSD), or an optical disc, etc.

[0142] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A short-term disaster weather forecast pushing method, characterized in that, The method comprises: Collecting weather radar echo intensity data and ground observation station wind and rain data, grouping the radar echo intensity data by Marching Square algorithm to obtain multiple groups, and generating latitude and longitude contour lines corresponding to each group; filtering the effectiveness of each group to generate an effective group set; Based on the effective group set, calculate the local wind and rain signal, construct the wind and rain signal time series, determine the strong weather start and end time according to the wind and rain signal time series, and generate a strong weather marker signal; Based on the strong weather marker signal and the radar echo intensity data, weather prediction processing is performed to obtain an AI prediction result; based on the effective group set, optical flow extrapolation prediction processing is performed to obtain an optical flow prediction result, and the AI prediction result and the optical flow prediction result are combined to obtain a fusion prediction result graph through Kalman filter fusion processing; Based on the fusion prediction result graph and the latitude and longitude contour lines, the precipitation rate is obtained through the conversion of precipitation intensity processing, the maximum cumulative precipitation value is calculated according to the precipitation rate, and the warning instruction is generated in combination with the preset grading threshold.

2. The method of claim 1, wherein, The radar echo intensity data is grouped by Marching Square algorithm to obtain multiple groups, and the latitude and longitude contour lines corresponding to each group are generated; Filtering the effectiveness of each group to generate an effective group set, comprising: Grouping the radar echo intensity data according to a preset echo intensity threshold sequence by Marching Square algorithm to obtain multiple groups; Each group is respectively subjected to contour extraction processing to generate the latitude and longitude contour lines corresponding to each group; Extracting the unit time precipitation data of the associated ground station of each group from the ground observation station wind and rain data; Each group is subjected to effectiveness screening processing, and groups with maximum echo intensity greater than a first preset echo threshold and unit time precipitation data greater than a first preset precipitation threshold are retained to generate the effective group set, which contains the contour line coordinate set, internal observation station list, maximum echo intensity and maximum unit time precipitation information of each effective group.

3. The method of claim 1, wherein, Based on the effective group set, calculate the local wind and rain signal, construct the wind and rain signal time series, determine the strong weather start and end time according to the wind and rain signal time series, and generate a strong weather marker signal, comprising: Based on the effective group set, the maximum unit time precipitation information of each effective group in the target area is extracted for each collection time, and the maximum value in the maximum unit time precipitation information of each effective group in the target area is taken as the local wind and rain signal corresponding to the collection time; Arrange the local wind and rain signals corresponding to multiple collection times in chronological order to perform time series construction processing to obtain the wind and rain signal time series with a preset time resolution; According to the wind and rain signal time series, strong weather start determination processing is performed, and when the wind and rain signal in the wind and rain signal time series of a first continuous preset time length is greater than or equal to a second preset precipitation threshold, a strong weather start marker is generated. According to the wind and rain signal time sequence, a strong weather termination marker is generated when the wind and rain signal in the wind and rain signal time sequence of a second continuous preset time length is equal to a third preset precipitation threshold; The strong weather start marker and the strong weather termination marker are integrated to generate the strong weather marker signal containing information about whether the current weather is in a strong weather process, the start time and the termination time.

4. The method of claim 1, wherein, Based on the strong weather marker signal and the radar echo intensity data, weather prediction processing is performed to obtain an AI prediction result, including: When the strong weather marker signal is in an active state, a preset frame of data of the radar echo intensity data is obtained through state recognition processing of the strong weather marker signal; The preset frame of data is taken as input data, and multi-scale spatial feature extraction processing is performed on the input data by Swim-Transformer to obtain multi-scale spatial features; Semantic feature extraction is performed on the multi-scale spatial features by a feature extraction network based on a U-Net architecture to obtain a high-dimensional feature map; A double-branch output model is constructed based on meteorological physical equations, the meteorological physical equations are composed of a local change term and an advection-induced intensity change term, the local change term is used to reflect the local change of radar echo intensity with time, and the advection-induced intensity change term is used to reflect the change of radar echo intensity caused by horizontal motion; The high-dimensional feature map is processed by the first branch and the second branch of the double-branch output model respectively to obtain corresponding advection term learning results and local term learning results; MOE weighted architecture is used to fuse the advection term learning results and the local term learning results to obtain model fusion features; An optimized AI model is obtained by optimizing the double-branch output model based on the radar echo intensity data through a generative adversarial network, wherein the generative adversarial network includes a discriminator and a generator, the discriminator is used to perform discriminative reinforcement processing on high-frequency signals in the radar echo intensity data according to a wavelet transform algorithm, and the generator is used to focus on strong echo regions in the radar echo intensity data using an attention mechanism based on echo intensity variance distribution; The model fusion features are input into the optimized AI model for weather prediction processing, and the AI prediction result of a preset prediction time is output.

5. The method of claim 4, wherein, Based on the effective group set, optical flow extrapolation prediction processing is performed to obtain an optical flow prediction result, and the AI prediction result and the optical flow prediction result are combined for Kalman filter fusion processing to obtain a fusion prediction result map, including: Based on the contour line coordinate set and the corresponding maximum echo intensity in the effective group set, optical flow extrapolation algorithm is used for extrapolation prediction processing to obtain an optical flow prediction result; The AI prediction result is obtained, and real-time observation data is collected; the real-time observation data includes meteorological radar echo real-time observation data and ground observation station wind and rain real-time observation data; The average absolute error of the AI prediction result at a preset time is calculated to obtain an AI prediction error. Calculate the average absolute error of the optical flow prediction result at the preset time, and obtain an optical flow prediction error; According to the AI prediction error and the optical flow prediction error, an error inverse weighting method is used to determine an AI prediction fusion weight and an optical flow prediction fusion weight; Using Kalman filtering technology, the AI prediction result and the optical flow prediction result are fused according to the AI prediction fusion weight and the optical flow prediction fusion weight to obtain a preliminary fusion result; The preliminary fusion result is corrected by combining the real-time observation data to obtain a corrected result, which is subjected to visual transformation processing to obtain the fusion prediction result graph of the preset fusion prediction length.

6. The method of claim 1, wherein, Based on the fusion prediction result graph and the latitude and longitude contour line, the precipitation rate is obtained by converting the precipitation intensity, and the maximum cumulative precipitation value is calculated according to the precipitation rate, and a warning instruction is generated in combination with a preset classification threshold, including: Based on the fusion prediction result graph and the latitude and longitude contour line corresponding to each effective group in the effective group set, a preset Z-R relationship is used to convert the maximum echo intensity corresponding to each effective group in the effective group set to obtain the precipitation rate of the area corresponding to each effective group. According to the precipitation rate of each effective group corresponding area, the cumulative precipitation value in different preset time is calculated respectively, and the maximum value of the cumulative precipitation value of each preset time is selected to obtain the maximum cumulative precipitation value of each effective group corresponding area in different preset time. The maximum cumulative precipitation value of each effective group corresponding area is compared with the preset classification threshold to generate a determination result of the corresponding level of each effective group corresponding area. Based on the fusion prediction result graph, extract disaster features, and judge the disaster features according to the preset disaster conditions to obtain a disaster judgment result. The disaster features include thunderstorm wind features and hail features. Based on the latitude and longitude contour line corresponding to each effective group in the effective group set and the disaster prediction range in the fusion prediction result graph, the influence range information of each effective group corresponding area is determined. Integrate the determination result, the disaster judgment result and the influence range information to generate the warning instruction containing disaster type, intensity level, influence area and effective period.

7. The method of claim 6, wherein, The mathematical expression of the preset Z-R relationship is as follows: wherein, is the radar reflectivity factor at the time coordinate, wherein, is the precipitation rate at the time coordinate, wherein, is the vertical thickness h of the echo at the time coordinate, wherein, and are dynamic adjustment coefficients, and , , and are base Z-R coefficients.

8. A short-term disaster weather forecast push system, characterized by, The system comprises: A data acquisition and grouping generation module is used to acquire meteorological radar echo intensity data and ground observation station wind and rain data, group the radar echo intensity data by Marching Square algorithm to obtain multiple groups, and generate latitude and longitude contour lines corresponding to each group. Each group is filtered for effectiveness to generate an effective group set; A time sequence construction and strong weather marking module is used to calculate local wind and rain signals based on the effective group set, construct a wind and rain signal time sequence, determine the start and end time of strong weather according to the wind and rain signal time sequence, and generate a strong weather marking signal; The double-model fusion prediction module is configured to perform weather prediction processing based on the strong weather marking signal and the radar echo intensity data to obtain an AI prediction result; perform optical flow extrapolation prediction processing based on the effective group set to obtain an optical flow prediction result; and combine the AI prediction result and the optical flow prediction result through Kalman filter fusion processing to obtain a fusion prediction result map. The hierarchical early warning instruction generation module is configured to perform conversion and precipitation intensity processing based on the fusion prediction result map and the latitude and longitude contour line to obtain a precipitation rate, calculate a maximum cumulative precipitation value according to the precipitation rate, and generate an early warning instruction in combination with a preset hierarchical threshold. 9.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-8 when the computer program is executed by the processor. The computer program is executed by the processor to implement the steps of the method in any one of claims 1 to 7.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the steps of the method in any one of claims 1 to 7.

Images