A short-time heavy rainfall early warning method and device, electronic equipment and medium

Abstract

The application discloses a short-time heavy rainfall early warning method and device, electronic equipment and medium, and belongs to the meteorological forecasting technical field. Based on the quality-controlled automatic station real-time data and sounding profile data, a real-time environmental data set in a preset time period is constructed, and based on numerical prediction field data, a plurality of time-effective prediction trend environmental data sets in a future period are constructed; based on the real-time environmental data set, the prediction trend environmental data set and topographic data, a short-time heavy rainfall environmental parameter matrix is constructed, and a trigger window state is determined according to a preset short-time heavy rainfall trigger window discrimination rule; the short-time heavy rainfall environmental parameter matrix and the state of the trigger window are input into the physical constraint prompt word template, a large language model is called to generate an interpretation recognition result; and according to the state of the trigger window and the interpretation recognition result, a short-time heavy rainfall intelligent early warning result is generated. The application improves the refinement, intelligence and business application capability of short-time heavy rainfall early warning.

Description

Technical Field

[0001] This invention relates to the field of meteorological forecasting technology, and in particular to a method, device, electronic equipment, and medium for short-term heavy rainfall early warning. Background Technology

[0002] Short-duration heavy rainfall is one of the most common and potentially hazardous weather events in severe convective weather, leading to urban flooding, flash floods, geological disasters, and transportation risks. Located on the southeast coast, the Minnan region is influenced by multiple factors, resulting in short-duration heavy rainfall characterized by its sudden onset, small spatial scale, rapid development and evolution, and significant local variations. Traditional methods for identifying and warning of severe convection mainly rely on numerical weather prediction, radar echo extrapolation, real-time monitoring by automatic weather stations, radiosonde environment analysis, and forecasters' subjective comprehensive judgment. While these methods have a certain operational foundation, they still have shortcomings in identifying short-duration heavy rainfall triggers, increasing lead times, and interpreting information in complex underlying surface areas.

[0003] In existing technologies, common methods for identifying or issuing early warnings of short-duration heavy precipitation typically include the following steps: 1. Acquiring multi-source meteorological data, such as numerical weather prediction products, radar data, automatic weather station observation data, and radiosonde data; 2. Interpolating or resampling the multi-source data onto a unified grid to form a meteorological analysis field or forecast field with a unified spatiotemporal resolution; 3. Identifying areas of heavy precipitation based on radar echo intensity, echo movement direction, echo extrapolation results, or actual precipitation changes; 4. Judging the probability of short-duration heavy precipitation occurrence by combining empirical thresholds or statistical models; 5. Generating risk levels or early warning prompts based on indicators such as precipitation intensity, affected area, and duration. Forecasters then correct the system output results by combining weather patterns, historical experience, and local climate characteristics, ultimately forming short-duration heavy precipitation early warning conclusions, service materials, or decision-making recommendations.

[0004] The above methods can provide some reference in routine business operations, but they still have the following shortcomings.

[0005] First, the spatiotemporal matching process of multi-source data is prone to introducing physical biases. Automatic weather station data typically has high temporal resolution, but its spatial distribution is limited by the network layout. Radiosonde data provides relatively complete vertical structure information, but there are usually only a limited number of occurrences per day, making it difficult to reflect the rapid evolution of the boundary layer. Numerical weather forecast data provides continuous spatial fields and future trend information, but its ability to characterize local triggering processes is limited. Existing methods typically use spatial interpolation, temporal interpolation, or grid resampling to uniformly process multi-source data, but this process may disrupt the original physical structure of different data, leading to the smoothing or weakening of small-scale triggering signals such as abrupt changes in boundary layer temperature and humidity, local convergence, the advance and retreat of sea-land wind fronts, and orographic uplift.

[0006] Second, existing methods lack sufficient connection between the "current real environment" and the "future evolution trend." The occurrence of short-duration heavy precipitation depends not only on the current boundary layer state but also on the evolution of thermal, dynamic, and water vapor conditions over the next few hours. Traditional early warning methods often rely heavily on one type of data, such as relying solely on radar extrapolation, single-valued forecast fields, or radiosondes at a specific time, making it difficult to form a continuous diagnostic chain of "current real environment - future forecast trend." When boundary layer temperature and humidity conditions, low-level convergence, or sea-land wind fronts change rapidly within a short period, static or low-frequency updated diagnostic results are prone to lag, affecting the accuracy of trigger window identification.

[0007] Third, the local triggering factors in complex underlying surface areas are insufficiently expressed. The Minnan region has diverse topography and underlying surface types, including coastal areas, hills, valleys, and plateaus. Short-duration heavy precipitation is often triggered by processes such as sea-land wind fronts, valley winds, valley convergence, orographic lifting, and low-level water vapor accumulation. Existing technologies mostly rely on general convective parameters and empirical thresholds for judgment, resulting in insufficient refinement of regional triggering factors. This makes it difficult to accurately characterize localized features such as the depth of sea-land wind front intrusion, enhanced valley convergence, slope lifting potential, and coastal ventilation corridors, leading to insufficient applicability of early warning results in complex terrain and sea-land interface areas.

[0008] Fourth, existing intelligent identification methods lack sufficient physical interpretation capabilities. In recent years, machine learning and deep learning methods have been used for severe convection identification and precipitation forecasting. However, most methods focus on the statistical mapping between sample characteristics and precipitation results, with outputs typically presented as probabilities, grades, or precipitation areas, making it difficult to clearly explain the physical causes of short-term heavy precipitation. This problem limits the understandability, traceability, and correctability of intelligent identification results in operational forecasting.

[0009] Fifth, existing large language models are prone to randomness and instability. While large language models possess strong text understanding and reasoning capabilities, directly inputting meteorological data or some environmental parameters into the model can easily lead to inconsistencies in interpretation, incomplete physical chains, omissions of key factors, and unstable risk representations. Existing methods have not fully addressed the coordination issues between meteorological physical constraints, parameter matrix input formats, trigger chain reasoning rules, and structured output templates, making it difficult to guarantee the model's stable and controllable generation of interpretive conclusions in operational early warning scenarios. Insufficient expressiveness makes it difficult to meet the needs of forecasters for review, correction, and user-oriented interpretive services.

[0010] In summary, existing technologies lack a method for short-term heavy rainfall early warning to improve the precision, intelligence, and operational application capabilities of short-term heavy rainfall early warning. Summary of the Invention

[0011] In view of this, it is necessary to provide a method, device, electronic equipment and medium for short-term heavy rainfall early warning, so as to improve the precision, intelligence and operational application capabilities of short-term heavy rainfall early warning.

[0012] To achieve the above objectives, in a first aspect, the present invention provides a short-term heavy rainfall early warning method, apparatus, electronic device, and medium, comprising: Acquire multi-source meteorological and topographic data for a target area within a preset time period. The multi-source meteorological data includes: real-time observation data from automatic weather stations, radiosonde data, and numerical weather prediction data. The topographic data includes: digital elevation model (DEM) data. The multi-source meteorological and topographic data were subjected to quality control and standardization to obtain quality-controlled automatic station real-time data, radiosonde profile data, numerical prediction field data, and basic topographic feature data. Based on the quality-controlled automatic station real-time data and sounding profile data, a real-time environment dataset for a preset time period is constructed. Based on the numerical forecast field data, a forecast trend environment dataset for multiple time periods in the future is constructed. The real-time environment dataset is updated based on the forecast trend environment dataset to obtain the updated real-time environment dataset. Based on the updated real-time environmental dataset and the topographic data, a short-term heavy precipitation environmental parameter matrix is ​​constructed. The short-term heavy precipitation environmental parameter matrix has a first dimension of diagnostic units, a second dimension of environmental parameters, and a third dimension of rolling diagnostic time and forecast lead time. Based on the short-term heavy precipitation environmental parameter matrix, the trigger window state is determined according to the preset short-term heavy precipitation trigger window discrimination rule; A physical constraint cue word template is constructed, which includes a fixed input format, a fixed analysis order, and a fixed output format, and is used to constrain the large language model to reason according to the meteorological physical chain. The short-term heavy precipitation environmental parameter matrix and the state of the trigger window are input into the physical constraint prompt word template, and the large language model is called to generate an interpretable recognition result. Based on the state of the trigger window and the interpretation recognition result, a short-term heavy rainfall intelligent early warning result is generated.

[0013] In one possible implementation, the real-time observation data from the automatic station includes at least one of the following: temperature, relative humidity, wind direction, wind speed, air pressure, and precipitation. The radiosonde data includes at least one of the following: temperature, dew point temperature, wind direction, wind speed, air pressure, and relative humidity at different altitudes. The numerical forecast data includes: hourly forecasts of temperature, humidity, wind field, geopotential height, vertical velocity, precipitation and / or water vapor flux for the next 24 hours. The digital elevation model (DEM) data includes: DEM elevation, slope, aspect, valley distribution, plateau boundaries, and / or coastal topographic features.

[0014] In one possible implementation, the near-surface atmospheric state signal reflected by the real-time data from automatic stations in the real-time environmental dataset includes at least one of the following: near-surface temperature change rate, humidity change rate, wind direction change, wind speed change, surface convergence index, and local temperature and humidity disturbance. The vertical stratification parameters reflected by the sounding profile data in the real-world environmental dataset include at least one of the following: Convective Active Potential Energy (CAPE), Convective Inhibition Energy (CIN), Lifted Condensation Height (LCL), Free Convection Height (LFC), Lower Layer Stability, Inversion Intensity, Wet Layer Thickness, 0-3 km Vertical Wind Shear, and 0-6 km Vertical Wind Shear.

[0015] In one possible implementation, the thermodynamic parameters in the forecast trend environmental dataset include at least one of CAPE, CIN, LCL, LFC, and lower-level stability; the dynamic parameters include at least one of lower-level wind field, vertical wind shear, surface or lower-level convergence, storm relative helicity (SRH), and vertical velocity; the water vapor parameters include at least one of relative humidity, wet layer thickness, water vapor flux, and water vapor flux convergence; and the stratification parameters include at least one of temperature lapse rate, inversion intensity, and dry intrusion height.

[0016] In one possible implementation, constructing a short-term heavy precipitation environmental parameter matrix based on the updated real-time environmental dataset and the topographic data includes: Extract terrain trigger features of the target area based on the DEM terrain elevation data in the terrain data; The updated real-world environment dataset and terrain-triggered features are fused in parameter space according to their physical meaning to obtain a multi-source environmental parameter set. A short-term heavy precipitation environmental parameter matrix is ​​constructed based on the aforementioned multi-source environmental parameter set.

[0017] In one possible implementation, the fixed analysis order in the physical constraint prompt word template is as follows: thermodynamic condition analysis, water vapor condition analysis, dynamic condition analysis, boundary layer signal analysis, terrain triggering condition analysis, inhibition factor analysis, and uncertainty source analysis. The fixed output format in the physical constraint prompt template includes the following fields: trigger probability, possible trigger time window, main favorable factors, main inhibiting factors, key physical chain, risk enhancement or reduction trend, key focus area, and business prompt text.

[0018] In one possible implementation, the intelligent early warning results for short-term heavy rainfall include: The short-term heavy rainfall intelligent early warning results shall be output in at least one of the following formats: current environmental diagnosis conclusions, risk trends for the next 24 hours, key areas of concern, possible triggering time periods, main triggering mechanisms, main inhibitory factors, sources of uncertainty, and business recommendations; the results shall be output in at least one of the following formats: structured text, tables, graphs, or interface data.

[0019] Secondly, the present invention also provides a short-term heavy rainfall early warning device, comprising: The data acquisition module is used to acquire multi-source meteorological data and topographic data of the target area within a preset time period. The multi-source meteorological data includes: real-time observation data from automatic weather stations, radiosonde data, and numerical weather prediction data. The topographic data includes: digital elevation model (DEM) data. The data standardization processing module is used to perform quality control and standardization processing on the multi-source meteorological data and topographic data to obtain quality-controlled automatic station real-time data, radiosonde profile data, numerical prediction field data and basic topographic feature data. The real-world environment dataset acquisition module is used to construct a real-world environment dataset within a preset time period based on the quality-controlled automatic station real-world data and sounding profile data, construct a forecast trend environment dataset with multiple timeframes in the future based on the numerical forecast field data, and update the real-world environment dataset based on the forecast trend environment dataset to obtain the updated real-world environment dataset. The short-term heavy precipitation environmental parameter matrix construction module is used to construct a short-term heavy precipitation environmental parameter matrix based on the updated real-time environmental dataset and the topographic data. The short-term heavy precipitation environmental parameter matrix has a first dimension of diagnostic units, a second dimension of environmental parameters, and a third dimension of rolling diagnostic time and forecast lead time. The trigger window state determination module is used to determine the trigger window state based on the short-term heavy precipitation environmental parameter matrix and according to the preset short-term heavy precipitation trigger window discrimination rules. The prompt word template construction module is used to construct physical constraint prompt word templates. The physical constraint prompt word templates include fixed input format, fixed analysis order and fixed output format, which are used to constrain the large language model to reason according to the meteorological physical chain. The interpretable recognition result generation module is used to input the short-term heavy precipitation environmental parameter matrix and the state of the trigger window into the physical constraint prompt word template, and call the large language model to generate interpretable recognition results; The early warning module is used to generate intelligent early warning results for short-term heavy rainfall based on the state of the trigger window and the interpreted recognition results.

[0020] Thirdly, the present invention also provides an electronic device, including a memory and a processor, wherein, The memory is used to store programs; The processor, coupled to the memory, is used to execute the program stored in the memory to implement the steps in the short-term heavy rainfall early warning method described in any of the above implementations.

[0021] Fourthly, the present invention also provides a computer-readable storage medium for storing a computer-readable program or instructions, which, when executed by a processor, can implement the steps of the short-term heavy rainfall early warning method described in any of the above implementations.

[0022] The beneficial effects of this invention are as follows: This invention provides a short-term heavy precipitation early warning method that acquires multi-source meteorological and topographic data for a target area within a preset time period. The multi-source meteorological data includes: automatic weather station observation data, radiosonde data, and numerical weather prediction data. The topographic data includes: digital elevation model (DEM) data. The multi-source meteorological and topographic data undergo quality control and standardization processing to obtain quality-controlled automatic weather station observation data, radiosonde profile data, numerical weather prediction field data, and basic topographic feature data. Based on the quality-controlled automatic weather station observation data and radiosonde profile data, a real-time environmental dataset for the preset time period is constructed. Based on the numerical weather prediction field data, a forecast trend environmental dataset with multiple timeframes for future periods is constructed. The real-time environmental dataset is updated based on the forecast trend environmental dataset to obtain an updated real-time environmental dataset. Based on the updated real-time environmental dataset and the topographic data, a short-term heavy precipitation environmental parameter matrix is ​​constructed. This method can reduce the physical bias caused by traditional interpolation, resampling, and unified grid processing, and better preserves... This invention preserves the true environmental characteristics reflected by different data sources. The short-term heavy precipitation environmental parameter matrix uses diagnostic units as the first dimension, environmental parameters as the second dimension, and rolling diagnostic time and forecast lead time as the third dimension. Based on this short-term heavy precipitation environmental parameter matrix, the trigger window state is determined according to preset short-term heavy precipitation trigger window discrimination rules. A physical constraint prompt word template is constructed, which includes a fixed input format, a fixed analysis order, and a fixed output format. This template is used to constrain the large language model to reason according to the meteorological physical chain. The short-term heavy precipitation environmental parameter matrix and the trigger window state are input into the physical constraint prompt word template. The large language model is then called to generate interpretable recognition results. Based on the trigger window state and the interpretable recognition results, a short-term heavy precipitation intelligent early warning result is generated. By using a fixed input parameter matrix, a fixed inference framework, and a fixed output format, the inference process of the large language model is physically constrained, reducing random generation, generalized expressions, and non-physical interpretations, making its output more stable, consistent, and traceable. This invention transforms automatic weather station data, radiosonde data, numerical forecasts, and topographic data into a physically consistent environmental parameter system. It can achieve rolling diagnosis of short-term heavy precipitation triggering environment through collaborative updates of real-time environment and forecast trends, and generate stable, readable, and traceable interpretable identification results using a large language model, thereby improving the refinement, intelligence and operational application capabilities of short-term heavy precipitation early warning. Attached Figure Description

[0023] Figure 1 A flowchart illustrating an embodiment of a short-term heavy rainfall early warning method provided by the present invention; Figure 2 This is a schematic flowchart of an embodiment of a short-term heavy rainfall early warning device provided by the present invention; Figure 3A schematic diagram of an embodiment of the electronic device provided by the present invention. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0025] In the description of the embodiments of the present invention, unless otherwise stated, "multiple" means two or more. "And / or" describes the relationship between related objects, indicating that there can be three relationships. For example, A and / or B can represent three situations: A exists alone, A and B exist simultaneously, and B exists alone.

[0026] The terms "first," "second," etc., used in the embodiments of this invention are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a technical feature defined with "first" or "second" may explicitly or implicitly include at least one of that feature.

[0027] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0028] Before demonstrating the embodiments, the following terms will be explained.

[0029] CAPE, short for Convective Available Potential Energy, refers to the total positive work done by the environment on an air parcel as it rises from its free convection height to its equilibrium height under the influence of buoyancy. It is a core indicator for measuring the potential unstable energy in the atmosphere that can be used for convection development, and a higher value is more conducive to the outbreak of severe convective weather.

[0030] CIN, short for Convective Inhibition, refers to the energy consumed by a surface air parcel to overcome the negative buoyancy zone before it rises to the free convection height. It represents the obstruction of convection development by the stable stratification of the lower layer. When the value is too high, the air parcel is unlikely to break through the stratification and trigger convection.

[0031] LCL, short for Lifted Condensation Level, refers to the critical height at which unsaturated air parcels near the ground reach saturation and begin to condense water vapor after being lifted by external forces and cooled adiabatically. It is an important reference parameter for determining cloud base height.

[0032] LFC, short for Level of Free Convection, refers to the lowest altitude at which an air mass, after rising and breaking through the convection inhibition zone, can accelerate upward autonomously due to its temperature being higher than the ambient temperature and positive buoyancy. The lower this altitude, the easier it is to trigger deep convection activity.

[0033] SRH, short for Storm Relative Helicity, is a core dynamic parameter in the diagnosis of severe convection. It refers to the component of the three-dimensional vorticity vector in the direction of the storm relative to the inflow. Essentially, it quantitatively measures the degree of rotation of the environmental wind field within a certain air layer and the total amount of environmental vorticity that can be input into the convective storm.

[0034] This invention provides a method, device, electronic equipment, and medium for short-term heavy rainfall early warning, which are described below.

[0035] Figure 1 This is a schematic flowchart of an embodiment of the short-term heavy rainfall early warning method provided by the present invention, as shown below. Figure 1 As shown, short-term heavy rainfall warning methods include: S101. Obtain multi-source meteorological data and topographic data of the target area within a preset time period. The multi-source meteorological data includes: real-time observation data from automatic weather stations, radiosonde data, and numerical weather prediction data. The topographic data includes: digital elevation model (DEM) data. S102. Perform quality control and standardization processing on the multi-source meteorological data and topographic data to obtain quality-controlled automatic station real-time data, radiosonde profile data, numerical forecast field data and basic topographic feature data. S103. Based on the real-time data from the automatic weather station and the sounding profile data after quality control, construct a real-time environment dataset within a preset time period. Based on the numerical forecast field data, construct a forecast trend environment dataset with multiple timeframes for future periods. Update the real-time environment dataset based on the forecast trend environment dataset to obtain the updated real-time environment dataset. S104. Based on the updated real-time environmental dataset and the terrain data, construct a short-term heavy precipitation environmental parameter matrix. The short-term heavy precipitation environmental parameter matrix has a first dimension of diagnostic units, a second dimension of environmental parameters, and a third dimension of rolling diagnostic time and forecast lead time. S105. Based on the short-term heavy rainfall environmental parameter matrix, determine the trigger window status according to the preset short-term heavy rainfall trigger window discrimination rule; S106. Construct a physical constraint cue word template, which includes a fixed input format, a fixed analysis order, and a fixed output format, to constrain the large language model to reason according to the meteorological physical chain. S107. Input the short-term heavy precipitation environmental parameter matrix and the state of the trigger window into the physical constraint prompt word template, and call the large language model to generate an interpretable recognition result; S108. Generate a short-term heavy rainfall intelligent early warning result based on the state of the trigger window and the interpretation recognition result.

[0036] Compared with existing technologies, this embodiment provides a short-term heavy precipitation early warning method that acquires multi-source meteorological and topographic data for a target area within a preset time period. The multi-source meteorological data includes automatic weather station observation data, radiosonde data, and numerical weather prediction data. The topographic data includes digital elevation model (DEM) data. The method performs quality control and standardization on the multi-source meteorological and topographic data to obtain quality-controlled automatic weather station observation data, radiosonde profile data, numerical weather prediction field data, and basic topographic feature data. Based on the quality-controlled automatic weather station observation data and radiosonde profile data, a real-time environmental dataset for the preset time period is constructed. Based on the numerical weather prediction field data, a forecast trend environmental dataset with multiple timeframes for future periods is constructed. The real-time environmental dataset is updated based on the forecast trend environmental dataset to obtain an updated real-time environmental dataset. Based on the updated real-time environmental dataset and the topographic data, a short-term heavy precipitation environmental parameter matrix is ​​constructed. This method can reduce the physical bias caused by traditional interpolation, resampling, and unified grid processing, and better preserves... The short-term heavy precipitation environmental parameter matrix reflects the real environmental characteristics of different data sources. It uses diagnostic units as the first dimension, environmental parameters as the second dimension, and rolling diagnostic time and forecast lead time as the third dimension. Based on this matrix, the trigger window state is determined according to preset short-term heavy precipitation trigger window discrimination rules. A physical constraint prompt word template is constructed, which includes a fixed input format, a fixed analysis order, and a fixed output format. This template constrains the large language model to reason according to the meteorological physical chain. The short-term heavy precipitation environmental parameter matrix and the trigger window state are input into the physical constraint prompt word template. The large language model is then called to generate interpretable recognition results. Based on the trigger window state and the interpretable recognition results, a short-term heavy precipitation intelligent early warning result is generated. By using a fixed input parameter matrix, a fixed inference framework, and a fixed output format, the inference process of the large language model is physically constrained, reducing random generation, generalized expressions, and non-physical interpretations, making its output more stable, consistent, and traceable. This invention transforms automatic weather station data, radiosonde data, numerical forecasts, and topographic data into a physically consistent environmental parameter system. It can achieve rolling diagnosis of short-term heavy precipitation triggering environment through collaborative updates of real-time environment and forecast trends, and generate stable, readable, and traceable interpretable identification results using a large language model, thereby improving the refinement, intelligence and operational application capabilities of short-term heavy precipitation early warning.

[0037] The short-term heavy rainfall warning method provided in this embodiment can be applied to a short-term heavy rainfall warning system, which can be a software system running on a terminal device. The terminal device can be a tablet computer, in-vehicle device, augmented reality (AR) / virtual reality (VR) device, laptop computer, ultra-mobile personal computer (UMPC), netbook, personal digital assistant (PDA), mobile phone, etc. This embodiment does not impose any restrictions on the specific type of terminal device.

[0038] In a specific embodiment of the present invention, the short-term heavy rainfall early warning method specifically includes the following steps: Step 1: Acquire multi-source meteorological and topographic data for the target area Multi-source data for the target area during and before the designated diagnostic time are acquired, including hourly observations from automatic weather stations, radiosonde data, numerical weather prediction data, and topographic elevation data. Hourly observations from automatic weather stations include parameters such as temperature, relative humidity, wind direction, wind speed, air pressure, and precipitation. Radiosonde data includes vertical profiles of temperature, dew point temperature, wind direction, wind speed, air pressure, and relative humidity at different altitudes. Numerical weather prediction data includes forecast fields for temperature, humidity, wind field, geopotential height, vertical velocity, precipitation, and water vapor flux for the next 24 hours, broken down into 3-hourly intervals. Topographic elevation data includes DEM elevation, slope, aspect, valley distribution, plateau boundaries, and coastal topographic features. All of these data serve as the raw input data for the short-duration heavy precipitation environmental diagnosis.

[0039] In some embodiments of the present invention, the real-time observation data from the automatic weather station includes at least one of the following: temperature, relative humidity, wind direction, wind speed, air pressure, and precipitation. The radiosonde data includes at least one of the following: temperature, dew point temperature, wind direction, wind speed, air pressure, and relative humidity at different altitudes. The numerical forecast data includes: hourly forecasts of temperature, humidity, wind field, geopotential height, vertical velocity, precipitation and / or water vapor flux for the next 24 hours. The digital elevation model (DEM) data includes: DEM elevation, slope, aspect, valley distribution, plateau boundaries, and / or coastal topographic features.

[0040] Step 2: Perform quality control and standardization processing on multi-source data. Quality control is performed on the data acquired in Step 1. For automatic weather station data, missing values, anomalous jump values, and observations exceeding reasonable meteorological ranges are removed, and consistency checks are performed on continuous time data. For radiosonde data, the integrity of vertical layers is checked, and any anomalous abrupt changes in temperature, humidity, and wind profiles are examined. The radiosonde data is then organized into a unified vertical hierarchical structure. For numerical weather prediction data, 3-hourly forecast fields for the next 24 hours within the target area are extracted, and variable names, units, and time identifiers are standardized. For DEM data, cropping, coordinate standardization, and spatial range matching are performed. This step outputs the quality-controlled automatic weather station data, radiosonde profile data, numerical weather prediction field data, and basic topographic feature data.

[0041] Step 3: Constructing a Real-World IC Dataset Based on the automated weather station (AWS) data and radiosonde data after quality control, a real-time environmental dataset for the diagnostic time is constructed. Specifically, this includes: calculating near-surface temperature change rate, humidity change rate, wind direction change, wind speed change, surface convergence index, and local temperature and humidity disturbances using AWS data on temperature, humidity, wind direction, wind speed, and air pressure; calculating environmental parameters such as CAPE, CIN, LCL, LFC, lower-level stability, inversion intensity, wet layer thickness, 0–3 km vertical wind shear, and 0–6 km vertical wind shear using radiosonde data; and combining the rapid near-surface change signals reflected by the AWS with the vertical stratification structure reflected by the radiosonde data to form the current true atmospheric state. This step outputs the real-time environmental dataset.

[0042] In some embodiments of the present invention, the near-surface atmospheric state signal reflected by the real-time data from automatic stations in the real-time environmental dataset includes at least one of the following: near-surface temperature change rate, humidity change rate, wind direction change, wind speed change, surface convergence index, and local temperature and humidity disturbance. The vertical stratification parameters reflected by the sounding profile data in the real-world environmental dataset include at least one of the following: Convective Active Potential Energy (CAPE), Convective Inhibition Energy (CIN), Lifted Condensation Height (LCL), Free Convection Height (LFC), Lower Layer Stability, Inversion Intensity, Wet Layer Thickness, 0-3 km Vertical Wind Shear, and 0-6 km Vertical Wind Shear.

[0043] Step 4: Construct the Forecast Trend Environment FC Dataset Based on numerical weather prediction data, a forecast trend environmental dataset for the next 24 hours is constructed, with each 3-hour interval. Specifically, this includes extracting thermal, dynamic, water vapor, and stratification parameters from the numerical weather prediction field on a time-by-time basis. Thermal parameters include CAPE, CIN, LCL, LFC, and lower-level stability; dynamic parameters include lower-level wind field, vertical wind shear, surface or lower-level convergence, SRH, and vertical velocity; water vapor parameters include relative humidity, wet layer thickness, water vapor flux, and water vapor flux convergence; and stratification parameters include temperature lapse rate, inversion intensity, and dry intrusion height. These parameters are organized according to a 0-24 hour timeframe, with each 3-hour interval, to form an FC dataset reflecting future environmental evolution trends.

[0044] In some embodiments of the present invention, the thermodynamic parameters in the forecast trend environmental dataset include at least one of CAPE, CIN, LCL, LFC, and lower-level stability; the dynamic parameters include at least one of lower-level wind field, vertical wind shear, surface or lower-level convergence, storm relative helicity (SRH), and vertical velocity; the water vapor parameters include at least one of relative humidity, wet layer thickness, water vapor flux, and water vapor flux convergence; and the stratification parameters include at least one of temperature lapse rate, inversion intensity, and dry intrusion height.

[0045] Step 5: Establish a rolling diagnostic chain for coordinated updates between IC and FC. The real-world environment dataset obtained in step three is time-linked with the forecast trend environment dataset obtained in step four to form a rolling diagnostic chain of "current real-world environment - future forecast trend". Specifically, the IC (Indexed Indicator) at the current diagnostic time is used as the initial environmental state, and the forecast field data for the next 24 hours, updated every 3 hours, is used as the trend extension. Whenever new real-world data from automatic weather stations arrives, the real-world dataset is updated according to a set period, and the timeliness of the future forecast dataset is re-matched, forming a new rolling diagnostic chain. This process does not change the numerical forecast field itself, but rather, through real-world environment replacement and trend re-linking, enables the system to continuously identify the formation, intensification, maintenance, or weakening of short-term heavy precipitation triggering conditions.

[0046] In a specific embodiment of the present invention, the update cycle is set to 3 hours, that is, the real-time dataset is updated every 3 hours. The numerical forecast is updated twice a day, at 8:00 am and 5:00 pm. The forecast at 8:00 am is for 11:00 am, 2:00 pm, 5:00 pm, 8:00 pm, 11:00 pm, 2:00 pm, and 5:00 pm. Starting at 5:00 pm, the new numerical forecast and real-time dataset are used to predict the conditions at 8:00 pm, 11:00 pm, 2:00 pm, 5:00 pm, 8:00 pm, 11:00 pm, 2:00 pm, and 5:00 pm.

[0047] Step 6: Extract terrain and underlying surface trigger features Based on DEM topographic elevation data, topographic and underlying surface triggering features of the target area are extracted. Specifically, this includes: calculating slope and aspect based on elevation data; identifying valley attributes based on relative elevation differences and runoff patterns; identifying plateau boundaries based on elevation abrupt changes and topographic boundaries; identifying coastal ventilation corridors based on topographic orientation and coastal openness; and calculating orographic uplift potential based on the angle between low-level wind direction and slope aspect. This step outputs slope factor, valley index, plateau boundary factor, coastal ventilation corridor factor, and orographic uplift potential factor for subsequent local triggering factor identification.

[0048] Step 7: Construct a multi-source environmental parameter set using a parameter space fusion method. This method fuses real-time datasets, forecast datasets, and terrain-triggered features in a parameter space according to their physical meaning, without resorting to traditional methods such as unified gridding, temporal interpolation, or spatial resampling. Specifically, it uses stations, regional feature points, or pre-defined diagnostic units (e.g., grid points) as objects, unifying near-surface real-time parameters from automatic weather stations, vertical stratigraphic parameters from radiosondes, future trend parameters from numerical weather prediction, and terrain-triggered parameters from DEMs into a single parameter set. Different data sources retain their original physical meanings, establishing correspondences only at the parameter level. This step outputs a multi-source environmental parameter set for short-duration heavy precipitation identification.

[0049] In some embodiments of the present invention, constructing a short-term heavy precipitation environmental parameter matrix based on the real-time environmental dataset, the forecast trend environmental dataset, and the topographic data includes: Extract terrain trigger features of the target area based on the DEM terrain elevation data in the terrain data; The real-world dataset, forecast dataset, and terrain-triggered features are fused in parameter space according to their physical meaning to obtain a multi-source environmental parameter set. A short-term heavy precipitation environmental parameter matrix is ​​constructed based on the aforementioned multi-source environmental parameter set.

[0050] Step 8: Construct the environmental parameter matrix for short-duration heavy precipitation Based on the multi-source environmental parameter set obtained in step seven, a short-term heavy precipitation environmental parameter matrix is ​​constructed. This matrix uses diagnostic units as rows, environmental parameters as columns, and rolling diagnostic time and future forecast lead time as time dimensions. The matrix fields include at least thermal parameters, dynamic parameters, water vapor parameters, boundary layer parameters, and topographic parameters. Thermal parameters include CAPE, CIN, LCL, LFC, and lower-level stability; dynamic parameters include surface convergence index, lower-level wind shear, 0–3 km wind shear, 0–6 km wind shear, and SRH; water vapor parameters include moist layer thickness, water vapor flux, water vapor flux convergence, and relative humidity; boundary layer parameters include near-surface temperature and humidity disturbances, wind direction change, land-sea wind front index, and local convergence enhancement signal; and topographic parameters include slope, valley index, topographic lifting potential, plateau boundary factor, and coastal ventilation corridor factor. This matrix serves as the structured input for subsequent trigger window discrimination and intelligent interpretation reasoning.

[0051] Step 9: Identify regional triggering factors in complex underlying surfaces Based on the short-duration heavy precipitation environmental parameter matrix obtained in step eight, local triggering factors in the target area are identified. Specifically, these include: calculating the land-sea wind front index based on signals such as the near-surface wind direction shift from land to sea or other favorable directions, temperature decrease, humidity increase, and enhanced surface convergence; calculating the valley convergence enhancement factor based on valley attributes, near-surface wind convergence, wind speed changes, and accumulated humidity; calculating the orographic lifting potential based on slope, aspect, the angle between low-level wind direction and slope aspect, and humidity conditions; and calculating the coastal ventilation corridor factor based on the consistency between the coastal ventilation corridor and the direction of low-level water vapor transport. These factors collectively characterize the effects of land-sea wind front advancement, enhanced valley convergence, and orographic forcing lifting on the triggering of short-duration heavy precipitation.

[0052] To ensure the repeatability of the thresholds, normalization scales, and weights described below, this embodiment establishes a reference sample set D_ref. D_ref includes all valid hourly samples from the month of the target diagnosis and the month before and after it within the past five full years; if the data is less than five years, all available data are used. P_p(X) represents the p-th percentile value of variable X in all finite valid samples in D_ref. To avoid a denominator of 0, this embodiment sets strictly greater than 0 normalization lower limits: ε_T, ε_RH, ε_C, ε_WS, and ε_Q represent the minimum allowable values ​​for the normalization scales of temperature, relative humidity, surface convergence, wind speed, and water vapor transport, respectively, and are taken as 0.1℃, 1 percentage point, 10^-6 s^-1, 0.1 m / s, and 10^-4 kg / kg·m / s.

[0053] When D_ref covers at least 3 complete years, Δτ = 3 hours is taken; T_0 = max{P_75[max(0,T_i(t-Δτ)-T_i(t))],ε_T}; RH_0=max{P_75[max(0,RH_i(t)-RH_i(t-Δτ))],ε_RH}; C_0=max{P_75[max(0,C_i(t))],ε_C}; WS_0=max{P_75[max(0,WS_i(t-Δτ)-WS_i(t))],ε_WS}; WS_min=max{P_25[WS_i(t)|WS_i(t)>0],ε_WS}; WS_ref=max{P_75[WS_low(i,t)],ε_WS}; Q_0=max{P_75[|Q_L(i,t)|],ε_Q}; RH_min=max{P_25[RH_low(i,t)],70%}, RH_trigger=max{P_75[RH_i(t)],80%}.

[0054] When D_ref is less than 3 full years, the default values ​​of T_0=2℃, RH_0=10 percentage points, C_0=10^-4 s^-1, WS_0=2 m / s, WS_min=1 m / s, WS_ref=10 m / s, Q_0=0.12 kg / kg·m / s, RH_min=70%, and RH_trigger=80% are used. Business units may use local thresholds validated with independent historical samples instead of the above values, but should record the threshold versions and keep them unchanged within the same rolling diagnostic cycle.

[0055] The formula for calculating the land-sea front index is as follows: S_dir(i,t) = max_{φ∈Ω_i}{[A(ψ_i(t-Δτ), φ+180°) × A(ψ_i(t), φ)]^(1 / 2)} (1) S_T(i,t) = clip{[T_i(t-Δτ)-T_i(t)] / T_0} (2) S_H(i,t) = clip{[RH_i(t)-RH_i(t-Δτ)] / RH_0} (3) SBFI_i(t) = 100 × [0.35S_dir(i,t) + 0.20S_T(i,t) + 0.20S_H(i,t) +0.25S_C^+(i,t)] (4) In the formula, i represents the diagnostic unit or station number, t represents the current diagnostic time, and Δτ represents the backtracking time interval; unless otherwise specified, the azimuth angles in this paper are all set with true north as 0° and increase in a clockwise direction, and mod represents the modulo operation; φ represents the candidate sea breeze inflow direction, φ_0,i represents the azimuth angle from the coastline point closest to diagnostic unit i to diagnostic unit i, and Ω_i represents the set of candidate sea breeze inflow directions obtained by discretizing φ_0,i as the center, with 22.5° to the left and right and 5° intervals; when the diagnostic unit i is more than 100 kilometers away from the coastline, let SBFI_i(t)=0; ψ_i(t) represents the near-surface airflow direction angle of diagnostic unit i at time t; if the input data uses meteorological wind direction, i.e., the angle of arrival of the wind, then ψ_i(t) = [meteorological wind direction + 180°] mod The transformation is performed using 360°; d(θ_1,θ_2)=min{|θ_1-θ_2|,360°-|θ_1-θ_2|} represents the minimum angle between two direction angles; A(θ_1,θ_2)=max{0,cos[d(θ_1,θ_2)]} represents the direction consistency function, and the angle is converted to radians when calculating the cosine value; S_dir(i,t) represents the normalized evaluation of the change in airflow from the offshore direction at time t-Δτ to the candidate sea breeze inflow direction at time t. The scores are: T_i(t) and RH_i(t) represent the air temperature and relative humidity of diagnostic unit i at time t, respectively; T_0 and RH_0 represent the normalized scale of temperature drop and the normalized scale of humidity increment, respectively; clip(x) = min{1, max{0, x}} indicates that the input value is truncated to the interval [0, 1]; S_T(i, t) and S_H(i, t) represent the score for temperature drop and the score for relative humidity increase, respectively; the ground convergence index C_i(t) is calculated according to C_i(t) = -( 1 / N)×Σ_{j∈N_i}{[V_j(t)-V_i(t)]·r_{i→j} / D_ij} is calculated, where N_i is the neighborhood set described in equation (21), N is the number of neighborhood diagnostic units, V_i(t) and V_j(t) are the horizontal wind vectors of diagnostic units i and j respectively, r_{i→j} is the unit direction vector from i to j, D_ij is the horizontal distance between i and j in meters, a positive value indicates convergence; C_0 represents the positive convergence normalization scale, S _C^+(i,t)=clip[max{0,C_i(t)} / C_0] represents the positive convergence score; 0.35, 0.20, 0.20 and 0.25 in equation (13) represent the preset weights of the four sub-scores of direction change, temperature decrease, relative humidity increase and positive convergence, respectively, and the sum of the weights is 1; SBFI_i(t) represents the land-sea front index, and the coefficient 100 in the equation is used to scale the weighted result to 0 to 100. The larger the value, the stronger the land-sea front trigger signal.

[0056] The formula for calculating the valley convergence enhancement factor is as follows: G_dir(i,t) = clip{(1 / N) × Σ_{j=1}^N max[0, u_j(t) · r_{j→i}]} (5) S_dec(i,t) = I[WS_i(t) ≥ WS_min] × clip{[WS_i(t-Δτ)-WS_i(t)] / WS_0} (6) H_acc(i,t) = clip{Σ_{k=1}^M max[0, RH_i(t_k)-RH_i(t_{k-1})] / RH_0} (7) Valley_Conv_i(t) = 100 × K_valley,i × [0.40G_dir(i,t) + 0.30S_dec(i,t) + 0.30H_acc(i,t)] (8) K_valley,i = clip[D_valley,i / D_0] × clip[R_valley,i / R_0] (9) D_valley,i = H_rim,i - H_floor,i (10) R_valley,i = L_valley,i / max(W_valley,i, ε_L) (11) N_i = {j | j≠i,D_ij ≤ R_N},N = |N_i| (12) In the formula, G_dir(i,t) represents the wind direction inflow score of diagnosis unit i at time t; N_i represents the set of adjacent stations or adjacent diagnosis units involved in the calculation, N represents the number of diagnosis units in the set, j represents the serial number of the adjacent diagnosis unit; u_j(t) represents the unit vector pointing to the airflow destination after the horizontal wind vector of diagnosis unit j is normalized; when WS_j(t) < WS_min or the wind data is missing, this diagnosis unit does not participate in the summation and N is updated accordingly, r_{j→i} represents the unit direction vector pointing from diagnosis unit j to diagnosis unit i, the symbol "·" represents the vector dot product; D_ij represents the horizontal distance between diagnosis unit i and j, R_N represents the neighborhood radius, in this embodiment, R_N is taken as 30 kilometers; when there are fewer than 3 valid adjacent diagnosis units within 30 kilometers, the search radius is extended to 60 kilometers, and if the number is still less than 3, set G_dir(i,t)=0 and output an insufficient data flag; S_dec(i,t) represents the wind speed attenuation score, WS_i(t) represents the near-surface wind speed of diagnosis unit i at time t, WS_min represents the minimum valid wind speed threshold, WS_0 represents the wind speed attenuation normalization scale; I[condition] represents the indicator function, which takes the value of 1 when the condition is satisfied, and 0 otherwise; H_acc(i,t) represents the humidity accumulation score within the statistical window; t_k and t_{k-1} represent two adjacent sampling times, k represents the serial number of the humidity increment, M represents the number of adjacent valid humidity increments within the interval [t-Δτ, t]; when the data is at equal time intervals and there is no missing measurement, M=Δτ / δt, where δt represents the data time resolution, for hourly data and a 3-hour statistical window, M=3; RH_i(t_k) represents the relative humidity of diagnosis unit i at time t_k, RH_0 represents the humidity increment normalization scale; Valley_Conv_i(t) represents the valley convergence enhancement factor, with a value range from 0 to 100; 0.40, 0.30 and 0.30 represents the preset weights for wind direction inflow, wind speed attenuation, and humidity accumulation; K_valley,i represents the valley topography coefficient, D_valley,i represents the relative valley cutting depth (D_0 is 100 meters), and R_valley,i represents the valley aspect ratio (R_0 is 3). Within a DEM window centered on diagnostic unit i with a radius of 10 kilometers, connected grids with elevations no higher than the 20th percentile and slopes no greater than 5° are defined as valley floor grids. H_floor,i represents the median elevation of the valley floor area containing diagnostic unit i or its nearest connected valley floor region. The data is analyzed as follows: Grids with elevations not lower than the 80th percentile within the same window are defined as valley edge grids, where H_rim,i represents the median elevation of the valley edge grids. The center coordinates of each grid within the valley floor connected region are projected onto the first and second principal axes obtained from principal component analysis, respectively. L_valley,i is the difference between the maximum and minimum values ​​of the projected coordinates of the first principal axis, W_valley,i is the difference between the maximum and minimum values ​​of the projected coordinates of the second principal axis, and ε_L is the side length of a single grid cell in the DEM. When no valley floor connected region meets the conditions, K_valley,i = 0.

[0057] The formula for calculating the potential for topographic uplift is as follows: OLP_i(t) = 100 × S_β(i)× S_WS(i,t) × A_up(i,t) × S_M(i,t) (13) w_oro(i,t) = WS_low(i,t)×sinβ_i × A_up(i,t) × S_M(i,t) (14) OLP_code(t)=(1 / n_code)×Σ_{i∈I_code} OLP_i(t), n_code=|I_code| (15) In the formula, OLP_i(t) represents the topographic lifting potential index of diagnostic unit i at time t, which is also the topographic lifting field written into the short-term heavy precipitation environmental parameter matrix, with a value range of 0 to 100; Slope_PCT75(i) represents the 75th percentile value of the slope percentage within a 10-kilometer radius DEM window centered on diagnostic unit i, β_i=arctan[Slope_PCT75(i) / 100] represents the corresponding slope angle; S_β(i)=clip[sinβ_i / sinβ_0] represents the slope score, which is the highest in this formula. In this example, β_0 = 15°; WS_low(i,t) represents the arithmetic mean of the 925 hPa and 850 hPa wind speeds, using the available layer if one is missing; WS_ref represents the normalized scale of the low-level wind speed; S_WS(i,t) = clip[WS_low(i,t) / WS_ref] represents the low-level wind speed score; for grids with a slope of not less than 2° within the DEM window, a circular average is performed on their downhill direction with the slope as the weight, and the resulting direction is denoted as α_aspect,i; when the magnitude of the composite vector of the circular average is less than 0.When 1, use the downhill direction of the grid with the largest slope within the window; if a valid direction still cannot be obtained, decrement A_up(i,t) as a missing value; α_up,i=(α_aspect,i+180°) mod 360° represents the uphill direction; V_925(i,t) and V_850(i,t) represent the horizontal wind vectors at 925 hPa and 850 hPa respectively; ψ_low,i(t)=Dir{[V_925(i,t)+V_850(i,t)] / 2} represents the average airflow direction angle of the lower layer after the synthesis of the two horizontal wind vectors, and the available layer is used when one layer is missing; A_up(i,t)=max{0,cos[d(ψ_low,i(t),α_up,i)]} represents the consistency coefficient between the lower airflow and the uphill direction; RH_low(i,t) represents the arithmetic mean of the relative humidity at 925 hPa and 850 hPa, and the available layer is used when one layer is missing; RH_min represents the lower limit of relative humidity at which the scoring of lower-level water vapor sufficiency begins; S_M(i,t)=clip{[RH_ `low(i,t)-RH_min] / max[100%-RH_min,ε_RH]}` represents the low-level water vapor sufficiency score; `w_oro(i,t)` represents the estimated topographic forced lifting velocity in meters per second, used as an auxiliary physical quantity of `OLP_i(t)` rather than as a separate matrix field; `code` represents the unique code of the warning grid or administrative region, `I_code` represents the set of valid diagnostic units covered by this code, `n_code=|I_code|`, and `OLP_code(t)` represents the arithmetic mean of the `OLP_i(t)` of each valid diagnostic unit within the coded region; when the 925 hPa and 850 hPa wind fields are both missing, the relative humidity is both missing, or `A_up(i,t)` cannot be determined, both `OLP_i(t)` and `w_oro(i,t)` output a missing flag, not replaced by 0.

[0058] The formula for calculating the coastal ventilation corridor factor is as follows: Q_L(i,t)=Σ_{k=1}^Kω_k q_k(i,t) V_k(i,t), and Σ_{k=1}^Kω_k= 1 (16) ψ_Q(i,t) = Dir[Q_L(i,t)]; A_corr(i,t) = max{0, cos[d(ψ_Q(i,t), α_corr,i)]} (17) S_Q(i,t) = clip[|Q_L(i,t)| / Q_0] (18) CVCF_i(t) = 100 × K_i × A_corr(i,t) × S_Q(i,t) (19) In the formula, Q_L(i,t) represents the low-level integrated water vapor transport vector of diagnostic unit i at time t; K represents the number of low-level vertical layers participating in the synthesis. In this embodiment, three layers of 925, 850, and 700 hPa are selected, so K=3; k represents the layer number; q_k(i,t) represents the specific humidity of the k-th layer in kilograms per kilogram; V_k(i,t) represents the horizontal wind vector of the k-th layer in meters per second. Therefore, the unit of Q_L(i,t) is kilograms per kilogram multiplied by meters per second; the weights ω_k corresponding to the three layers are 0.50, 0.30, and 0.20 respectively; if a layer is missing data, the layer is deleted and the remaining weights are adjusted accordingly. The sum of their weights is renormalized to make Σω_k=1; Q_L(i,t) outputs a missing measurement flag when all three levels are missing; Dir[·] represents the vector airflow direction angle obtained by the atan2 function, and ψ_Q(i,t) represents the direction of low-level integrated water vapor transport; coastal ventilation corridors are extracted from the DEM window with a radius of 50 kilometers centered on diagnostic unit i: grids with an elevation not exceeding 200 meters and a slope not exceeding 5° are used as candidate corridor grids. If the connected area of ​​the candidate grids simultaneously contacts the coastline and extends to a position not exceeding 2 kilometers from diagnostic unit i, then I_conn,i=1, otherwise I_conn, i=0; α_corr,i represents the direction of the first principal axis obtained by principal component analysis of the connected region, and the direction is uniformly pointed to the land side; the center of each grid in the corridor connected region is projected onto the first principal axis, and L_corr,i is the difference between the maximum and minimum values ​​of the projected coordinates; A_corridor,i represents the total area of ​​the grid in the corridor connected region, and W_corr,i=A_corridor,i / max(L_corr,i,ε_L) represents the equivalent average width; the corridor effectiveness coefficient is calculated according to K_i=I_conn,i×clip[L_corr,i / L_0]×cl ip[W_corr,i / W_0] is calculated. In this embodiment, L_0 = 20 km and W_0 = 5 km. d(·) represents the minimum angle between two direction angles. A_corr(i,t) represents the consistency coefficient between the water vapor transport direction and the corridor axis. |Q_L(i,t)| represents the magnitude of the low-level integrated water vapor transport vector. Q_0 represents the water vapor transport normalization scale. S_Q(i,t) represents the water vapor transport intensity score. CVCF_i(t) represents the coastal ventilation corridor factor, with a value range of 0 to 100. When I_conn,i = 0, both K_i and CVCF_i(t) are 0.

[0059] To ensure consistency between the output fields in the environmental parameter matrix and the early warning examples and the above formula, other near-surface diagnostic indicators are calculated as follows: ΔT_i(t) = T_i(t)-T_i(t-Δτ); ΔTd_i(t) = Td_i(t)-Td_i(t-Δτ) (20) TM_Jump_i(t) = ΔT_i(t) + ΔTd_i(t) (21) BL_Heating_i(t) = ΔT_i(t) / Δτ (22) Moist_Trigger_i(t) = I[RH_i(t) ≥ RH_trigger] (23) Surface_Conv_i(t) = 10^4 × C_i(t) (24) In the formula, Td_i(t) represents the dew point temperature of diagnostic unit i at time t, ΔT_i(t) and ΔTd_i(t) represent the changes in air temperature and dew point temperature within the statistical window, respectively, both in degrees Celsius; TM_Jump_i(t) represents the temperature-humidity jump index, in degrees Celsius, and its value is equal to the sum of the changes in air temperature and dew point temperature; BL_Heating_i(t) represents the boundary layer heating intensity, in degrees Celsius per hour; Moist_Trigger_i(t) represents the humidity trigger index, which is 1 when the relative humidity reaches RH_trigger, otherwise it is 0; Surface_Conv_i(t) represents the surface convergence value for display in the environmental parameter matrix and business text, which is 10^4 times the actual surface convergence index C_i(t), and the reporting unit is 10^-4 s^-1, with positive values ​​indicating convergence.

[0060] The missing data handling rules are as follows: For equation (4), when at least three of the four sub-scores are valid, SBFI_i(t) is recalculated using the original weights of the valid sub-scores; if there are fewer than three, a missing data flag is output. For equation (8), when at least two of the three sub-scores are valid, Valley_Conv_i(t) is recalculated using the original weights of the valid sub-scores; if there are fewer than two, a missing data flag is output. For equations (13) and (14), a missing data flag is output when any multiplication factor is missing. For equation (16), when at least one vertical level is valid, it is recalculated using the remaining weights; if all three levels are missing, Q_L, S_Q, and CVCF are all output with missing data flags. For equations (20) to (24), when any of the required observations is missing, the corresponding index is output with a missing data flag. All missing data results are accompanied by the quality control field QC_Flag=Missing to distinguish them from a physical value of 0.

[0061] The regionalized triggering factors calculated in step nine are written back into the short-term heavy precipitation environmental parameter matrix. The specific matrix form is as follows: In a specific embodiment of the present invention, the final short-term heavy precipitation environmental parameter matrix is ​​as follows: IC Short-Term Heavy Precipitation Environmental Parameter Matrix Time: 2026-07-26 17:00 Statistics window: 3 hours [Site] Hua'an National Basic Meteorological Station (58928) - 3-hour cumulative precipitation: 9.9 mm Current temperature: 25.6 ℃ - Temperature change (within the window): -1.0 ℃ - Current average wind speed: 1.0 m / s - Wind speed variation (within the window): -0.4 m / s - Wind direction: 46.0° - Surface convergence index (Surface_Conv, unit: 10^-4 s^-1): -7.8e-02 (positive value indicates convergence) - Temperature and humidity jump index (TM_Jump): -1.8 (ΔT=-1.0 ℃, ΔTd=-0.8 ℃) - Orographic lifting potential (OLP): Not calculated (low-level wind field and low-level humidity data are not included in the current IC sample). - Valley Convergence Intensity (Valley_Conv): 0.0 - Boundary layer heating intensity (BL_Heating): -0.3 ℃ / h - Moisture Trigger Index: 1.0 -------------------------------------------------- [Station] Anxi National Basic Meteorological Station (58929) - 3-hour cumulative precipitation: 11.0 mm Current temperature: 25.2 ℃ - Temperature change (within the window): -3.6 ℃ - Current average wind speed: 3.8 m / s - Wind speed variation (within the window): -0.2 m / s - Wind direction: 62.0° - Surface convergence index (Surface_Conv, unit: 10^-4 s^-1): 4.4e-01 (positive value indicates convergence) - Temperature and humidity jump index (TM_Jump): -3.7 (ΔT=-3.6 ℃, ΔTd=-0.1 ℃) - Orographic lifting potential (OLP): Not calculated (low-level wind field and low-level humidity data are not included in the current IC sample). - Valley Convergence Intensity (Valley_Conv): 0.0 - Boundary layer heating intensity (BL_Heating): -1.2 ℃ / h - Moisture Trigger Index: 1.0 Step 10: Determine the trigger window for short-term heavy precipitation based on the environmental parameter chain. Based on the environmental parameter matrix obtained in step nine, a rule for determining the trigger window of short-duration heavy precipitation is constructed. The specific method is as follows: To determine whether the CIN (Concentration Intensity Index) weakens, whether the CAPE (Capacity for Emergence) is maintained or strengthened, whether the LCL (Liquidity Coefficient) and LFC (Large Fiber Coefficient) decrease, whether the lower-level moist layer thickens, whether water vapor flux convergence strengthens, whether surface or lower-level convergence intensifies, whether the sea-land wind front advances to the target area, whether valley convergence or orographic lifting signals strengthen, and whether vertical wind shear is conducive to convection organization. When multiple favorable factors synergistically strengthen in the same or adjacent time periods, and the main suppressing factor weakens, the short-term heavy precipitation trigger window is considered to be open. When favorable factors weaken or inhibitory factors strengthen, the trigger window is closed or the risk is reduced. This step outputs the trigger window status, the possible trigger time period, and the main control factors.

[0062] Step 11: Construct a physical constraint prompt template To address the interpretation and recognition process of a large language model, a physical constraint prompt template is constructed. This template includes a fixed input format, a fixed analysis order, and a fixed output format. The input format consists of the environmental parameter matrix generated in step eight and the trigger window discrimination results generated in step ten. The analysis order is as follows: thermal conditions, water vapor conditions, dynamic conditions, boundary layer signals, terrain triggering conditions, suppression factors, and sources of uncertainty. The output format includes trigger probability, potential triggering time window, main favorable factors, main suppression factors, key physical chains, risk enhancement or reduction trends, key areas of concern, and operational prompt text. This template constrains the large language model to reason according to the meteorological physical chain, avoiding randomization and non-physicalization of the output content.

[0063] In some embodiments of the present invention, the fixed analysis order in the physical constraint prompt word template is as follows: thermal condition analysis, water vapor condition analysis, dynamic condition analysis, boundary layer signal analysis, terrain triggering condition analysis, inhibition factor analysis, and uncertainty source analysis. The fixed output format in the physical constraint prompt template includes the following fields: trigger probability, possible trigger time window, main favorable factors, main inhibiting factors, key physical chain, risk enhancement or reduction trend, key focus area, and business prompt text.

[0064] Step 12: Generate interpretable recognition results using a large language model The environmental parameter matrix obtained in step nine and the trigger window discrimination result obtained in step ten are input into the physical constraint prompt word template constructed in step eleven, and the large language model is called for interpretive inference. The large language model does not directly perform numerical weather prediction, but rather judges the probability of short-term heavy precipitation triggering, the triggering time period, the main control factors, and the risk change trend based on the physical relationships between the input parameters. For example, when a combination of signals such as weakening CIN, thickening of the moist layer, enhanced low-level convergence, advancing sea-land wind front, and enhanced potential for orographic lifting occurs, the model generates the interpretive conclusion that "the triggering conditions are tending to open, and the risk of short-term heavy precipitation in the future is increasing." This step outputs the structured interpretation result of the strong convection triggering mechanism.

[0065] Step 13: Generate intelligent early warning results for short-term heavy rainfall Based on the trigger window discrimination results obtained in step ten and the interpreted identification results obtained in step twelve, a short-term heavy precipitation intelligent early warning result is generated. The product includes current environmental diagnostic conclusions, risk trends for the next 24 hours, key areas of concern, possible trigger time periods, main triggering mechanisms, main suppression factors, sources of uncertainty, and operational recommendations. The early warning product can be output in the form of structured text, tables, graphs, or interface data, and can be used for short-term heavy precipitation monitoring and early warning, severe convective weather consultations, grassroots meteorological services, case reviews, and rolling test evaluations.

[0066] In some embodiments of the present invention, the intelligent early warning results for short-term heavy rainfall include: The short-term heavy rainfall intelligent early warning results shall be output in at least one of the following formats: current environmental diagnosis conclusions, risk trends for the next 24 hours, key areas of concern, possible triggering time periods, main triggering mechanisms, main inhibitory factors, sources of uncertainty, and business recommendations; the results shall be output in at least one of the following formats: structured text, tables, graphs, or interface data.

[0067] In a specific embodiment of the present invention, the final warning result is as follows: Chief AI Meteorological Service Analyst, Haicang District Meteorological Bureau | 2026-07-29 14:00 | Convection Potential Analysis

I. IC Real-World Overview

[0068] - The BL_Heating of Xianyue Mountain in Huli District reached 1.2℃ / h, and multiple stations such as Kaiyuan Middle School in Siming District, Xiamen No.1 Middle School, Qinyuan Park in Gulangyu, and the Third Institute of Oceanography in Binhai District had a heating rate of >1.0℃ / h, forming a heating center in the urban area.

[0069] - The temperature rise and humidity increase were significant at multiple points, including Huli Heshan Shuangshi Middle School TM_Jump (ΔT=0.9℃ / ΔTd=3.4℃), Huli Jinshan Xiajinwan 3.1, and Haicang Dongfu Wenpushan 2.2, covering a wide range.

[0070] - The Surface_Conv of Lianqian Liancheng Community in Siming District is -0.27 (unit: 10^-4 s^-1, divergence), while that of Xike in Tongan District is 0.16 (unit: 10^-4 s^-1) and that of Haijiao in Zhenlonghai District is 0.11 (unit: 10^-4 s^-1). Local positive convergence still exists, but its intensity is weak.

[0071] - Triggering condition status: Local establishment (significant heating and humidification in urban areas, but the convergence signal is scattered and has not formed a systematic structure).

[0072] II. FC Forecast - The peak CAPE occurred at 14:00 at the northern grid point (24.75°N, 118.0°E) at 3006.5 J / kg. After 14:00, the CAPE in the whole region remained above 2000 J / kg, and the heat remained strong.

[0073] - Abundant moisture, TCWV between 53-59mm, Q850 generally 12-14g / kg, far exceeding the convection threshold.

[0074] - KI reaches a peak temperature of 35.7℃ at 14:00, and with CAPE > 2000 J / kg, it has a high instability index.

[0075] - At 14:00, the northern grid point of W850 (118.0 / 24.75) was -3.51, indicating a weak upward trend. However, after 17:00, most grid points of W850 turned positive or slightly negative (from -0.12 to 1.26), indicating a weakening of upward momentum. At 17:00, some grid points of W700 turned positive (1.3-5.37), indicating a tendency for the momentum to be suppressed.

[0076] - The first >0.5mm crain3 appeared at 17:00 at the northern grid point (118.0 / 24.75) at 1.52mm, but the entire area returned to zero after 20:00.

[0077] - Forecast support strength: moderate to strong, with a high-risk time window from 14:00 to 17:00 on the 29th (favorable dynamic conditions in the early morning, weakening in the afternoon).

[0078] III. IC Calibration FC - Precipitation timing: No precipitation at IC, FC has a crin3 signal > 0.5mm at 17:00, and the window remains around 17:00.

[0079] - Heating deviation: IC heating rate in urban areas is 1.0-1.2℃ / h, and FC CAPE is already >2500J / kg at 14:00. The actual heat reserve is sufficient and there is no obvious deviation. Maintain this.

[0080] - Moisture deviation: In IC, TM_Jump > 2-4 at multiple stations in Huli, Haicang, and Xiang'an (significant humidification), consistent with TCWV > 53mm and high Q850 in FC. The actual humidification rate is consistent with or slightly stronger than the forecast, and will be maintained.

[0081] - Dynamic deviation: The IC only showed localized weak convergence (Tong'an Xike 0.16), and the FC's W850 turned positive after 17:00. The actual convergence was weak and the time window for the FC dynamic to turn negative was narrow. Overall, the triggering condition depends on urban thermal forcing being greater than systemic uplift.

[0082] - Calibration conclusion: Maintained. The revised high-risk period is 14:00-17:00 on the 29th, but attention should be paid to the possibility that weaker actual convergence may reduce the triggering efficiency.

[0083] IV. Early Warning Recommendations - Trigger probability: High (maintained by calibration). The thermal + moisture conditions of IC have been established, and the high CAPE and sufficient moisture of FC support the potential for eruption, but the dynamic layer is weak, and triggering depends on the thermal rise of the boundary layer.

[0084] - Triggering mechanism: Continuous heating in urban areas (BL_Heating>1℃ / h) + temperature and humidity increase (TM_Jump>3) → boundary layer stratification instability (CAPE>2500J / kg) → near-surface convergence (small area in Tong'an and Longhai) in the afternoon from 14:00 to 17:00, combined with the 17:00 precipitation signal in the north of FC, triggers convection.

[0085] - High-risk period: 14:00-17:00 on the 29th, with the peak around 17:00. The key areas are northern Haicang, northern Jimei and central and northern Tongan (northwestern mountain frontier).

[0086] - Rolling correction: The conclusion at 11:00 remains "high probability of triggering". The key changes are: the thermal humidification signal (TM_Jump) in IC is more prevalent than at 11:00, but the convergence has not improved; in FC, W850 / 700 turns positive or weakly negative after 17:00, and the dynamic support weakens, but Crain3 still remains at 1.52mm at 17:00. The actual convergence conditions are weaker than expected, which may delay or weaken the triggering.

[0087] - Warning Recommendations: It is recommended to maintain a Level IV emergency response and intensify monitoring of Haicang, northern Jimei, and central and northern Tongan from 2 PM to 5 PM. Pay close attention to Crain3 signals > 0.5 mm and localized convergence enhancements before 5 PM (such as in the Tongan-Xike area). If FC dynamics do not significantly strengthen by 5 PM, lower the probability of triggering in subsequent periods.

[0088]

Overall Conclusion

[0089] In summary, this embodiment forms a rolling diagnostic chain by interleaving the real-time environmental dataset with the forecast trend environmental dataset, and updates the real-time environment and re-matches the forecast lead time when new real-time conditions arrive. This achieves continuous tracking of "current real atmospheric conditions - future forecast trends", overcomes the limitation of the separation between real-time analysis and forecast application in traditional methods, and enables the evolution trend of short-term heavy precipitation triggering conditions to be monitored dynamically in real time.

[0090] The parametric spatial fusion method organizes environmental parameters from different sources with diagnostic units as the object, without performing uniform gridding, temporal interpolation, or spatial resampling. This allows each data source to retain its original physical meaning, fundamentally avoiding the systematic physical bias introduced by forced spatial transformation in traditional multi-source data fusion. This ensures that the fused environmental parameter set can truly reflect the actual spatial configuration and physical state of the atmosphere, thereby significantly improving the accuracy and precision of trigger identification.

[0091] By constructing a structured short-term heavy precipitation environmental parameter matrix, thermal, dynamic, water vapor, boundary layer and topographic parameters are uniformly incorporated into a multi-dimensional feature space. Trigger window discrimination is performed based on the physical logic of multi-factor synergistic enhancement and suppression factor weakening, so that the discrimination rules have clear meteorological and physical basis, realizing objective and automatic trigger window state determination and reducing the unreliability of human subjective judgment.

[0092] By using physical constraint cue word templates to drive large language models for interpretation and reasoning, the model does not directly perform numerical weather forecasting. Instead, it performs logical reasoning based on the physical relationships between the input environmental parameters, generating structured interpretation results that include trigger probability, time window, main control factors, physical chains, and risk change trends. This makes the early warning output objective, interpretable, and operationally feasible, significantly improving the efficiency of forecasting personnel in understanding early warning information and their decision support capabilities.

[0093] To better implement the short-term heavy rainfall warning method in this embodiment of the invention, based on a short-term heavy rainfall warning method, correspondingly, as follows: Figure 2 As shown, this embodiment of the invention also provides a short-term heavy rainfall warning device. A short-term heavy rainfall warning device 200 includes: Data acquisition module 201 is used to acquire multi-source meteorological data and topographic data of the target area within a preset time period. The multi-source meteorological data includes: real-time observation data from automatic weather stations, radiosonde data, and numerical weather prediction data. The topographic data includes: digital elevation model (DEM) data. The data standardization processing module 202 is used to perform quality control and standardization processing on the multi-source meteorological data and topographic data to obtain quality-controlled automatic station real-time data, radiosonde profile data, numerical forecast field data and basic topographic feature data. The real-time environment dataset acquisition module 203 is used to construct a real-time environment dataset within a preset time period based on the quality-controlled automatic station real-time data and sounding profile data, construct a forecast trend environment dataset with multiple time-series forecasts within a future time period based on the numerical forecast field data, and update the real-time environment dataset based on the forecast trend environment dataset to obtain the updated real-time environment dataset. The short-term heavy precipitation environmental parameter matrix construction module 204 is used to construct a short-term heavy precipitation environmental parameter matrix based on the updated real-time environmental dataset and the topographic data. The short-term heavy precipitation environmental parameter matrix has a first dimension of diagnostic units, a second dimension of environmental parameters, and a third dimension of rolling diagnostic time and forecast lead time. The trigger window state determination module 205 is used to determine the trigger window state based on the short-term heavy precipitation environmental parameter matrix and according to the preset short-term heavy precipitation trigger window discrimination rule. The prompt word template construction module 206 is used to construct a physical constraint prompt word template. The physical constraint prompt word template includes a fixed input format, a fixed analysis order, and a fixed output format, which is used to constrain the large language model to reason according to the meteorological physical chain. The interpretable recognition result generation module 207 is used to input the short-term heavy precipitation environmental parameter matrix and the state of the trigger window into the physical constraint prompt word template, and call the large language model to generate interpretable recognition results; The early warning module 208 is used to generate a short-term heavy rainfall intelligent early warning result based on the state of the trigger window and the interpretation recognition result.

[0094] The short-term heavy rainfall warning device 200 provided in the above embodiments can realize the technical solution described in the above embodiment of the short-term heavy rainfall warning method. The specific implementation principle of each module or unit can be found in the corresponding content in the above embodiment of the short-term heavy rainfall warning method, which will not be repeated here.

[0095] like Figure 3 As shown, the present invention also provides an electronic device 300. The electronic device 300 includes a processor 301, a memory 302, and a display 303. Figure 3 Only some components of the electronic device 300 are shown, but it should be understood that it is not required to implement all of the components shown, and more or fewer components may be implemented instead.

[0096] In some embodiments, processor 301 may be a central processing unit (CPU), microprocessor, or other data processing chip, used to run program code stored in memory 302 or process data, such as a short-term heavy rainfall early warning method in this invention.

[0097] In some embodiments, processor 301 may be a single server or a group of servers. The server group may be centralized or distributed. In some embodiments, processor 301 may be local or remote. In some embodiments, processor 301 may be implemented on a cloud platform. In some embodiments, the cloud platform may include a private cloud, public cloud, hybrid cloud, community cloud, distributed cloud, internal cloud, multi-cloud, or any combination thereof.

[0098] In some embodiments, memory 302 may be an internal storage unit of electronic device 300, such as a hard disk or memory of electronic device 300. In other embodiments, memory 302 may also be an external storage device of electronic device 300, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc. equipped on electronic device 300.

[0099] Furthermore, the memory 302 may include both internal storage units of the electronic device 300 and external storage devices. The memory 302 is used to store application software and various types of data installed on the electronic device 300.

[0100] In some embodiments, display 303 may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, or an OLED (Organic Light-Emitting Diode) touchscreen. Display 303 is used to display information from electronic device 300 and to display a visual user interface. Components 301-303 of electronic device 300 communicate with each other via a system bus.

[0101] In one embodiment, when the processor 301 executes a short-term heavy rainfall warning program stored in the memory 302, the following steps can be implemented: Acquire multi-source meteorological and topographic data for a target area within a preset time period. The multi-source meteorological data includes: real-time observation data from automatic weather stations, radiosonde data, and numerical weather prediction data. The topographic data includes: digital elevation model (DEM) data. The multi-source meteorological and topographic data were subjected to quality control and standardization to obtain quality-controlled automatic station real-time data, radiosonde profile data, numerical prediction field data, and basic topographic feature data. Based on the quality-controlled automatic station real-time data and sounding profile data, a real-time environment dataset for a preset time period is constructed. Based on the numerical forecast field data, a forecast trend environment dataset for multiple time periods in the future is constructed. The real-time environment dataset is updated based on the forecast trend environment dataset to obtain the updated real-time environment dataset. Based on the updated real-time environmental dataset and the topographic data, a short-term heavy precipitation environmental parameter matrix is ​​constructed. The short-term heavy precipitation environmental parameter matrix has a first dimension of diagnostic units, a second dimension of environmental parameters, and a third dimension of rolling diagnostic time and forecast lead time. Based on the aforementioned short-term heavy rainfall environmental parameter matrix, the trigger window status is determined according to the preset short-term heavy rainfall trigger window discrimination rules; A physical constraint cue word template is constructed, which includes a fixed input format, a fixed analysis order, and a fixed output format, and is used to constrain the large language model to reason according to the meteorological physical chain. The short-term heavy rainfall environmental parameter matrix and the state of the trigger window are input into the physical constraint prompt word template, and the large language model is called to generate an interpretable recognition result. Based on the state of the trigger window and the interpretation recognition result, a short-term heavy rainfall intelligent early warning result is generated.

[0102] It should be understood that when the processor 301 executes a short-term heavy rainfall warning program in the memory 302, in addition to the functions mentioned above, it can also perform other functions, as can be found in the description of the corresponding method embodiments above.

[0103] Furthermore, the embodiments of the present invention do not specifically limit the type of electronic device 300 mentioned. Electronic device 300 can be a mobile phone, tablet computer, personal digital assistant (PDA), wearable device, laptop computer, or other portable electronic device. Exemplary embodiments of portable electronic devices include, but are not limited to, portable electronic devices running iOS, Android, Microsoft, or other operating systems. The aforementioned portable electronic device can also be other portable electronic devices, such as a laptop computer with a touch-sensitive surface (e.g., a touch panel). It should also be understood that in some other embodiments of the present invention, electronic device 300 may not be a portable electronic device, but rather a desktop computer with a touch-sensitive surface (e.g., a touch panel).

[0104] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware, and the program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.

[0105] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for issuing early warnings of short-term heavy rainfall, characterized in that, include: Acquire multi-source meteorological and topographic data for a target area within a preset time period. The multi-source meteorological data includes: real-time observation data from automatic weather stations, radiosonde data, and numerical weather prediction data. The topographic data includes: digital elevation model (DEM) data. The multi-source meteorological and topographic data were subjected to quality control and standardization to obtain quality-controlled automatic station real-time data, radiosonde profile data, numerical prediction field data, and basic topographic feature data. Based on the real-time data from the automatic weather station and the sounding profile data after quality control, a real-time environment dataset is constructed within a preset time period. Based on the numerical forecast field data, a forecast trend environment dataset with multiple timeframes is constructed for future time periods. The real-time environment dataset is updated based on the forecast trend environment dataset to obtain the updated real-time environment dataset. Based on the updated real-time environmental dataset and the topographic data, a short-term heavy precipitation environmental parameter matrix is ​​constructed. The short-term heavy precipitation environmental parameter matrix has a first dimension of diagnostic units, a second dimension of environmental parameters, and a third dimension of rolling diagnostic time and forecast lead time. Based on the short-term heavy precipitation environmental parameter matrix, the trigger window state is determined according to the preset short-term heavy precipitation trigger window discrimination rule; A physical constraint cue word template is constructed, which includes a fixed input format, a fixed analysis order, and a fixed output format, and is used to constrain the large language model to reason according to the meteorological physical chain. The short-term heavy precipitation environmental parameter matrix and the state of the trigger window are input into the physical constraint prompt word template, and the large language model is called to generate an interpretable recognition result. Based on the state of the trigger window and the interpretation recognition result, a short-term heavy rainfall intelligent early warning result is generated.

2. The short-term heavy rainfall early warning method according to claim 1, characterized in that, The real-time observation data from the automatic weather station includes at least one of the following: temperature, relative humidity, wind direction, wind speed, air pressure, and precipitation. The radiosonde data includes at least one of the following: temperature, dew point temperature, wind direction, wind speed, air pressure, and relative humidity at different altitudes. The numerical forecast data includes: hourly forecasts of temperature, humidity, wind field, geopotential height, vertical velocity, precipitation and / or water vapor flux for the next 24 hours. The digital elevation model (DEM) data includes: DEM elevation, slope, aspect, valley distribution, plateau boundaries, and / or coastal topographic features.

3. The short-term heavy rainfall early warning method according to claim 1, characterized in that, The near-surface atmospheric state signals reflected by the real-time data from automatic stations in the real-time environmental dataset include at least one of the following: near-surface temperature change rate, humidity change rate, wind direction change, wind speed change, surface convergence index, and local temperature and humidity disturbance. The vertical stratification parameters reflected by the sounding profile data in the real-world environmental dataset include at least one of the following: Convective Active Potential Energy (CAPE), Convective Inhibition Energy (CIN), Lifted Condensation Height (LCL), Free Convection Height (LFC), Lower Layer Stability, Inversion Intensity, Wet Layer Thickness, 0-3 km Vertical Wind Shear, and 0-6 km Vertical Wind Shear.

4. The short-term heavy rainfall early warning method according to claim 1, characterized in that, The thermal parameters in the forecast trend environmental dataset include at least one of CAPE, CIN, LCL, LFC, and lower-level stability; the dynamic parameters include at least one of lower-level wind field, vertical wind shear, surface or lower-level convergence, storm relative helicity (SRH), and vertical velocity; the water vapor parameters include at least one of relative humidity, wet layer thickness, water vapor flux, and water vapor flux convergence; and the stratification parameters include at least one of temperature lapse rate, inversion intensity, and dry intrusion height.

5. The short-term heavy rainfall early warning method according to claim 1, characterized in that, The construction of a short-term heavy precipitation environmental parameter matrix based on the updated real-time environmental dataset and the topographic data includes: Extract terrain trigger features of the target area based on the DEM terrain elevation data in the terrain data; The updated real-world environment dataset and terrain-triggered features are fused in parameter space according to their physical meaning to obtain a multi-source environmental parameter set. A short-term heavy precipitation environmental parameter matrix is ​​constructed based on the aforementioned multi-source environmental parameter set.

6. The short-term heavy rainfall early warning method according to claim 1, characterized in that, The fixed analysis order in the physical constraint prompt word template is as follows: thermal condition analysis, water vapor condition analysis, dynamic condition analysis, boundary layer signal analysis, terrain triggering condition analysis, inhibition factor analysis, and uncertainty source analysis. The fixed output format in the physical constraint prompt template includes the following fields: trigger probability, possible trigger time window, main favorable factors, main inhibiting factors, key physical chain, risk enhancement or reduction trend, key focus area, and business prompt text.

7. The short-term heavy rainfall early warning method according to claim 1, characterized in that, The intelligent early warning results for short-term heavy precipitation include: The short-term heavy rainfall intelligent early warning results shall be output in at least one of the following formats: current environmental diagnosis conclusions, risk trends for the next 24 hours, key areas of concern, possible triggering time periods, main triggering mechanisms, main inhibitory factors, sources of uncertainty, and business recommendations; the results shall be output in at least one of the following formats: structured text, tables, graphs, or interface data.

8. A short-term heavy rainfall early warning device, characterized in that, include: The data acquisition module is used to acquire multi-source meteorological data and topographic data of the target area within a preset time period. The multi-source meteorological data includes: real-time observation data from automatic weather stations, radiosonde data, and numerical weather prediction data. The topographic data includes: digital elevation model (DEM) data. The data standardization processing module is used to perform quality control and standardization processing on the multi-source meteorological data and topographic data to obtain quality-controlled automatic station real-time data, radiosonde profile data, numerical prediction field data and basic topographic feature data. The real-world environment dataset acquisition module is used to construct a real-world environment dataset within a preset time period based on the quality-controlled automatic station real-world data and sounding profile data, construct a forecast trend environment dataset with multiple timeframes in the future based on the numerical forecast field data, and update the real-world environment dataset based on the forecast trend environment dataset to obtain the updated real-world environment dataset. The short-term heavy precipitation environmental parameter matrix construction module is used to construct a short-term heavy precipitation environmental parameter matrix based on the updated real-time environmental dataset and the topographic data. The short-term heavy precipitation environmental parameter matrix has a first dimension of diagnostic units, a second dimension of environmental parameters, and a third dimension of rolling diagnostic time and forecast lead time. The trigger window state determination module is used to determine the trigger window state based on the short-term heavy precipitation environmental parameter matrix and according to the preset short-term heavy precipitation trigger window discrimination rules. The prompt word template construction module is used to construct physical constraint prompt word templates. The physical constraint prompt word templates include fixed input format, fixed analysis order and fixed output format, which are used to constrain the large language model to reason according to the meteorological physical chain. The interpretable recognition result generation module is used to input the short-term heavy precipitation environmental parameter matrix and the state of the trigger window into the physical constraint prompt word template, and call the large language model to generate interpretable recognition results; The early warning module is used to generate intelligent early warning results for short-term heavy rainfall based on the state of the trigger window and the interpreted recognition results.

9. An electronic device, characterized in that, Including memory and processor, among which, The memory is used to store programs; The processor, coupled to the memory, is used to execute the program stored in the memory to implement the steps in the short-term heavy rainfall warning method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, Used to store computer-readable programs or instructions, which, when executed by a processor, can implement the steps in the short-term heavy rainfall early warning method according to any one of claims 1 to 7.

Images