A Deep Reinforcement Learning-Based Method and System for Urban Meteorological Disaster Data Identification

Abstract

This invention relates to the field of data analysis technology, and provides a method and system for identifying urban meteorological disaster data based on deep reinforcement learning, which effectively improves model performance to enhance the accuracy and real-time performance of meteorological disaster monitoring. The method includes: acquiring a meteorological monitoring data set for a target urban area; performing meteorological data preprocessing on the meteorological monitoring data set to obtain a preprocessed set of meteorological spatiotemporal features; calling a trained deep reinforcement learning recognition model; performing dynamic disaster pattern matching processing on the meteorological spatiotemporal feature set to generate a meteorological disaster identification result set for the target urban area; generating a disaster response strategy set based on the meteorological disaster identification result set; performing dynamic strategy optimization processing on the deep reinforcement learning recognition model based on the disaster response strategy set to obtain an optimized deep reinforcement learning recognition model; and deploying the optimized deep reinforcement learning recognition model to a meteorological disaster monitoring system.

Description

Technical Field

[0001] This invention belongs to the field of data analysis technology, specifically relating to a method and system for identifying urban meteorological disaster data based on deep reinforcement learning. Background Technology

[0002] With the acceleration of urbanization, the threat of meteorological disasters to urban operational safety is becoming increasingly prominent. Current urban meteorological disaster analysis technologies mainly rely on discrete data collection from meteorological observation stations and numerical forecasting model extrapolation, achieving risk warnings by constructing a disaster level assessment index system. Traditional techniques often employ machine learning algorithms to match patterns from historical disaster cases, combined with geographic information systems to display the scope of disaster impact. However, these techniques struggle to effectively integrate the temporal variation characteristics of meteorological elements with the distribution characteristics of geographic space, leading to the loss of spatiotemporal correlation information during disaster identification. Machine learning models based on static training sets lack the dynamic response capability to real-time meteorological changes, are prone to misjudgments during disaster pattern matching, and are ill-suited to the complex urban environment where multiple disasters are coupled and evolve. Therefore, how to effectively improve model performance to enhance the accuracy and real-time performance of meteorological disaster monitoring is a pressing technical problem that needs to be addressed. Summary of the Invention

[0003] This invention provides a method and system for identifying urban meteorological disaster data based on deep reinforcement learning, which can effectively improve model performance and enhance the accuracy and real-time performance of meteorological disaster monitoring.

[0004] In a first aspect, embodiments of the present invention provide a method for identifying urban meteorological disaster data based on deep reinforcement learning, applied to an urban meteorological disaster data identification system. The method includes: acquiring a meteorological monitoring data set for a target urban area, the meteorological monitoring data set including multi-dimensional meteorological time-series data and corresponding geospatial coding data; performing meteorological data preprocessing on the meteorological monitoring data set to obtain a preprocessed meteorological spatiotemporal feature set, the meteorological spatiotemporal feature set including meteorological element correlation features, geospatial distribution features, and meteorological change trend features; calling a trained deep reinforcement learning identification model to perform dynamic disaster pattern matching processing on the meteorological spatiotemporal feature set to generate a meteorological disaster identification result set for the target urban area, the meteorological disaster identification result set including disaster type identifiers, disaster impact ranges, and disaster evolution prediction paths; generating a disaster response strategy set based on the meteorological disaster identification result set, the disaster response strategy set including emergency resource scheduling strategies, risk area delineation strategies, and early warning information release strategies; performing dynamic strategy optimization processing on the deep reinforcement learning identification model based on the disaster response strategy set to obtain an optimized deep reinforcement learning identification model, and deploying the optimized deep reinforcement learning identification model to the meteorological disaster monitoring system.

[0005] Secondly, embodiments of the present invention provide an urban meteorological disaster data identification system, which includes a processor and a memory, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the above-described method.

[0006] Thirdly, embodiments of the present invention provide a computer-readable storage medium including a computer program, which, when run on an urban meteorological disaster data identification system, causes the urban meteorological disaster data identification system to perform the steps of the above-described method.

[0007] In this invention, by acquiring multi-dimensional meteorological time-series data and geospatial coding data, a comprehensive understanding of the meteorological conditions of the target urban area can be achieved. Through meteorological data preprocessing, the correlation characteristics, geospatial distribution characteristics, and meteorological change trend characteristics of meteorological elements are mined, providing rich and valuable information for subsequent analysis. Utilizing a deep reinforcement learning recognition model for dynamic disaster pattern matching, a set of meteorological disaster identification results containing disaster type identifiers, impact ranges, and evolution prediction paths can be accurately generated. Based on this, a set of disaster response strategies covering emergency resource allocation, risk area delineation, and early warning information dissemination can effectively address meteorological disasters. Furthermore, dynamic strategy optimization of the deep reinforcement learning recognition model and its deployment in the meteorological disaster monitoring system can improve model performance, thereby enhancing the accuracy and real-time performance of meteorological disaster monitoring. Attached Figure Description

[0008] Figure 1 This is a flowchart illustrating a method for identifying urban meteorological disaster data based on deep reinforcement learning, as provided in an embodiment of the present invention.

[0009] Figure 2 This is a schematic diagram of the structure of an urban meteorological disaster data identification system provided in an embodiment of the present invention. Detailed Implementation

[0010] See Figure 1 This invention provides a method for identifying urban meteorological disaster data based on deep reinforcement learning. This method can be applied to urban meteorological disaster data identification systems. The specific process is as follows: steps S110-S150.

[0011] Step S110: Obtain a meteorological monitoring data set for the target city area, wherein the meteorological monitoring data set includes multi-dimensional meteorological time-series data and corresponding geospatial coding data.

[0012] In this embodiment, a large urban area is used as the target area, and multiple meteorological monitoring stations are distributed within the target area to continuously collect various meteorological data. The multi-dimensional meteorological time-series data includes records of changes in various meteorological elements over time. For example, temperature is recorded every half hour starting at 6:00 AM, resulting in a set of data such as [18℃, 18.5℃, 19℃, 19.5℃, 20℃, ...]. Humidity is recorded synchronously, forming a data sequence such as [55%, 56%, 57%, 58%, 59%, ...]. In addition, elements such as air pressure, wind speed, and precipitation are also recorded at the same time intervals, collectively constituting the multi-dimensional meteorological time-series data.

[0013] Optionally, regarding geospatial coding data, the urban area is divided into grids of 1 square kilometer each, with each grid having a unique code. For each grid, the terrain elevation is recorded; for example, grid G10 has a terrain elevation of 45 meters. The distribution of urban buildings is also recorded, showing that buildings occupy 40% of the grid area, primarily residential buildings. Water system coverage is also recorded; this grid contains a stream, and the water system covers 5% of the area. This grid data constitutes the geospatial coding data.

[0014] Step S120: Perform meteorological data preprocessing on the meteorological monitoring data set to obtain a preprocessed meteorological spatiotemporal feature set, which includes meteorological element correlation features, geographic spatial distribution features, and meteorological change trend features.

[0015] In one implementation, meteorological data preprocessing is performed on the meteorological monitoring data set to obtain a preprocessed meteorological spatiotemporal feature set. This meteorological spatiotemporal feature set includes meteorological element correlation features, geographic spatial distribution features, and meteorological change trend features, including:

[0016] Step S121: Perform abnormal data cleaning on the multi-dimensional meteorological time series data, remove data fragments exceeding the preset meteorological element fluctuation threshold, generate cleaned meteorological time series data, perform spatiotemporal alignment processing on the cleaned meteorological time series data, unify meteorological data with different time resolutions to a preset time base, and divide the spatial grid according to the geospatial coding data to generate spatiotemporally aligned meteorological grid data, extract meteorological element association features from the spatiotemporally aligned meteorological grid data, the meteorological element association features include temperature-humidity coupling features, air pressure gradient change features, and wind speed-precipitation synergistic features.

[0017] In this embodiment, for abnormal data cleaning, taking temperature as an example, by analyzing historical data, the normal temperature fluctuation range of the urban area is between 10℃ and 35℃, and the temperature change between adjacent half-hour intervals generally does not exceed 2℃. Therefore, the temperature fluctuation threshold is set to 2℃. In the currently collected temperature time series data, if the difference between two adjacent data points exceeds 2℃, such as a sudden change from 20℃ to 23℃, this data segment is determined to be abnormal and removed, thus obtaining the cleaned meteorological time series data.

[0018] When performing spatiotemporal alignment, the preset time base is the hour. For data with different time resolutions, such as wind speed data recorded every 15 minutes, they are integrated according to the hour. For example, between 9:00 and 10:00, there are four wind speed values ​​recorded every 15 minutes [5 m / s, 6 m / s, 7 m / s, 8 m / s]. The wind speed at the hour 10:00 is the average of these four values, i.e., (5+6+7+8)÷4=6.5 m / s. According to the grid division of geospatial coded data, the meteorological data within each grid is integrated together to form spatiotemporally aligned meteorological grid data.

[0019] When extracting the correlation features of meteorological elements, for the temperature-humidity coupling feature, for example, within the temperature range of 15℃-25℃, the relationship between humidity and temperature is approximately humidity = 2 × temperature - 20. Based on this, the temperature-humidity coupling feature is calculated by combining the temperature and humidity values ​​in the current meteorological grid data. For the pressure gradient change feature, the pressure difference between adjacent grids is calculated. For example, if the pressures of adjacent grids G1 and G2 are 1008 hPa and 1005 hPa respectively, and the grid spacing is 1 km, then the pressure gradient is (1008-1005) ÷ 1 = 3 hPa / km. Regarding the wind speed-precipitation synergy feature, statistics show that the probability of precipitation increases significantly when the wind speed is greater than 8 m / s. By analyzing the precipitation situation under different wind speeds in the current meteorological grid data, the wind speed-precipitation synergy feature is derived.

[0020] Step S122: Perform spatial topology parsing processing on the geospatial encoded data to generate geospatial distribution features that include terrain elevation correlation features, urban building distribution features, and water system coverage features.

[0021] In this embodiment, the terrain elevation correlation features are analyzed by examining the differences and directions of elevation change between adjacent grids. For example, the terrain elevation of grid G3 is 50 meters, and the terrain elevation of the adjacent grid G4 is 48 meters. The elevation correlation between them is represented as an elevation difference of -2 meters, with the direction of change being a decrease from G3 to G4. Urban building distribution features are determined by detailed statistics on the building area and building type ratio within each grid. For example, the building area in grid G5 is 0.5 square kilometers, of which commercial buildings account for 30% and residential buildings account for 70%. Water system coverage features record the specific details of the water system within the grid. For instance, a river flows through grid G6 with a width of 50 meters, flowing from west to east, and the water system coverage area accounts for 15% of the grid area.

[0022] Step S123: Perform trend fitting processing on the spatiotemporally aligned meteorological grid data based on a sliding time window to generate meteorological change trend features that include the cumulative change rate of meteorological elements, the probability of sudden meteorological events, and the characteristics of meteorological state transition.

[0023] In this embodiment, the sliding time window is set to 6 hours. Taking temperature as an example, in the current spatiotemporally aligned meteorological grid data, the temperature data within this sliding time window from 8:00 AM to 2:00 PM is [20℃, 21℃, 22℃, 23℃, 24℃, 25℃]. The cumulative rate of change of meteorological elements is calculated as follows: Cumulative rate of change of temperature = (final temperature - initial temperature) ÷ time span = (25 - 20) ÷ 6 ≈ 0.83℃ / hour.

[0024] To determine the probability of a meteorological event's sudden occurrence, we analyze historical data and current meteorological grid data to determine the frequency of a specific meteorological event (such as a rainstorm) occurring under similar weather conditions. For example, given the current temperature, humidity, and wind speed conditions, and based on historical data, if rainstorms occurred 10 times in the past 100 similar weather conditions, then the probability of the current meteorological event (rainstorm) occurring is 10 ÷ 100 = 10%.

[0025] Regarding the characteristics of meteorological state transitions, the transitions between different meteorological states can be analyzed. For example, if the current meteorological state is sunny, by analyzing historical data and current meteorological trends, it can be found that there is a 30% probability of it turning into cloudy weather and a 10% probability of it turning into light rain within the next 6 hours, thus determining the characteristics of meteorological state transitions.

[0026] Step S124: Generate the meteorological spatiotemporal feature set by combining the meteorological element association features, the geographic spatial distribution features, and the meteorological change trend features.

[0027] In this embodiment, the meteorological element correlation features, geospatial distribution features, and meteorological change trend features obtained above are integrated. For example, meteorological element correlation features such as temperature-humidity coupling features, pressure gradient change features, and wind speed-precipitation synergy features are combined; geospatial distribution features such as topographic elevation correlation features, urban building distribution features, and water system coverage features are combined; and meteorological change trend features such as cumulative change rate of meteorological elements, probability of meteorological events, and meteorological state transition features are combined. These three sets of features are then spliced ​​together to form a set of meteorological spatiotemporal features. During the splicing process, those skilled in the art can standardize the dimensions and scales of different features using existing technologies, based on the actual situation.

[0028] Step S130: Call the trained deep reinforcement learning recognition model to perform dynamic disaster pattern matching processing on the meteorological spatiotemporal feature set, and generate a meteorological disaster recognition result set for the target urban area. The meteorological disaster recognition result set includes disaster type identifier, disaster impact range and disaster evolution prediction path.

[0029] Optionally, step 130 includes:

[0030] Step S131: Input the meteorological element association features into the feature encoding layer of the deep reinforcement learning recognition model to generate meteorological element feature vectors.

[0031] In this embodiment, the previously extracted meteorological element association features, such as temperature-humidity coupling features, pressure gradient change features, and wind speed-precipitation synergistic features, are input into the feature encoding layer of the deep reinforcement learning recognition model. The feature encoding layer processes these features; for example, the temperature-humidity coupling feature is digitized. For instance, the temperature-humidity coupling feature is represented by a two-dimensional array, where the first element represents the linear relationship coefficient between temperature and humidity, and the second element represents a constant term. The feature encoding layer converts this two-dimensional array into a vector of length 10 using existing encoding algorithms. Similar encoding processing is performed on the pressure gradient change features and the wind speed-precipitation synergistic features. Finally, these encoded feature vectors are concatenated to form the meteorological element feature vector.

[0032] Step S132: Input the geospatial distribution features into the spatial relationship parsing layer of the deep reinforcement learning recognition model to generate a spatial topological relationship matrix, which contains meteorological influence weights between different geographical units.

[0033] In this embodiment, geospatial distribution features are input into the spatial relationship analysis layer. Taking topographic elevation correlation features as an example, the spatial relationship analysis layer analyzes the impact of topographic elevation differences between different grids on meteorology. For example, if the topographic elevation difference between grid A and grid B is large, historical data analysis reveals that this elevation difference makes it easier for airflow from grid A to flow to grid B, thus affecting precipitation distribution. Based on this degree of influence, a meteorological influence weight of 0.6 is assigned to the area from grid A to grid B. Similarly, the influence of urban building distribution features and water system coverage features on meteorological elements between different grids is analyzed. For example, a large number of high-rise buildings within a grid can change local wind speed and direction. Based on the degree of influence, meteorological influence weights are assigned to the relevant grids, ultimately forming a spatial topological relationship matrix representing the meteorological influence weights between different geographic units (grids).

[0034] Step S133: Input the meteorological change trend features into the time series analysis layer of the deep reinforcement learning recognition model to generate a meteorological state transition probability map.

[0035] In this embodiment, the meteorological state transition features from the meteorological change trend characteristics are input into the time series analysis layer. For example, the meteorological state transition features record the probability that the current sunny state will change to other meteorological states (such as cloudy, light rain, etc.) in different time periods. Based on this information, the time series analysis layer combines historical meteorological data and the current meteorological trend to generate a meteorological state transition probability map. This probability map is represented in matrix form, where the rows and columns of the matrix represent different meteorological states, and the matrix elements represent the probability of transitioning from one meteorological state to another. For example, the probability of transitioning from sunny to cloudy in the matrix corresponds to a value of 0.4, and the probability of transitioning from sunny to light rain corresponds to a value of 0.1, etc.

[0036] Step S134: The policy network of the deep reinforcement learning recognition model is used to jointly evaluate the meteorological element feature vector, the spatial topology matrix and the meteorological state transition probability map to generate a dynamic evaluation result that includes the matching degree of disaster triggering conditions, the similarity of disaster diffusion paths and the score of disaster evolution stage.

[0037] In this embodiment, the policy network of the deep reinforcement learning recognition model comprehensively analyzes the previously generated meteorological element feature vectors, spatial topological relationship matrix, and meteorological state transition probability map. The policy network first analyzes the meteorological element feature vectors to determine whether the current combination of meteorological elements meets the triggering conditions for a certain disaster. For example, when temperature-humidity coupling characteristics, pressure gradient change characteristics, etc., reach a certain threshold, a rainstorm disaster may be triggered. The matching degree of the disaster triggering conditions is calculated by comparing it with a preset disaster triggering condition template. For example, the matching degree between the current meteorological element feature vector and the rainstorm disaster triggering condition template is 0.7.

[0038] Next, the policy network combines the spatial topology matrix and the meteorological state transition probability map to analyze possible disaster spread paths. For example, based on the meteorological influence weights between different grids and the meteorological state transition probability, it predicts which grids rainwater might spread from to which grids if a rainstorm disaster occurs, and calculates the similarity between the current situation and the preset rainstorm disaster spread path template, for example, 0.6.

[0039] Simultaneously, the strategy network assesses the current evolutionary stage of the disaster based on the current meteorological conditions and trends, and assigns a disaster evolution stage score. For example, if the current rainstorm disaster is determined to be in the initial stage, the score is 0.5. These assessment results are then combined to form a dynamic assessment result.

[0040] Step S135: Based on the value network of the deep reinforcement learning recognition model, perform disaster impact value prediction on the dynamic assessment results, and generate a set of meteorological disaster identification results including disaster economic loss prediction value, personnel safety risk index and infrastructure vulnerability level.

[0041] In this embodiment, the value network of the deep reinforcement learning recognition model predicts the value of disaster impact based on dynamic evaluation results. For the predicted economic loss, the value network analyzes historical disaster data and the current economic situation of the urban area. For example, given that similar past rainstorm disasters had similar triggering conditions, diffusion paths, and evolutionary stage scores, the average economic loss was 5 million yuan. Based on the current economic growth of the urban area and the economic density of the affected area, and using existing calculation methods (such as linear regression models adjusted with relevant parameters), the predicted economic loss for this rainstorm disaster is 6 million yuan.

[0042] For the personnel safety risk index, the value network considers factors such as population density and building type in the disaster-stricken area. For example, if the disaster-stricken area has a high population density and many old buildings, by analyzing the probability of casualties in similar situations in historical data and combining it with the current meteorological disaster situation, the personnel safety risk index is calculated to be 0.6 (the higher the index, the greater the risk).

[0043] For infrastructure vulnerability levels, the value network analyzes the types of infrastructure (such as roads, bridges, and hydropower facilities) in the affected area, as well as their construction dates and maintenance conditions. For example, if the bridges in a certain area were built relatively early and are in average condition, and considering the impact of the current meteorological disaster on the bridges, the infrastructure vulnerability level of the area is assessed as medium. These prediction results are combined to form a set of meteorological disaster identification results.

[0044] Step S140: Generate a disaster response strategy set based on the meteorological disaster identification result set. The disaster response strategy set includes emergency resource scheduling strategy, risk area division strategy, and early warning information release strategy.

[0045] Optionally, step S140 includes:

[0046] Step S141: Match the disaster type identifier with a preset emergency resource type library to generate an emergency resource dispatch strategy that includes material storage location, transportation route planning and allocation priority.

[0047] In this embodiment, when the disaster type identifier in the meteorological disaster identification result set is a rainstorm disaster, a preset emergency resource type database is matched. The emergency resource type database records various required emergency resources for rainstorm disasters, such as flood control sandbags and water pumps. Regarding the location of material reserves, it is known that flood control sandbags are stored in three warehouses in the city: Warehouse A, Warehouse B, and Warehouse C. Warehouse A stores 5000 sandbags, Warehouse B stores 3000, and Warehouse C stores 2000. Ten water pumps are stored in the city's maintenance center. Transportation route planning is based on geospatial information and real-time traffic conditions. For example, if the disaster area is mainly in the eastern part of the city, and Warehouse A is closer to the disaster area and has good traffic conditions, then transporting flood control sandbags from Warehouse A is prioritized. The planned transportation route from Warehouse A to the disaster area passes through main roads X and Y. Allocation priority is determined based on the severity of the disaster area and the urgency of the needs. Resources are allocated to the most severely affected areas first. For example, if the severity score of affected area A is 8 out of 10, and the severity score of affected area B is 6, then the allocation of flood control sandbags to affected area A will have a higher priority than that to affected area B. An emergency resource allocation strategy is generated based on these rules.

[0048] Step S142: Based on the disaster impact range superimposed with the urban population density distribution map, generate a risk area delineation strategy that includes high-risk area lockdown plans, personnel evacuation routes, and temporary resettlement point settings.

[0049] In this embodiment, the disaster impact range in the known meteorological disaster identification result set is used, and the rainstorm disaster impact range covers multiple areas of the city. This disaster impact range is overlaid and analyzed with the city's population density distribution map. For example, areas with high population density within the disaster area are mainly concentrated in several blocks, such as blocks M, N, and O. These blocks are prone to flooding during rainstorms due to their low-lying terrain, posing a high risk. A lockdown plan is developed for these high-risk areas. For example, traffic control is implemented in block M, prohibiting non-emergency vehicles from entering, and warning signs are set up. Evacuation routes are determined based on geospatial information and building distribution. For example, residents in block M evacuate to nearby higher ground via streets P and Q, with clear directional signs along the evacuation routes. Temporary resettlement points are set up in higher, safer areas with some infrastructure. For example, a temporary resettlement point is set up in a park near block M, which has sufficient space to accommodate evacuees and is equipped with temporary tents, drinking water supply facilities, etc. Following the above method, a risk area delineation strategy is generated.

[0050] Step S143: Based on the time series characteristics of the disaster evolution prediction path, generate an early warning information release strategy that includes the early warning information release level, the priority of the early warning coverage area, and the combination of information dissemination channels.

[0051] In this embodiment, the time series characteristics of the disaster evolution prediction path in the meteorological disaster identification result set are used. For example, the rainstorm disaster evolution prediction path shows that the disaster will gradually worsen within the next 3 hours. Based on this time series characteristic, the warning information issuance level is determined. As the disaster gradually worsens, the warning information issuance level is upgraded from yellow warning (relatively severe) to orange warning (severe). The priority of the warning coverage area is determined based on the disaster impact range and population density. Areas with high population density within the disaster area are given priority for warnings, such as the previously mentioned blocks M, N, and O. Multiple information dissemination channels are selected to ensure that information can be delivered to relevant personnel in a timely manner. For example, orange warning information is issued through television and radio, while warning text messages are sent to mobile phone users in the disaster area using mobile phone SMS platforms, and warning notices are posted on community bulletin boards in the disaster area. According to the above rules, a warning information issuance strategy is generated.

[0052] Step S144: Perform strategy conflict detection on the emergency resource scheduling strategy, risk area division strategy and early warning information release strategy to eliminate execution time conflicts and resource allocation conflicts between different strategies, and obtain an optimized disaster response strategy set.

[0053] In this embodiment, conflict detection is performed on the generated emergency resource scheduling strategy, risk area delineation strategy, and early warning information dissemination strategy. For example, the emergency resource scheduling strategy plans to transport flood control sandbags from warehouse A to the disaster area during a certain time period, but the risk area delineation strategy stipulates that traffic control will be implemented in that area during that time period, resulting in an execution time conflict. By adjusting the transportation time and postponing it until after the traffic control is lifted, the execution time conflict is resolved.

[0054] Regarding resource allocation conflicts, the early warning information dissemination strategy requires the use of some vehicles to post notices in disaster-stricken areas, but the emergency resource dispatch strategy allocates these vehicles to transport supplies. By reallocating vehicle resources and assigning some spare vehicles for notice posting, the resource allocation conflict is resolved. After detecting and adjusting these conflicts, execution time conflicts and resource allocation conflicts between different strategies are eliminated, resulting in an optimized set of disaster response strategies. This optimized set of strategies ensures that various measures can be implemented in a coordinated and efficient manner when responding to meteorological disasters, minimizing the losses caused by the disaster.

[0055] Step S145: Encode the optimized disaster response strategy set into an execution instruction format and push it to the corresponding emergency management terminal device.

[0056] In this embodiment, the optimized disaster response strategy set is encoded to conform to a format that emergency management terminal equipment can recognize and execute. For example, information such as the location of material reserves, transportation route planning, and allocation priority in the emergency resource dispatch strategy is converted according to specific encoding rules. For example, the location of material reserves is represented by numerical codes, with warehouse A represented by "01", warehouse B by "02", and warehouse C by "03"; transportation routes are digitized using road numbers and node information; allocation priorities are represented by numbers 1-5, with 1 being the highest priority. These encoded information are combined into a string, such as "01, 101-102-103, 3; 02, 104-105, 2", indicating that materials are transported from warehouse A along route 101-102-103 with an allocation priority of 3, and materials are transported from warehouse B along route 104-105 with an allocation priority of 2.

[0057] For the risk zone delineation strategy, information such as the high-risk area lockdown plan, evacuation routes, and temporary resettlement site setup are also coded. For example, a high-risk area block M is represented by "M", and the lockdown plan is represented by the code "L1". Evacuation routes are coded using street numbers and direction information, such as "PN" indicating evacuation from street P to the north. The temporary resettlement park is represented by "P1". These coded information are combined into strings like "M, L1; PN; P1".

[0058] The early warning information release strategy is also coded accordingly. The early warning information release level is represented by a specific character. An orange warning is represented by "O". The priority of the warning coverage area is based on the combination of area number and priority number. For example, "M, 1; N, 2; O, 3" means that street M has a priority of 1, street N has a priority of 2, and street O has a priority of 3. The combination of information dissemination channels is represented by a code. For example, television is represented by "T", radio by "R", SMS by "S", and bulletin board by "B", which are combined into the string "O, M, 1; N, 2; O, 3, TRSB".

[0059] These encoded emergency resource allocation strategies, risk zone delineation strategies, and early warning information dissemination strategies are combined into a complete and optimized disaster response strategy set. Then, using network communication technology, the encoded instructions are pushed to corresponding emergency management terminal devices, such as computers in the emergency command center and handheld terminals for rescue personnel. Upon receiving the instructions, these terminal devices can automatically decode them and execute the corresponding response measures.

[0060] Step S150: Based on the disaster response strategy set, perform dynamic strategy optimization on the deep reinforcement learning recognition model to obtain an optimized deep reinforcement learning recognition model, and deploy the optimized deep reinforcement learning recognition model to the meteorological disaster monitoring system.

[0061] Optionally, step S150 includes:

[0062] Step S151: Extract actual material arrival time data from the emergency resource scheduling strategy, extract personnel evacuation route execution efficiency data from the risk area division strategy, and generate a set of strategy execution performance indicators.

[0063] In this embodiment, during the implementation of the emergency resource dispatch strategy, positioning devices and time recording devices are installed on material transport vehicles to obtain real-time data on the actual arrival time of materials. For example, if flood control sandbags depart from warehouse A and are scheduled to arrive at the disaster area at 10:00 AM, but actually arrive at 10:30 AM, this 30-minute time difference is recorded as part of the actual material arrival time data. Corresponding actual arrival time data is recorded for different batches and types of materials.

[0064] During the implementation of the risk zone delineation strategy, sensors and statistical devices are installed along evacuation routes to acquire data on the efficiency of these routes. For example, on evacuation routes P and Q in block M, the number of people passing through per unit of time is counted. For instance, if the plan is to evacuate 500 people per hour and the actual evacuation is 450 people per hour, the efficiency of the evacuation routes is calculated by combining the ratio of actual evacuees to planned evacuees (450 ÷ 500 = 0.9) with the difference between evacuation time and estimated evacuation time.

[0065] By combining the actual arrival time data of supplies and the execution efficiency data of personnel evacuation routes, a set of strategy execution effectiveness indicators is formed, which reflects the effectiveness and efficiency of disaster response strategies in actual implementation.

[0066] Step S152: Input the set of policy execution performance indicators into the backpropagation interface of the policy network of the deep reinforcement learning recognition model, and calculate the weight gradient deviation matrix of the disaster triggering condition matching degree.

[0067] In this embodiment, the generated set of policy execution performance indicators is input into the backpropagation interface of the policy network of the deep reinforcement learning recognition model. The policy network backpropagation interface calculates the weight gradient bias of the disaster triggering condition matching degree based on the input indicator set and the model's current parameters. For example, the current model's predicted disaster triggering condition matching degree may deviate from the actual situation; in reality, a rainstorm disaster should be triggered under certain meteorological conditions, but the model's predicted matching degree is low. The backpropagation algorithm calculates the gradient bias of each weight parameter that causes this deviation.

[0068] The specific calculation process is as follows: First, based on the data in the strategy execution performance index set, determine the error between the actual disaster occurrence and the model's predicted disaster triggering conditions. For example, an actual rainstorm disaster may occur, but the model's predicted matching degree is only 0.4, while the ideal matching degree should be close to 1. Then, based on this error, combined with the model structure and current weight parameters, use the backpropagation algorithm to propagate the error from the output layer to the input layer, calculating the contribution of each weight parameter to the error, i.e., the weight gradient. Finally, organize these weight gradients into a matrix, forming the weight gradient bias matrix. Each element in this matrix represents the gradient bias of the corresponding weight parameter, reflecting the direction and magnitude of adjustment required for that weight parameter.

[0069] Step S153: Adjust the meteorological influence weight allocation ratio between different geographical units in the spatial topology matrix based on the weight gradient deviation matrix to generate an optimized spatial topology matrix.

[0070] In this embodiment, the spatial topology matrix is ​​adjusted based on the weight gradient deviation matrix. For example, the weight gradient deviation matrix shows that the meteorological influence weight between a certain geographic unit (grid) and its neighboring grids contributes significantly to the deviation in the matching degree of disaster triggering conditions. For example, if the current value of this weight is 0.5, it is decided to adjust it to 0.6 based on the direction and magnitude of the weight gradient deviation.

[0071] Other weights in the spatial topology matrix are adjusted in a similar manner. By traversing the weight gradient bias matrix, the meteorological impact weights in the spatial topology matrix are adjusted according to the situation of each element. The principle of adjustment is to reduce the weights that cause deviations in the matching degree of disaster triggering conditions and increase the weights that help improve the matching degree. After the above adjustments, an optimized spatial topology matrix is ​​generated. This optimized matrix more accurately reflects the meteorological impact relationships between different geographical units and can better support the model's identification and prediction of disasters.

[0072] Step S154: Perform convolution alignment operation on the actual response delay data of the early warning information release strategy and the time series of the meteorological state transition probability map to generate the value network loss function correction coefficient.

[0073] In this embodiment, the actual response delay data of the early warning information release strategy is obtained. For example, the early warning information release strategy plans to issue an orange warning one hour before the disaster occurs, but the actual release time is delayed by 30 minutes, i.e., the actual response delay data is 30 minutes. This actual response delay data is then convolved and aligned with the time series of the meteorological state transition probability map.

[0074] The time series of the meteorological state transition probability map records the probability of meteorological state transitions at different times. For example, using existing convolutional alignment processing, the actual response delay data can be matched and calculated with the time series of the meteorological state transition probability map. For instance, the time series of the meteorological state transition probability map records probability values ​​at 15-minute intervals; the 30-minute actual response delay is mapped to this time series and calculated with the corresponding probability values. The calculation method could be weighted summation, etc. A value is obtained through weighted summation, and this value is then transformed using existing functions (such as normalization) to generate a correction coefficient for the value network loss function. This correction coefficient is used to adjust the calculation of the loss function in the value network to more accurately reflect the difference between the actual situation and the model prediction.

[0075] Step S155: Update the output parameters of the spatial relationship parsing layer using the optimized spatial topology matrix, and adjust the error calculation weight of the disaster economic loss prediction value using the value network loss function correction coefficient.

[0076] In this embodiment, the optimized spatial topology matrix is ​​applied to the spatial relationship parsing layer of the deep reinforcement learning recognition model to update its output parameters. For example, the spatial relationship parsing layer originally generated output results based on the old spatial topology matrix. Now, the new optimized matrix is ​​used as input, and the relevant calculation parameters and weights are adjusted so that the spatial relationship parsing layer can output results that more accurately reflect the geospatial meteorological influence relationships. Simultaneously, the error calculation weights for disaster economic loss predictions are adjusted using the value network loss function correction coefficient. For example, when calculating the error of disaster economic loss predictions, the value network originally set the weight for a certain factor (such as the economic growth of the disaster-stricken area) to 0.3. Based on the value network loss function correction coefficient and combined with the difference analysis between the actual situation and the model prediction, this weight is adjusted to 0.4. Through these adjustments, the value network can more reasonably consider various factors when calculating the error of disaster economic loss predictions, improving the accuracy of model predictions.

[0077] Step S156: Perform joint convergence verification on the updated parameters of the policy network and the value network. When the accuracy of the disaster evolution prediction path is improved to a preset threshold, output the optimized deep reinforcement learning recognition model.

[0078] In this embodiment, the updated parameters of the policy network and value network are jointly converged. By continuously inputting new meteorological spatiotemporal feature data, the updated model is used for meteorological disaster identification and related prediction. For example, meteorological data from different time periods are input multiple times, allowing the model to predict disaster evolution paths. The degree of agreement between the model's predicted disaster evolution paths and the actual situation is statistically analyzed, and the accuracy is calculated. A preset threshold is set at 80%. When, after multiple verifications, the accuracy of the model's predicted disaster evolution paths consistently and stably increases to 80% or higher, the parameter adjustment of the model is considered to have achieved a good effect, and at this point, the optimized deep reinforcement learning recognition model is output.

[0079] In an optional embodiment, the method further includes:

[0080] Step S210: Monitor the data identification delay index and disaster underreporting rate index of the meteorological disaster monitoring system in real time.

[0081] In this embodiment, a monitoring module is set up in the meteorological disaster monitoring system to collect and analyze data identification delay indicators and disaster underreporting rate indicators in real time. For the data identification delay indicator, the time taken from when meteorological monitoring data enters the system to when the deep reinforcement learning identification model outputs a set of meteorological disaster identification results is recorded. For example, if a set of meteorological data is detected entering the system at 9:00 AM, and the model outputs identification results at 9:15 AM, then the data identification delay time is 15 minutes. Multiple such data identification delay times are continuously recorded, and the average value is calculated as the current data identification delay indicator.

[0082] The disaster underreporting rate is calculated by comparing it with the actual number of meteorological disasters that occur. For example, if 10 meteorological disasters actually occur within a certain period, but the meteorological disaster monitoring system only identifies 8, then the disaster underreporting rate = (actual number of disasters - number of identified disasters) ÷ actual number of disasters = (10 - 8) ÷ 10 = 20%. By monitoring these indicators in real time, the performance status of the meteorological disaster monitoring system can be understood promptly.

[0083] Step S220: When the data recognition delay index exceeds the preset response threshold, the deep reinforcement learning recognition model is lightened by compressing the number of neurons in the feature encoding layer, reducing the hidden layer dimension of the policy network, and quantizing the parameter precision of the value network.

[0084] In this embodiment, the preset response threshold is set to 20 minutes. When the monitored data recognition latency exceeds 20 minutes, lightweight processing of the deep reinforcement learning recognition model begins. For the feature encoding layer, the number of neurons is compressed; for example, if the feature encoding layer originally had 100 neurons, the number is reduced to 80 according to a certain compression ratio (e.g., 20%). During the reduction of neurons, the contribution of each neuron to the model performance is analyzed, prioritizing the retention of neurons with larger contributions and removing those with smaller contributions. The hidden layer dimension of the policy network is reduced; for example, if the policy network originally had 3 hidden layers with 50 neurons each, the number of hidden layers is reduced to 2, and the number of neurons in each hidden layer is adjusted to 40, according to the lightweight requirements. By readjusting the network structure and weight parameters, the policy network can maintain a certain level of performance even with the reduced hidden layer dimension. The parameter precision of the value network is quantized; for example, if the parameters of the value network were originally represented as 32-bit floating-point numbers, they are now quantized as 16-bit floating-point numbers. During the quantization process, existing quantization algorithms are used to minimize the impact of quantization on model performance, ensuring that the model can still accurately predict the value of disaster impact even when parameter accuracy is reduced.

[0085] Step S230: When the disaster underreporting rate exceeds the preset safety threshold, the incremental learning mode is activated, and the latest acquired meteorological disaster case data is input into the deep reinforcement learning recognition model for local parameter fine-tuning.

[0086] In this embodiment, the preset safety threshold is set to 15%. When the monitored disaster underreporting rate exceeds 15%, the incremental learning mode is activated. For example, if a new meteorological disaster case data is recently acquired, and this case has some unique combinations of meteorological elements and geospatial features, this case data is input into the deep reinforcement learning recognition model. Then, the model performs local parameter fine-tuning based on the new data. For example, for the special meteorological element association features in the new case, the relevant encoding parameters in the feature encoding layer are adjusted so that the model can better encode and recognize these features. For the spatial relationship parsing layer, the weight parameters in the spatial topology relationship matrix are fine-tuned according to the characteristics of the geospatial features in the new case to more accurately reflect the impact of geospatial factors on meteorological disasters. In terms of the value network, the calculation weights of relevant parameters such as the disaster economic loss prediction value, personnel safety risk index, and infrastructure vulnerability level are adjusted according to the disaster impact in the new case, thereby improving the model's ability to identify new or special meteorological disasters.

[0087] Step S240: Dynamically adjust the model compression rate and incremental learning frequency through the delay-accuracy balance algorithm to maintain a preset balance between the generation efficiency and recognition accuracy of the meteorological disaster identification result set.

[0088] In this embodiment, the delay-accuracy balance algorithm dynamically adjusts the model compression ratio and incremental learning frequency based on real-time monitored data to identify delay and disaster underreporting rates, as well as the model's performance after lightweight processing and incremental learning. For example, when the data identification delay index decreases but remains close to the preset response threshold, while the disaster underreporting rate increases but does not exceed the preset safety threshold, the algorithm appropriately reduces the model compression ratio, decreasing the compression ratio of the number of neurons in the feature encoding layer and the dimension of the hidden layer in the policy network. Simultaneously, it increases the incremental learning frequency, increasing the input of new meteorological disaster case data and the number of parameter fine-tuning operations, thereby improving identification accuracy while ensuring generation efficiency. Conversely, when the data identification delay index is far below the preset response threshold, and the disaster underreporting rate is also low, the algorithm appropriately increases the model compression ratio to further reduce the model's computational resource consumption. Simultaneously, it reduces the incremental learning frequency, minimizing unnecessary parameter fine-tuning, to achieve a preset balance between generation efficiency and identification accuracy. This dynamic adjustment mechanism ensures that the meteorological disaster monitoring system can operate efficiently and accurately under various practical conditions.

[0089] In yet another optional embodiment, the method further includes:

[0090] Step S310: Construct a meteorological disaster identification and verification matrix, which includes a historical disaster case library, a set of simulated disaster scenarios, and real-time monitoring and comparison data.

[0091] In this embodiment, the historical disaster case database collects various meteorological disaster cases that have occurred in the urban area over the past many years. Each case records meteorological monitoring data at the time of the disaster, including multi-dimensional meteorological time-series data and geospatial coding data, as well as information such as the corresponding disaster type, impact range, and evolution path. For example, a rainstorm disaster case was recorded, with temperature time-series data of [25℃, 24℃, 23℃, ...], humidity time-series data of [70%, 72%, 75%, ...], and geospatial coding data showing that the affected area was mainly in the low-lying areas of the city, the disaster type was rainstorm, the impact range covered several blocks, and the disaster evolution path was from a certain area to the surrounding areas.

[0092] The simulated disaster scenario set is a collection of various possible meteorological disaster scenarios generated through computer simulation. Using meteorological models and geographic information systems (GIS), different meteorological parameters and geospatial conditions are input to generate multiple simulated disaster scenarios. For example, simulating a hurricane disaster scenario involves setting initial meteorological parameters such as wind speed, wind direction, and air pressure, as well as geospatial parameters such as the topography and building distribution of an urban area. The simulation then tracks the formation, development, and impact of the hurricane in the region, recording meteorological data and disaster-related information during the simulation process.

[0093] Real-time monitoring and comparison data refers to the comparison information between actual meteorological data and model predictions collected in real time during the operation of the meteorological disaster monitoring system. For example, if the model predicts a certain meteorological disaster based on current meteorological data, the system records the differences between the actual meteorological conditions and the model-predicted disaster type, impact range, evolution path, etc. These historical disaster case databases, simulated disaster scenario sets, and real-time monitoring and comparison data are combined to form a meteorological disaster identification and verification matrix.

[0094] Step S320: Compare the meteorological disaster identification result set output by the deep reinforcement learning identification model with the meteorological disaster identification verification matrix in multiple dimensions. The multiple dimensions comparison includes: disaster type matching degree verification, impact range overlap degree verification, and evolution path similarity verification.

[0095] In this embodiment, when verifying the disaster type matching degree, the disaster type identifier output by the deep reinforcement learning recognition model is compared with the disaster types in the historical disaster case library and the simulated disaster scenario set in the meteorological disaster recognition verification matrix. For example, if the disaster type output by the model is "rainstorm", all cases marked as "rainstorm" are searched in the historical disaster case library and the simulated disaster scenario set. The ratio of the number of times the model correctly identifies it as "rainstorm" to the total number of comparisons is used as the disaster type matching degree. For example, if a total of 50 comparisons are made and the model correctly identifies it as "rainstorm" 40 times, then the disaster type matching degree = 40 ÷ 50 = 0.8.

[0096] For verifying the overlap of impact areas, the disaster impact area output by the model is compared with the data in the verification matrix. For example, if the model predicts that the impact area of ​​a rainstorm disaster covers 5 city blocks, corresponding cases are found in the historical disaster case database and the simulated disaster scenario set. The ratio of the overlap area between the model-predicted impact area and the actual or simulated impact area to the total area of ​​both is calculated. For example, if the overlap area is 3 city blocks and the total area is 8 city blocks, then the overlap of impact areas = 3 ÷ 8 = 0.375.

[0097] In terms of evolution path similarity verification, the disaster evolution prediction paths output by the model are compared with the data in the validation matrix. For cases in the historical disaster case database, detailed records of their actual disaster evolution processes are obtained, including information such as the development location and intensity changes of the disaster at different time points. For the simulated disaster scenario set, simulated disaster evolution path data are also available.

[0098] The disaster evolution prediction path output by the model is analyzed in detail according to time series and spatial location information, and compared point-to-point and time-to-time comparisons are made with the evolution paths in historical cases and simulated scenarios. For example, the path predicted by the model for the disaster to spread from region A to region B in the next hour is compared with the spread paths of similar disasters in historical cases within the same time period. By calculating factors such as the overlap of key locations on the path, the consistency of the path direction, and the degree of matching of the disaster development speed in different time periods, a similarity score is obtained.

[0099] By verifying and comparing these three dimensions (disaster type matching degree, overlap of impact range, and similarity of evolution path), the accuracy and reliability of the output results of the deep reinforcement learning recognition model can be comprehensively and accurately evaluated.

[0100] Step S330: When there are abnormal identification results in the comparison results that exceed the preset error threshold, the model interpretability analysis module is triggered to generate a visual analysis report containing feature weight distribution map, strategy decision path tracing and value assessment basis.

[0101] In this embodiment, the preset error threshold is set according to the actual application requirements and the accuracy requirements of the model. For example, the preset error threshold is set to 0.7 for disaster type matching degree; 0.4 for overlap of influence range; and 0.5 for evolution path similarity.

[0102] When the value of a certain dimension in the comparison results is lower than the corresponding preset error threshold, it is determined that there is an abnormal identification result, and the model interpretability analysis module is triggered.

[0103] The process of generating the feature weight distribution map is as follows: The model interpretability analysis module deeply analyzes the weight contribution of each feature (such as temperature-humidity coupling features, pressure gradient change features, and topographic elevation correlation features) to the final disaster identification result when the deep reinforcement learning identification model processes a set of meteorological spatiotemporal features. Based on existing algorithms or calculation methods, the weight value of each feature can be determined and displayed graphically. For example, a bar chart can be used, with the horizontal axis representing different feature names and the vertical axis representing feature weight values, intuitively showing which features play a key role in the model's decision-making and which features have relatively low weights.

[0104] In terms of tracing the strategic decision-making path, the module starts from the model's input layer and analyzes the processing procedures and decision-making basis of each layer (such as the feature encoding layer, spatial relationship analysis layer, time series analysis layer, policy network, and value network) along the data propagation path within the model. It records how data is transformed, features are extracted and combined, and how the final decision is based on these processing results in each layer. For example, in the policy network, it analyzes how joint policy evaluation is performed based on meteorological element feature vectors, spatial topological relationship matrices, and meteorological state transition probability maps, and specifically which calculation methods and rules are used to determine the disaster triggering condition matching degree, disaster spread path similarity, and disaster evolution stage score. Through this tracing, the complete process of the model making decisions is clearly demonstrated.

[0105] The generation of value assessment criteria involves a detailed analysis of the various factors and calculation methods used by the value network in generating disaster economic loss predictions, personnel safety risk indices, and infrastructure vulnerability levels within the meteorological disaster identification result set. For example, it demonstrates how the value network combines historical data, current meteorological conditions, geospatial information, and the model's own parameter settings to calculate disaster economic loss predictions. It explains which factors, such as population density and building structure, are considered when calculating the personnel safety risk index, and how these factors are integrated and calculated using specific algorithms and model structures. For the infrastructure vulnerability level assessment, it elucidates which infrastructure attributes (such as construction date, materials, and maintenance status) are used, and the impact mechanisms of meteorological disasters on these infrastructures.

[0106] These feature weight distribution maps, strategy decision-making path tracing, and value assessment basis are integrated into a single report in a visual manner, forming a visual analysis report that combines intuitive and easy-to-understand graphics, charts, and text descriptions.

[0107] Step S340: Identify the root cause feature dimensions of the deviation based on the location model according to the visualization analysis report, and optimize the preprocessing process of the meteorological spatiotemporal feature set based on the root cause feature dimensions.

[0108] In this embodiment, by carefully studying the visualization analysis report, the root cause of the model identification bias was identified. For example, the feature weight distribution map revealed that the weight of the temperature-humidity coupling feature did not match expectations in the model decision-making process, leading to a deviation in disaster type matching. Further analysis of the strategy decision path tracing information revealed a problem with the encoding method of the temperature-humidity coupling feature at the feature encoding layer. This resulted in the feature not being correctly expressed and utilized in subsequent model processing, thus confirming that the temperature-humidity coupling feature dimension was one of the root causes of the model identification bias.

[0109] Based on this fundamental feature dimension, the preprocessing workflow for the meteorological spatiotemporal feature set was specifically optimized. For the extraction of temperature-humidity coupled features, the algorithms and parameter settings used were re-examined. For example, a simple linear relationship model was initially used to calculate the temperature-humidity coupled features, but analysis revealed that the actual temperature-humidity relationship is more complex and may involve nonlinear relationships. Therefore, the calculation method was adjusted, introducing more complex nonlinear models, such as multinomial regression models or neural network models, to fit the relationship between temperature and humidity.

[0110] During the data collection and processing phase, the sampling frequency and accuracy of temperature and humidity data are increased. For example, instead of recording temperature and humidity data every half hour, the frequency is increased to every 15 minutes, and the accuracy of the measuring equipment is improved to reduce measurement errors. At the same time, more stringent quality control is implemented on the collected data to remove outliers and erroneous data.

[0111] Regarding feature encoding, a more suitable encoding scheme is designed for the new temperature-humidity coupled feature calculation method. For example, an encoding method that can better preserve feature information can be adopted, such as encoding the temperature-humidity coupled feature into a multi-dimensional vector, where each dimension of the vector represents a different feature attribute or feature relationship, so that the model can more accurately understand and utilize the feature in subsequent processing.

[0112] By optimizing the preprocessing flow of meteorological spatiotemporal feature sets based on the root feature dimension, the quality and accuracy of features can be improved, thereby enhancing the recognition performance of deep reinforcement learning recognition models, reducing model recognition bias, and improving the overall accuracy and reliability of meteorological disaster monitoring systems.

[0113] Optionally, the method further includes:

[0114] Step S410: Deploy a distributed model update architecture, including edge computing nodes and a cloud coordination center, in the meteorological disaster monitoring system.

[0115] In this embodiment, the meteorological disaster monitoring system covers a large urban area. To improve system processing efficiency and meet the meteorological data processing needs of different areas, a distributed model update architecture is deployed. Edge computing nodes are distributed throughout the city, close to meteorological monitoring stations. These edge computing nodes have certain computing and storage capabilities. For example, edge computing nodes are set up in the east, west, south, and north areas of the city. Each edge computing node is equipped with a high-performance processor and a certain capacity of memory and storage devices to meet the needs of real-time meteorological data processing. The cloud coordination center is located in the data center and has powerful computing and storage resources. It is responsible for receiving data uploaded by each edge computing node, coordinating and managing the overall data, and updating and optimizing the model. The cloud coordination center is connected to each edge computing node through a high-speed network to ensure rapid data transmission and interaction. The edge computing nodes and the cloud coordination center cooperate to form an organic whole. The edge computing nodes are responsible for processing local meteorological data in real time, reducing the burden on the cloud coordination center; the cloud coordination center analyzes and processes global data, updates the model, and promptly sends the updated model parameters to the edge computing nodes to ensure the consistency and accuracy of the entire system.

[0116] Step S420: Deploy a lightweight recognition model on the edge computing node to process the meteorological spatiotemporal feature set of the area in real time and generate preliminary recognition results.

[0117] In this embodiment, lightweight recognition models are deployed on each edge computing node. These lightweight recognition models are simplified and optimized versions of deep reinforcement learning recognition models, designed to run quickly on resource-constrained edge computing nodes. The lightweight recognition models also receive a set of meteorological spatiotemporal features of their local area as input. For example, an edge computing node located in the eastern part of a city acquires real-time meteorological monitoring data for that area and generates a set of meteorological spatiotemporal features after preprocessing. The lightweight recognition model processes these feature sets, using its internal feature encoding layer, spatial relationship analysis layer, and time series analysis layer for analysis and computation. In the feature encoding layer, a simplified encoding algorithm is used to encode the associated features of meteorological elements. For example, for temperature-humidity coupling features, a faster linear encoding method is used to encode them into a shorter vector. The spatial relationship analysis layer uses a pre-calculated simplified spatial topology matrix to quickly analyze the impact of geospatial factors on meteorology. The time series analysis layer generates a preliminary estimate of the probability of meteorological state transitions through simple statistical methods and models. After the above processing, the lightweight identification model generates preliminary identification results, including a preliminary judgment on the type of meteorological disaster that may occur and a rough estimate of the scope of the disaster's impact. For example, the preliminary identification results may indicate that there is a high probability of a rainstorm disaster in the eastern district in the next few hours, and the impact may extend to several city blocks. Edge computing nodes store these preliminary identification results in real time and upload them to the cloud coordination center when necessary.

[0118] Step S430: Deploy the full recognition model in the cloud coordination center, and receive the abnormal feature data and high-risk recognition results uploaded by each edge computing node for review and analysis.

[0119] In this embodiment, the full-scale identification model deployed in the cloud coordination center is a complete, complex, and highly accurate deep reinforcement learning identification model. It possesses more powerful computing capabilities and richer parameters, enabling deeper and more accurate analysis of meteorological data. After processing the meteorological spatiotemporal feature set to generate preliminary identification results, each edge computing node evaluates the results. If certain feature data is found to be abnormal, or if the preliminary identification results indicate a high-risk situation (such as the potential for severe meteorological disasters), these abnormal feature data and high-risk identification results are uploaded to the cloud coordination center. For example, when processing meteorological data, the edge computing node in the western part of the city finds that a set of meteorological element correlation features differs significantly from previous data, initially judging it to be a possible new meteorological disaster pattern, and thus uploads this set of abnormal feature data to the cloud coordination center. Simultaneously, if the edge computing node initially identifies a region as potentially vulnerable to hurricane disasters, this is also a high-risk identification result and is promptly uploaded to the cloud coordination center.

[0120] After receiving the uploaded data, the cloud-based coordination center's full-volume identification model performs a review and analysis. Utilizing its more comprehensive feature encoding methods, more accurate spatial relationship analysis, and more sophisticated time-series analysis techniques, the full-volume identification model re-evaluates and analyzes anomalous feature data and high-risk identification results. For example, for uploaded anomalous meteorological element correlation features, the full-volume identification model analyzes them from multiple perspectives, considering the more complex interrelationships between different meteorological elements, as well as the impact of geospatial and time-series factors on these features, to determine whether these anomalous features truly represent a new meteorological disaster pattern or are caused by data errors. For high-risk identification results, the full-volume identification model further precisely calculates the probability of disaster occurrence, the scope of impact, and the evolution path, providing a more accurate basis for subsequent decision-making.

[0121] Step S440: Establish an edge-cloud model parameter synchronization mechanism. When the cloud-based full-scale identification model detects a new disaster mode, the updated policy network parameters are encrypted and transmitted to the corresponding edge computing node.

[0122] In this embodiment, to ensure the collaborative operation of the models on the edge computing nodes and the cloud coordination center, and to respond promptly to emerging meteorological disasters, an edge-cloud model parameter synchronization mechanism is established. During the review and analysis of data uploaded by each edge computing node, the cloud-based full-scale identification model adjusts and updates its model parameters if a new disaster pattern is detected. For example, when a new combination of meteorological elements is found to be associated with a specific disaster type, the cloud-based full-scale identification model adjusts the weight parameters of the corresponding features in the policy network to improve its ability to identify this new disaster pattern. The updated policy network parameters are encrypted. Existing encryption algorithms, such as AES encryption, are used to encrypt the parameters, ensuring security and confidentiality during transmission. The encrypted parameters are transmitted to the corresponding edge computing node via a high-speed network. For example, when the cloud-based full-scale identification model detects a new disaster pattern in the southern part of the city, it encrypts the updated policy network parameters and sends them to the edge computing node located in the southern area via the network.

[0123] Upon receiving the encrypted parameters, the edge computing node first decrypts them using a pre-shared key. After successful decryption, the updated policy network parameters are applied to the local lightweight identification model. Following a defined update process, the edge computing node replaces the old parameters with the new ones, ensuring the lightweight identification model can promptly identify new disaster patterns. Through this edge-cloud model parameter synchronization mechanism, the entire meteorological disaster monitoring system can quickly adapt to new meteorological disaster situations, improving the overall performance and adaptability of the system.

[0124] Step S450: Optimize the spatial coverage impact weight and new disaster adaptation impact weight of the meteorological disaster monitoring system through the coordinated operation of local response of edge computing nodes and global strategy optimization in the cloud.

[0125] In this embodiment, the localized response of edge computing nodes and the global strategy optimization in the cloud work together to optimize the relevant weights of the meteorological disaster monitoring system. During the real-time processing of the meteorological spatiotemporal feature set of the area and the generation of preliminary identification results, the edge computing nodes adjust the spatial coverage influence weights based on local conditions. For example, in the area where a certain edge computing node is located, due to the special geographical environment (such as mountains, rivers, and other terrain factors), certain meteorological elements have a more significant impact on disasters. The edge computing node will appropriately increase the spatial coverage influence weights corresponding to these key meteorological elements based on these local characteristics. For instance, an edge computing node in a mountainous area finds that terrain elevation has a significant impact on precipitation; when calculating the spatial coverage influence weights, it will increase the weight values ​​of meteorological elements related to terrain elevation, enabling the model to more accurately identify and predict meteorological disasters in that area.

[0126] Cloud-based global strategy optimization considers the system's performance and adaptability from an overall perspective. When the cloud-based full-scale identification model detects a new disaster pattern and updates the policy network parameters, it adjusts the weights of the new disaster's adaptive impact. For example, for a newly discovered meteorological disaster pattern, the cloud-based full-scale identification model analyzes the probability and severity of this disaster in different geographical regions, and adjusts the weights of different regions and meteorological elements' adaptive impact on the new disaster based on the analysis results. For areas prone to this new disaster, the weights of relevant meteorological elements and geospatial features in that region are increased so that the model can better identify and respond to this new disaster.

[0127] By leveraging the collaborative operation of localized response from edge computing nodes and global strategy optimization in the cloud, the spatial coverage impact weight and the new disaster adaptation impact weight of the meteorological disaster monitoring system are continuously adjusted and optimized. This collaborative optimization mechanism enables the system to maintain high identification accuracy and adaptability in different geographical regions and when facing various meteorological disaster situations, thereby improving the overall effectiveness of the meteorological disaster monitoring system.

[0128] Optionally, the method further includes:

[0129] Step S510: Perform feature matching between the disaster type identifier in the meteorological disaster identification result set and the causal chain nodes in the historical disaster case database to generate a disaster causal association vector.

[0130] In this embodiment, the meteorological disaster identification result set includes information such as the type identifier of the current meteorological disaster. The historical disaster case database stores a large number of past meteorological disaster cases, each of which records detailed information on the causal chain nodes of the disaster. For the disaster type identifier in the meteorological disaster identification result set, such as "rainstorm disaster," all cases marked as "rainstorm disaster" are searched in the historical disaster case database. For each matching case, its causal chain nodes are analyzed. Causal chain nodes may include meteorological factors (such as abnormal atmospheric circulation, water vapor transport, etc.), geospatial factors (such as topography, urban layout, etc.), and other related factors.

[0131] The features of these causal chain nodes are extracted and organized. For example, for a case of a rainstorm disaster, the causal chain node features might include: a specific atmospheric circulation pattern causing a large amount of water vapor to accumulate in the area, and the mountainous terrain surrounding the city blocking the diffusion of water vapor, resulting in precipitation being concentrated in that area. These features are then digitally represented, converted into vector form. For example, a 10-dimensional vector can be used to represent these features, with each dimension corresponding to a different causal chain node feature. For instance, the first dimension represents a feature value of the atmospheric circulation pattern, the second dimension represents a feature value of water vapor transport, and the third dimension represents a feature value of terrain elevation, etc.

[0132] By processing the causal chain node features of multiple matching historical disaster cases, and comprehensively calculating these vectors (e.g., by averaging or weighted averaging), a disaster causal correlation vector is generated. This vector comprehensively reflects the causal features of historical disasters related to the current disaster type, providing an important basis for subsequent analysis and decision-making.

[0133] Step S520: Input the disaster cause association vector into the graph reasoning module of the pre-constructed meteorological disaster knowledge graph to trigger the case strategy node associated with the geospatial distribution characteristics.

[0134] In this embodiment, the pre-constructed meteorological disaster knowledge graph contains rich knowledge and information related to meteorological disasters, as well as the relationships between these information. For example, the generated disaster cause association vectors are input into the graph reasoning module of the meteorological disaster knowledge graph. The graph reasoning module searches and matches in the knowledge graph based on the disaster cause features represented by the vectors. For example, if the disaster cause association vectors contain specific atmospheric circulation patterns and topographic features, the graph reasoning module will search for nodes and relationships associated with these features in the knowledge graph.

[0135] In the knowledge graph, case strategy nodes associated with geospatial distribution characteristics store response strategies and case information for meteorological disasters under different geospatial conditions. When the graph reasoning module finds matching nodes and relationships based on the disaster cause association vector, these case strategy nodes associated with geospatial distribution characteristics are triggered. For example, if it is found that the current disaster cause is similar to a certain type of rainstorm disaster in history, and that this type of rainstorm disaster has corresponding response strategies and case records in a specific geospatial distribution (such as low-lying urban areas), these related case strategy nodes will be triggered to obtain relevant information and strategies.

[0136] Step S530: Extract response strategy templates that satisfy the meteorological change trend characteristics from the case strategy nodes, and generate a set of derived strategies that include cross-case resource scheduling rules, multi-hazard chain priority ranking, and new disaster analogy thresholds.

[0137] In this embodiment, the triggered case strategy node contains multiple response strategy templates, which are designed for different meteorological conditions and geographic situations. Specifically, based on the current meteorological trend characteristics, a matching response strategy template is selected from the case strategy node. For example, if the current meteorological trend indicates that precipitation will continue to increase and wind speed will gradually increase, a response strategy template targeting this meteorological trend is searched in the case strategy node.

[0138] The selected response strategy templates were further organized and extracted. Regarding cross-case resource allocation rules, the methods and principles of resource allocation in responding to similar meteorological disasters were analyzed across different cases. For example, in multiple rainstorm disaster cases, the rule was summarized that when the disaster's impact area is large, priority should be given to allocating flood control materials from warehouses closer to the affected area with sufficient reserves.

[0139] Regarding the prioritization of multiple disaster chains, considering that in reality, multiple meteorological disasters may occur simultaneously, or one disaster may trigger a chain reaction of other disasters, the methods for determining the priority of different disaster chains in the case studies are analyzed. For example, when heavy rain may trigger floods and landslides, the priority principle is to prioritize responding to flood disasters to ensure the safety of people's lives, based on factors such as the population density and importance of infrastructure in the affected area, and then addressing landslide disasters.

[0140] The threshold for analogy to new disasters is determined by setting a standard based on the degree of similarity between current meteorological disasters and historical cases. For example, the similarity score between the current disaster and historical cases is calculated in terms of meteorological elements, geospatial characteristics, and disaster evolution trends. When the similarity score reaches 80% (the set threshold) or above, the current disaster is considered to have a high degree of similarity to historical cases, and the response strategies of historical cases can be used as a reference.

[0141] These cross-case resource scheduling rules, multi-hazard chain priority ranking, and new disaster analogy thresholds are organized and combined to generate a set of derivative strategies. This set of derivative strategies integrates the experience and wisdom of historical cases and can provide more comprehensive and targeted strategic guidance for responding to current meteorological disasters.

[0142] Step S540: Perform instruction-level fusion of the derived strategy set with the emergency resource scheduling strategy in the disaster response strategy set to generate an enhanced emergency resource scheduling strategy that includes a dynamic priority adjustment field and a resource conflict resolution protocol.

[0143] In this embodiment, the derived strategy set is deeply integrated with the emergency resource dispatch strategy in the disaster response strategy set. First, a detailed analysis is performed on each instruction and parameter in the emergency resource dispatch strategy, such as the location of material reserves, transportation route planning, and allocation priority.

[0144] For the dynamic priority adjustment field, the allocation priority in the emergency resource dispatch strategy is adjusted according to the multi-hazard chain priority ranking principle in the derived strategy set. For example, if the derived strategy set indicates that under current meteorological conditions, a certain disaster-stricken area requires a higher resource allocation priority due to the possibility of triggering more severe secondary disasters, then in the emergency resource dispatch strategy, the allocation priority of that area is raised from the original level 3 to level 1 (for example, priorities are divided into levels 1-5, with level 1 being the highest). Simultaneously, the priorities of each disaster-stricken area are continuously monitored and dynamically adjusted based on real-time meteorological trends and disaster development. For example, if it is found that the disaster situation in another area is worsening, potentially affecting more critical infrastructure and the safety of a large number of people, the priority of that area is promptly raised.

[0145] Regarding resource conflict resolution protocols, detailed solutions are developed by incorporating cross-case resource scheduling rules from the derived strategy set. For example, when multiple disaster-stricken areas simultaneously request the same emergency resource (such as flood control sandbags), the cross-case resource scheduling rules prioritize the needs of areas closer to the disaster source and with higher disaster severity. If different types of emergency resources have route conflicts during transportation, the priority transportation route is determined based on the urgency and importance of the resources. For example, the transportation of medical relief supplies takes precedence over general flood control supplies to ensure the safety of disaster-stricken people.

[0146] Through the aforementioned instruction-level fusion, effective information and rules from the derived strategy set are integrated into the emergency resource dispatch strategy, generating an enhanced emergency resource dispatch strategy. This strategy not only considers the current disaster situation but also incorporates the experience and response strategies from historical cases, enabling it to respond more flexibly and effectively to complex and ever-changing meteorological disaster situations and improve the efficiency and rationality of emergency resource dispatch.

[0147] Step S550: Update the infrastructure vulnerability level assessment parameters in the value network of the deep reinforcement learning recognition model using the instruction execution log of the enhanced emergency resource scheduling strategy.

[0148] In this embodiment, the enhanced emergency resource dispatch strategy generates detailed instruction execution logs during execution. These logs record the execution status of each dispatch instruction, including information such as the allocation time, arrival location, and actual usage of the materials.

[0149] Analyze these command execution logs to extract information related to infrastructure vulnerability. For example, records of deploying a large amount of power equipment to restore power in a disaster-stricken area indicate that the area's power infrastructure was significantly affected by the meteorological disaster and may have vulnerabilities. Assess the vulnerability level by statistically analyzing the demand and damage levels of different types of infrastructure (such as power, communications, and transportation) during emergency resource allocation.

[0150] Based on these assessment results, the infrastructure vulnerability level assessment parameters in the value network of the deep reinforcement learning identification model are updated. For example, if it is found that the communication infrastructure in a certain region requires a large amount of emergency resources for repair and protection in response to multiple meteorological disasters, it indicates that the communication infrastructure in that region is highly vulnerable, and the vulnerability level assessment parameters for the corresponding region's communication infrastructure in the value network are increased. Conversely, if a certain infrastructure demonstrates strong stability in disaster response and requires fewer emergency resources, its vulnerability level assessment parameters are appropriately decreased.

[0151] By continuously updating the infrastructure vulnerability assessment parameters in the value network using the instruction execution log of the enhanced emergency resource scheduling strategy, the deep reinforcement learning identification model can more accurately reflect the vulnerability of infrastructure under meteorological disasters in reality. This allows for a more reasonable consideration of infrastructure factors in future meteorological disaster identification and response strategy generation processes, improving the model's accuracy and practicality, and providing support for more effective meteorological disaster response.

[0152] In a non-limiting embodiment, the method further includes:

[0153] Step S610: Input the geospatial encoded data into the voxel modeling module of the 3D visualization sand table module to generate a 3D base model containing terrain elevation gradient and building outline. Map the wind speed-precipitation synergistic features in the meteorological spatiotemporal feature set to the atmospheric particle simulator of the 3D base model to generate a 3D dynamic scene with fluid dynamics properties.

[0154] In this embodiment, geospatial encoded data is input into the voxel modeling module of the 3D visualization sandbox module. The geospatial encoded data contains detailed information such as the terrain elevation and urban building distribution of each grid within the urban area. The voxel modeling module processes this data, dividing the entire urban area into tiny voxels (similar to pixels in 3D space).

[0155] For each voxel, its height value is determined based on the terrain elevation information in the geospatial coding data, thereby constructing a 3D terrain model that reflects the terrain elevation gradient. For example, if the grid terrain elevation corresponding to a certain voxel is 50 meters, this voxel is placed at the corresponding height position in the 3D model. Simultaneously, based on urban building distribution information, the outlines of buildings within the voxels are determined. If a voxel is located within a building area, the voxel is divided and processed accordingly based on the building's shape and boundary information to form the building's 3D outline. Through the processing of all voxels, a 3D base model containing the terrain elevation gradient and building outlines is generated.

[0156] Next, the wind speed-precipitation co-feedback features from the meteorological spatiotemporal feature set are mapped to the atmospheric particle simulator of the 3D baseline model. The wind speed-precipitation co-feedback features contain the correlation information between wind speed and precipitation at different locations and times. The atmospheric particle simulator simulates the motion and interaction of particles in the atmosphere. Wind speed information is mapped to the velocity and direction of atmospheric particles; for example, in an area with a wind speed of 5 m / s, atmospheric particles are set to move at the corresponding velocity and direction in that area. Precipitation information is mapped to the aggregation and sedimentation behavior of atmospheric particles; for example, in areas with high precipitation intensity, the aggregation density of atmospheric particles is increased, and the sedimentation velocity of the particles is set.

[0157] Through this mapping, the atmospheric particle simulator can simulate a three-dimensional dynamic scene with hydrodynamic properties based on a three-dimensional base model. In this scene, the interaction between wind speed and precipitation under different terrains and building outlines can be intuitively observed, such as changes in wind speed and precipitation distribution patterns near mountains, and phenomena such as airflow disturbances and precipitation convergence in densely built urban areas.

[0158] Step S620: Overlay the boundary coordinates of the disaster impact range of the meteorological disaster identification result set onto the three-dimensional dynamic scene to generate a multi-layer disaster impact voxel model with diffusion path growth animation. Import the evacuation path coordinates of the risk area division strategy into the collision detection module of the multi-layer disaster impact voxel model to generate an interactive evacuation simulation scene containing path availability indicators and congestion risk heat maps. Based on the modified path parameters in the interactive evacuation simulation scene, update the urban building distribution feature weights in the spatial relationship parsing layer of the deep reinforcement learning identification model in real time.

[0159] In this embodiment, the boundary coordinates of the disaster impact range from the meteorological disaster identification result set are overlaid on the already generated 3D dynamic scene with fluid dynamic properties. The meteorological disaster identification result set clarifies the boundaries of the areas that the disaster may affect, and these boundary coordinates are marked and displayed in the 3D dynamic scene. For example, for a rainstorm disaster, its impact range boundary coordinates are drawn in the 3D scene, forming a closed region representing the area that the rainstorm may affect. As time progresses, based on the disaster's evolution prediction path, the spread of the disaster impact range is dynamically displayed, generating a multi-layered disaster impact voxel model with a spread path growth animation. This model not only shows the current impact range of the disaster but also intuitively presents the development trend of the disaster.

[0160] Then, the coordinates of the evacuation paths from the risk zone delineation strategy are imported into the collision detection module of the multi-layer disaster impact voxel model. The evacuation path coordinates determine the routes taken by affected people to evacuate from dangerous areas to safe areas. The collision detection module detects the evacuation paths in relation to the disaster impact area, buildings, and other objects. For example, it detects whether the evacuation paths pass through areas experiencing flooding or whether they collide with buildings.

[0161] Based on the collision detection results, path availability indicators are generated. If an evacuation route does not conflict with a hazardous area or building, the path availability indicator is green, indicating that the route is available; if the route is partially or entirely located within a hazardous area or conflicts with obstacles such as buildings, the path availability indicator is red, indicating that the route is unavailable. Simultaneously, by analyzing factors such as pedestrian flow and speed along the evacuation routes, a congestion risk heatmap is generated. For example, in densely populated areas with slow evacuation speeds, red or orange indicates high congestion risk; in areas with relatively smooth pedestrian flow, green or light blue indicates low congestion risk. This generates an interactive evacuation simulation scenario that includes path availability indicators and a congestion risk heatmap.

[0162] In interactive evacuation simulation scenarios, operators can modify path parameters based on actual conditions, such as adjusting evacuation routes and changing evacuation start and end points. When path parameters are modified, the system feeds these changes back to the deep reinforcement learning recognition model in real time. Specifically, based on the relationship between the modified path and the distribution of urban buildings, the impact of urban buildings on evacuation and meteorological disasters is reassessed. For example, if the new evacuation route passes through areas with more high-rise buildings, it indicates that these buildings have a greater impact on evacuation, and the weights of urban building distribution features related to these buildings in the spatial relationship analysis layer of the deep reinforcement learning recognition model are increased accordingly. In this way, the deep reinforcement learning recognition model can adjust parameters in real time according to the actual evacuation simulation, improving its understanding and prediction capabilities of meteorological disasters and personnel evacuation in the urban environment, and further optimizing meteorological disaster response strategies.

[0163] In a non-limiting embodiment, the method further includes:

[0164] Step S710: Collect the actual path trajectory data of the material transportation terminal in the emergency resource scheduling strategy in real time, generate a set of scheduling efficiency indicators including route deviation, transportation speed fluctuation rate and node dwell time, and correct the road traffic weight coefficient in the spatial topology relationship matrix of the deep reinforcement learning recognition model according to the set of scheduling efficiency indicators to generate dynamic road network influence factors.

[0165] In this embodiment, to accurately evaluate the execution effect of the emergency resource dispatch strategy and optimize the deep reinforcement learning recognition model, real-time data on the actual path trajectory of the material transportation terminal is collected. By installing high-precision positioning systems and sensors on terminal equipment such as material transportation vehicles and ships, the location information and operating status data of the equipment can be obtained in real time.

[0166] The collected actual route trajectory data is analyzed and processed to generate a set of scheduling efficiency indicators. Route deviation is calculated by comparing the actual transportation route with the pre-planned route. For example, if the planned transportation route is from warehouse A through roads X and Y to disaster area Z, but the actual transportation route deviates from road Y, the route deviation is calculated by determining the geometrical deviation between the actual and planned routes. For example, if the planned route length is 10 kilometers and the actual route deviates from the planned route by 2 kilometers, then the route deviation = 2 ÷ 10 = 0.2. Transportation speed fluctuation rate is a statistical measure of speed changes during transportation. The speed of the transportation terminal is recorded at different time periods, and the average and standard deviation of the speed are calculated. For example, if the speed during transportation is 30 km / h, 40 km / h, and 35 km / h on different road segments, the average speed is calculated to be 35 km / h. The standard deviation of the speed is calculated using a specific algorithm to measure the transportation speed fluctuation rate. Node dwell time is the statistical measure of the dwell time of the transportation terminal at various intermediate nodes (such as warehouses, transfer stations, etc.). For example, a transport vehicle spends 30 minutes loading goods at warehouse A and 15 minutes at transfer station B for cargo allocation; these dwell times are recorded as part of the node dwell time. These indicators, such as route deviation, transport speed fluctuation rate, and node dwell time, are combined into a scheduling efficiency indicator set. Based on these indicators, the road traffic weight coefficients in the spatial topology matrix of the deep reinforcement learning recognition model are adjusted. For example, if a road has a high route deviation, it indicates potential traffic difficulties or other problems, and its traffic weight coefficient in the spatial topology matrix is ​​reduced; if a road has a low transport speed fluctuation rate, it indicates relatively stable traffic conditions, and its traffic weight coefficient is appropriately increased.

[0167] Through the above corrections, a dynamic road network impact factor is generated. This dynamic road network impact factor can reflect the actual situation of the current road network in the emergency resource transportation process in real time, providing more accurate road network information for the deep reinforcement learning recognition model in the subsequent meteorological disaster identification and response strategy generation, so that the model can more reasonably consider the impact of road traffic conditions on disaster response.

[0168] Step S720: Perform spatiotemporal overlay operation on the dynamic road network influence factor and the meteorological state transition probability map to update the diffusion direction probability distribution of the disaster evolution prediction path. Based on the updated diffusion direction probability distribution, adjust the high-risk area blockade boundary coordinates in the risk area division strategy, and backfeed the adjusted blockade boundary coordinates to the value network of the deep reinforcement learning recognition model for loss function compensation calculation.

[0169] In this embodiment, the generated dynamic road network influence factor is spatiotemporally superimposed with the meteorological state transition probability map. The meteorological state transition probability map describes the probability of different meteorological states transitioning at different times and in different spaces. The dynamic road network influence factor reflects the impact of the road network on resource transportation and personnel movement during meteorological disaster response.

[0170] The spatiotemporal overlay process combines road traffic information from the dynamic road network influence factors with spatial and temporal information from the meteorological state transition probability map. For example, during a certain time period, poor road conditions (shown as low traffic weight by the dynamic road network influence factors) restrict the flow of people and goods between disaster-stricken areas, affecting the speed and direction of meteorological disaster spread. Based on this impact, the probability of disaster spread direction in the corresponding area and time in the meteorological state transition probability map is adjusted. For instance, if the probability of a meteorological disaster spreading in a certain direction in a certain area was originally 0.6, due to road traffic issues, the probability of spread in that direction is adjusted to 0.4, while simultaneously increasing the probability of spread in other feasible directions.

[0171] The aforementioned spatiotemporal overlay operation updates the probability distribution of the disaster evolution prediction path's diffusion direction. Based on this updated probability distribution, the coordinates of the high-risk area blockade boundary in the risk zone delineation strategy are reassessed. If an increase in the probability of disaster diffusion in a new direction is detected, the blockade boundary of the high-risk area in that direction is correspondingly expanded. For example, if the original high-risk area blockade boundary only considered the northeastward diffusion of the disaster, the updated probability distribution reveals a significant increase in the southeastward diffusion probability, thus extending the high-risk area blockade boundary southeastward.

[0172] The adjusted boundary coordinates are fed back to the value network of the deep reinforcement learning recognition model for loss function compensation calculation. When calculating disaster losses (such as economic losses and casualties), the value network needs to consider the extent of high-risk areas. The adjusted boundary coordinates signify a change in the high-risk area. Based on the new high-risk area, the value network recalculates the potentially affected population, infrastructure, and other factors, thereby adjusting the loss function calculation. For example, if the high-risk area expands, potentially involving more residents and critical infrastructure, the value network will correspondingly increase indicators such as the predicted economic loss and the personnel safety risk index. Through this loss function compensation calculation, the deep reinforcement learning recognition model can more accurately reflect the actual impact and potential losses of meteorological disasters, providing a more reliable basis for subsequent decision-making and response strategy optimization.

[0173] This invention, through acquiring multi-dimensional meteorological time-series data and geospatial coding data, can comprehensively grasp the meteorological conditions of a target urban area. By preprocessing meteorological data, it mines the correlation characteristics, geospatial distribution characteristics, and meteorological change trend characteristics of meteorological elements, providing rich and valuable information for subsequent analysis. Utilizing a deep reinforcement learning recognition model for dynamic disaster pattern matching, it can accurately generate a set of meteorological disaster identification results, including disaster type identifiers, impact ranges, and evolution prediction paths. Based on this, the generated disaster response strategy set covers strategies such as emergency resource allocation, risk area delineation, and early warning information dissemination, effectively addressing meteorological disasters. Furthermore, dynamic strategy optimization of the deep reinforcement learning recognition model and its deployment in the meteorological disaster monitoring system can improve model performance and enhance the accuracy and real-time performance of meteorological disaster monitoring.

[0174] Based on the same inventive concept, embodiments of the present invention also provide an urban meteorological disaster data identification system. (See also...) Figure 2 As shown, it is a schematic diagram of the structure of a possible urban meteorological disaster data identification system provided in an embodiment of the present invention. Figure 2 The urban meteorological disaster data identification system 200 includes a processor 210 and a memory 220. The memory 220 stores computer programs executable by the processor 210. By executing the instructions stored in the memory 220, the processor 210 can perform the steps of the aforementioned urban meteorological disaster data identification method based on deep reinforcement learning.

[0175] Based on the same inventive concept, embodiments of the present invention provide a computer-readable storage medium including a computer program. When the computer program is run on an urban meteorological disaster data identification system, the computer program is used to cause the urban meteorological disaster data identification system to perform the steps of the above-described urban meteorological disaster data identification method based on deep reinforcement learning.

[0176] In the technical solutions involved in the above embodiments of the present invention, whether performing comparison calculations of multi-dimensional features or constructing composite parameters, if there are problems caused by significant differences in the number of dimensions, units of measurement, and semantic meanings of different features, those skilled in the art, based on their professional knowledge and past practical experience, can fully understand that these differences need to be properly handled so that the calculation results are accurate and comparable, and to avoid situations such as logical confusion and unclear mathematical meaning.

[0177] The formulas and calculation processes involved in the embodiments of this invention, whether used for multidimensional feature comparison or composite loss function construction, strictly adhere to the principle of dimensional correspondence. Each variable in the formula has a clear and explicit physical meaning, and its operational logic fully conforms to basic mathematical and physical logic. The calculation results are necessarily the reasonable results expected by this invention. Those skilled in the art are capable of effectively solving various problems arising from the number of dimensions, dimensional differences, etc., in the multidimensional feature comparison calculation and composite loss function construction in the embodiments, based on specific data conditions and business needs, by comprehensively utilizing the above-mentioned general technical means, thus ensuring the accuracy, reliability, and implementability of the technical solution of this invention.

Claims

1. A method for identifying urban meteorological disaster data based on deep reinforcement learning, characterized in that, The method includes: Obtain a set of meteorological monitoring data for the target urban area; Perform meteorological data preprocessing on the meteorological monitoring data set to obtain a preprocessed set of meteorological spatiotemporal features; The trained deep reinforcement learning recognition model is invoked to perform dynamic disaster pattern matching processing on the meteorological spatiotemporal feature set, thereby generating a meteorological disaster recognition result set for the target urban area. A set of disaster response strategies is generated based on the meteorological disaster identification result set. Based on the disaster response strategy set, the deep reinforcement learning recognition model is dynamically optimized to obtain an optimized deep reinforcement learning recognition model, which is then deployed to the meteorological disaster monitoring system. The disaster response strategy set includes emergency resource allocation strategies, risk area delineation strategies, and early warning information dissemination strategies. The dynamic strategy optimization process performed on the deep reinforcement learning recognition model based on the disaster response strategy set to obtain an optimized deep reinforcement learning recognition model includes: The actual material arrival time data is extracted from the emergency resource dispatch strategy, and the personnel evacuation route execution efficiency data is extracted from the risk area division strategy to generate a set of strategy execution performance indicators. The set of policy execution performance indicators is input into the backpropagation interface of the policy network of the deep reinforcement learning recognition model to calculate the weight gradient deviation matrix of the disaster triggering condition matching degree. Based on the weight gradient deviation matrix, the meteorological influence weight allocation ratio between different geographical units in the spatial topology matrix is ​​adjusted to generate an optimized spatial topology matrix. The actual response delay data of the aforementioned early warning information release strategy is convolved and aligned with the time series of the meteorological state transition probability map to generate the value network loss function correction coefficient. The optimized spatial topology matrix is ​​used to update the output parameters of the spatial relationship analysis layer, and the error calculation weight of the disaster economic loss prediction value is adjusted by the value network loss function correction coefficient. The updated parameters of the policy network and the value network are jointly verified for convergence. When the accuracy of the disaster evolution prediction path is improved to a preset threshold, the optimized deep reinforcement learning recognition model is output.

2. The method as described in claim 1, characterized in that, The meteorological monitoring dataset includes multi-dimensional meteorological time-series data and corresponding geospatial coding data. The meteorological spatiotemporal feature set includes meteorological element correlation features, geospatial distribution features, and meteorological change trend features. Performing meteorological data preprocessing on the meteorological monitoring dataset yields a preprocessed meteorological spatiotemporal feature set, including: The multi-dimensional meteorological time-series data is subjected to abnormal data cleaning processing to remove data fragments exceeding the preset meteorological element fluctuation threshold, generating cleaned meteorological time-series data. The cleaned meteorological time-series data is then subjected to spatiotemporal alignment processing to unify meteorological data with different time resolutions to a preset time base. Spatial grids are then divided according to the geospatial coding data to generate spatiotemporally aligned meteorological grid data. Meteorological element association features are extracted from the spatiotemporally aligned meteorological grid data. The meteorological element association features include temperature-humidity coupling features, pressure gradient change features, and wind speed-precipitation synergistic features. The geospatial encoded data is subjected to spatial topology parsing to generate geospatial distribution features that include terrain elevation correlation features, urban building distribution features, and water system coverage features. Based on the sliding time window, the spatiotemporally aligned meteorological grid data is subjected to trend fitting processing to generate meteorological change trend features that include the cumulative change rate of meteorological elements, the probability of sudden meteorological events, and the characteristics of meteorological state transition. The meteorological spatiotemporal feature set is generated by combining the meteorological element correlation characteristics, the geographic spatial distribution characteristics, and the meteorological change trend characteristics.

3. The method as described in claim 2, characterized in that, The method invokes the trained deep reinforcement learning recognition model to perform dynamic disaster pattern matching processing on the meteorological spatiotemporal feature set, generating a meteorological disaster recognition result set for the target urban area, including: The meteorological element association features are input into the feature encoding layer of the deep reinforcement learning recognition model to generate meteorological element feature vectors. The geospatial distribution features are input into the spatial relationship parsing layer of the deep reinforcement learning recognition model to generate a spatial topological relationship matrix, which includes meteorological influence weights between different geographic units. The meteorological change trend features are input into the time series analysis layer of the deep reinforcement learning recognition model to generate a meteorological state transition probability map. The policy network of the deep reinforcement learning recognition model performs joint policy evaluation on the meteorological element feature vector, the spatial topological relationship matrix, and the meteorological state transition probability map, generating dynamic evaluation results that include disaster triggering condition matching degree, disaster spread path similarity, and disaster evolution stage score. The value network of the deep reinforcement learning recognition model is used to predict the disaster impact value of the dynamic assessment results, generating a set of meteorological disaster identification results that includes predicted disaster economic losses, personnel safety risk index and infrastructure vulnerability level.

4. The method as described in claim 1, characterized in that, The meteorological disaster identification result set includes disaster type identifiers, disaster impact ranges, and disaster evolution prediction paths. The generation of a disaster response strategy set based on the meteorological disaster identification result set includes: Based on the disaster type identifier, a preset emergency resource type library is matched to generate an emergency resource dispatch strategy that includes material storage location, transportation route planning, and allocation priority. Based on the disaster impact range overlaid with the urban population density distribution map, a risk area delineation strategy is generated, which includes high-risk area lockdown plans, personnel evacuation routes, and the setting of temporary resettlement points. Based on the time series characteristics of the disaster evolution prediction path, an early warning information release strategy is generated, which includes the early warning information release level, the priority of the early warning coverage area, and the combination of information dissemination channels. The emergency resource scheduling strategy, risk area division strategy and early warning information release strategy are subjected to strategy conflict detection to eliminate execution time conflicts and resource allocation conflicts between different strategies, and an optimized disaster response strategy set is obtained. The optimized disaster response strategy set is encoded into an instruction format to be executed and pushed to the corresponding emergency management terminal device.

5. The method as described in claim 3, characterized in that, The method further includes: Real-time monitoring of the data identification delay index and disaster underreporting rate index of the meteorological disaster monitoring system; When the data recognition latency index exceeds a preset response threshold, the deep reinforcement learning recognition model is subjected to lightweight processing: the number of neurons in the feature encoding layer is compressed, the hidden layer dimension of the policy network is reduced, and the parameter precision of the value network is quantized. When the disaster underreporting rate exceeds the preset safety threshold, the incremental learning mode is activated, and the latest acquired meteorological disaster case data is input into the deep reinforcement learning recognition model for local parameter fine-tuning. By dynamically adjusting the model compression rate and incremental learning frequency through a delay-accuracy balance algorithm, the generation efficiency and recognition accuracy of the meteorological disaster identification result set are kept in a preset balance.

6. The method as described in claim 1, characterized in that, The method further includes: A meteorological disaster identification and verification matrix is ​​constructed, which includes a historical disaster case library, a set of simulated disaster scenarios, and real-time monitoring and comparison data. The meteorological disaster identification result set output by the deep reinforcement learning identification model is compared with the meteorological disaster identification verification matrix in multiple dimensions. The multi-dimensional comparison includes: disaster type matching degree verification, impact range overlap degree verification, and evolution path similarity verification. When there are abnormal identification results in the comparison results that exceed the preset error threshold, the model interpretation analysis module is triggered to generate a visual analysis report that includes feature weight distribution map, strategy decision path tracing and value assessment basis. The root cause feature dimension of the deviation is identified by the location model based on the visualization analysis report, and the preprocessing process of the meteorological spatiotemporal feature set is optimized based on the root cause feature dimension.

7. The method as described in claim 1, characterized in that, The method further includes: The meteorological disaster monitoring system is deployed with a distributed model update architecture that includes edge computing nodes and a cloud coordination center. A lightweight recognition model is deployed on the edge computing node to process the meteorological spatiotemporal feature set of the area in real time and generate preliminary recognition results; A full-scale identification model is deployed in the cloud coordination center to receive abnormal feature data and high-risk identification results uploaded by each edge computing node for review and analysis; Establish an edge-cloud model parameter synchronization mechanism. When the cloud-based full-scale identification model detects a new disaster mode, the updated policy network parameters will be encrypted and transmitted to the corresponding edge computing node. By leveraging the collaborative operation of localized response at edge computing nodes and global strategy optimization in the cloud, the spatial coverage impact weight and the new disaster adaptation impact weight of the meteorological disaster monitoring system are optimized.

8. The method as described in claim 1, characterized in that, The method further includes: The disaster type identifiers in the meteorological disaster identification result set are matched with the causal chain nodes in the historical disaster case database to generate a disaster causal association vector; The disaster cause correlation vector is input into the graph reasoning module of the pre-constructed meteorological disaster knowledge graph to trigger case strategy nodes associated with geospatial distribution characteristics; Extract response strategy templates that meet the meteorological change trend characteristics from the case strategy nodes, and generate a set of derived strategies that include cross-case resource scheduling rules, multi-hazard chain priority ranking, and new disaster analogy thresholds; The derived strategy set is fused with the emergency resource scheduling strategy in the disaster response strategy set at the instruction level to generate an enhanced emergency resource scheduling strategy that includes a dynamic priority adjustment field and a resource conflict resolution protocol. The instruction execution log of the enhanced emergency resource scheduling strategy is used to update the infrastructure vulnerability level assessment parameters in the value network of the deep reinforcement learning recognition model.

9. A system for identifying urban meteorological disaster data, characterized in that, It includes a processor and a memory, wherein the memory stores a computer program that, when executed by the processor, causes the processor to perform the steps of any one of the methods described in claims 1 to 8.

Images