Urban meteorological disaster data identification method and system based on deep reinforcement learning

Abstract

The invention relates to the technical field of data analysis, provides an urban meteorological disaster data identification method and system based on deep reinforcement learning, and is used for effectively improving the model performance so as to enhance the accuracy and real-time performance of meteorological disaster monitoring. The method comprises the steps of obtaining a meteorological monitoring data set of a target city area, executing meteorological data preprocessing operation on the meteorological monitoring data set to obtain a preprocessed meteorological spatial-temporal feature set, calling a trained deep reinforcement learning recognition model, and performing dynamic disaster mode matching processing on the meteorological spatial-temporal feature set to obtain a dynamic disaster mode recognition model. And generating a meteorological disaster recognition result set of the target city region, generating a disaster coping strategy set according to the meteorological disaster recognition result set, and performing dynamic strategy optimization processing on the deep reinforcement learning recognition model based on the disaster coping 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] The present invention belongs to the technical field of data analysis, and particularly relates to a method and system for identifying urban meteorological disaster data based on deep reinforcement learning. Background Art

[0002] With the acceleration of the urbanization process, the threat of meteorological disasters to the safe operation of cities has become increasingly prominent. The current urban meteorological disaster analysis technology mainly relies on the discrete data collection of meteorological observation stations and the deduction of numerical prediction models, and realizes risk early warning by constructing a disaster level evaluation index system. Traditional technologies mostly use machine learning algorithms to perform pattern matching on historical disaster cases and combine geographic information systems to display the disaster impact range. However, the above technologies are difficult to effectively integrate the temporal variation characteristics of meteorological elements and the distribution characteristics of geographical space, resulting in the loss of spatio-temporal correlation information during the disaster identification process; the machine learning model based on a static training set lacks the dynamic response ability to real-time meteorological changes, is prone to misjudgment during disaster pattern matching, and the above technologies are difficult to adapt to the situation of the coupled evolution of multiple disaster types in a complex urban environment. Therefore, how to effectively improve the model performance to enhance the accuracy and real-time nature of meteorological disaster monitoring is a technical problem that needs to be solved at present. Summary of the Invention

[0003] The present invention provides a method and system for identifying urban meteorological disaster data based on deep reinforcement learning, so as to effectively improve the model performance to enhance the accuracy and real-time nature of meteorological disaster monitoring.

[0004] In a first aspect, an embodiment of the present invention provides a method for identifying urban meteorological disaster data based on deep reinforcement learning, which is applied to an urban meteorological disaster data identification system. The method includes: obtaining a set of meteorological monitoring data for a target urban area, where the set of meteorological monitoring data includes multi-dimensional meteorological time series data and corresponding geographic space coding data; performing meteorological data preprocessing operations on the set of meteorological monitoring data to obtain a preprocessed set of meteorological spatio-temporal characteristics, where the set of meteorological spatio-temporal characteristics includes meteorological element correlation characteristics, geographic space distribution characteristics, and meteorological change trend characteristics; calling a trained deep reinforcement learning identification model to perform dynamic disaster pattern matching processing on the set of meteorological spatio-temporal characteristics to generate a set of meteorological disaster identification results for the target urban area, where the set of meteorological disaster identification results includes disaster type identifiers, disaster impact ranges, and disaster evolution prediction paths; generating a set of disaster response strategies based on the set of meteorological disaster identification results, where the set of disaster response strategies includes emergency resource scheduling strategies, risk area division strategies, and early warning information release strategies; performing dynamic strategy optimization processing on the deep reinforcement learning identification model based on the set of disaster response strategies to obtain an optimized deep reinforcement learning identification model, and deploying the optimized deep reinforcement learning identification model to a meteorological disaster monitoring system.

[0005] In a second aspect, an embodiment of the present invention provides 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 executes the steps of the above method.

[0006] In a third aspect, an embodiment of the present invention provides a computer-readable storage medium, which includes a computer program. When the computer program is run on an urban meteorological disaster data identification system, the computer program is used to enable the urban meteorological disaster data identification system to execute the steps of the above method.

[0007] In the implementation of the present invention, by acquiring multi-dimensional meteorological time series data and geographic space coding data, the meteorological conditions in the target urban area can be fully understood; by mining the meteorological element correlation characteristics, geographic space distribution characteristics and meteorological change trend characteristics through meteorological data preprocessing operations, it can provide rich and valuable information for subsequent analysis; using the deep reinforcement learning recognition model for dynamic disaster pattern matching, it can accurately generate a set of meteorological disaster identification results containing disaster type identification, impact range and evolution prediction path, and the disaster response strategy set generated based on this covers emergency resource scheduling, risk area division and early warning information release strategies, which can effectively respond to meteorological disasters. Furthermore, the deep reinforcement learning recognition model is dynamically optimized and deployed to the meteorological disaster monitoring system, which can improve the model performance and thus enhance the accuracy and real-time performance of meteorological disaster monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

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

[0009] Figure 2 The present invention provides a schematic diagram of the structure of a system for identifying urban meteorological disaster data. DETAILED DESCRIPTION

[0010] See also Figure 1 , which is a method for identifying urban meteorological disaster data based on deep reinforcement learning provided in an embodiment of the present invention. The method can be applied to an urban meteorological disaster data identification system. The specific process is as follows: step S110 to step S150.

[0011] Step S110: Acquire a meteorological monitoring data set of a target urban area, wherein the meteorological monitoring data set includes multi-dimensional meteorological time series data and corresponding geographic space coding data.

[0012] In this embodiment, a large urban area is used as the target area, and multiple meteorological monitoring stations are distributed in the target area to continuously collect various types of meteorological data. Multi-dimensional meteorological time series data contains records of changes in various meteorological elements over time. For example, the temperature element is recorded every half an hour starting from 6 o'clock in the morning to obtain a set of data such as [18℃, 18.5℃, 19℃, 19.5℃, 20℃, ...]; the humidity element is recorded synchronously to form 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 interval, together forming multi-dimensional meteorological time series data.

[0013] Optionally, in terms of geospatial coding data, the urban area is divided into grids of 1 square kilometer in size, and each grid has a unique code. For each grid, the terrain elevation is recorded, such as the grid numbered G10, which has a terrain elevation of 45 meters; the distribution of urban buildings is recorded, and the building area in the grid accounts for 40% of the grid area, and is mainly residential buildings; at the same time, the water system coverage is recorded, and there is a stream in the grid, and the water system coverage area accounts for 5%. The above grid data constitutes the geospatial coding data.

[0014] Step S120: performing meteorological data preprocessing operations on the meteorological monitoring data set to obtain a preprocessed meteorological spatiotemporal feature set, wherein the meteorological spatiotemporal feature set includes meteorological element association features, geographic space distribution features, and meteorological change trend features.

[0015] In one implementation, a meteorological data preprocessing operation is performed on the meteorological monitoring data set to obtain a preprocessed meteorological spatiotemporal feature set, wherein the meteorological spatiotemporal feature set includes meteorological element association features, geographic space distribution features, and meteorological change trend features, including: Step S121: performing abnormal data cleaning processing on the multi-dimensional meteorological time series data, eliminating data segments that exceed a preset meteorological element fluctuation threshold, generating cleaned meteorological time series data, performing spatiotemporal alignment processing on the cleaned meteorological time series data, unifying meteorological data of different time resolutions to a preset time reference, performing spatial grid division according to the geographic space coding data, generating spatiotemporal aligned meteorological grid data, extracting meteorological element correlation features in the spatiotemporal aligned meteorological grid data, the meteorological element correlation features including temperature-humidity coupling features, pressure gradient change features and wind speed-precipitation synergy features.

[0016] In this embodiment, for abnormal data cleaning, taking the temperature element as an example, by analyzing historical data, the normal fluctuation range of the temperature in this urban area is between 10°C and 35°C, and the temperature change in adjacent half-hours generally does not exceed 2°C. Therefore, the temperature fluctuation threshold is set to 2°C. In the currently collected temperature time-series data, if the difference between two adjacent data points exceeds 2°C, such as suddenly changing from 20°C to 23°C, then this data segment is determined to be abnormal and excluded, and the cleaned meteorological time-series data is obtained accordingly.

[0017] When performing spatio-temporal alignment processing, the preset time reference is the whole hour. For data with different time resolutions, such as wind speed data recorded every 15 minutes, it is integrated according to the whole hour. For example, between 9 o'clock and 10 o'clock, there are 4 wind speed values recorded every 15 minutes [5 m / s, 6 m / s, 7 m / s, 8 m / s], then the wind speed value at 10 o'clock is the average of these 4 values, that is, (5 + 6 + 7 + 8) ÷ 4 = 6.5 m / s. According to the grid division of the geospatial coding data, the meteorological data within each grid is integrated together to form the meteorological grid data after spatio-temporal alignment.

[0018] When extracting the correlation features of meteorological elements, for the temperature-humidity coupling feature, for example, in the temperature range of 15°C - 25°C, the relationship between humidity and temperature is approximately humidity = 2 × temperature - 20. Based on this, combined with the temperature and humidity values in the current meteorological grid data, the temperature-humidity coupling feature is calculated. For the air pressure gradient change feature, calculate the air pressure difference between adjacent grids. For example, the air pressures of adjacent grids G1 and G2 are 1008 hPa and 1005 hPa respectively, and the grid spacing is 1 km, then the air pressure gradient is (1008 - 1005) ÷ 1 = 3 hPa / km. In terms of the wind speed-precipitation synergy feature, it is statistically found that when the wind speed is greater than 8 m / s, the precipitation probability increases significantly. By analyzing the precipitation conditions at different wind speeds in the current meteorological grid data, the wind speed-precipitation synergy feature is obtained.

[0019] Step S122: Perform spatial topological relationship analysis and processing on the geospatial coding data to generate geospatial distribution features including terrain elevation correlation features, urban building distribution features, and water system coverage features.

[0020] In this embodiment, for the terrain elevation correlation feature, the terrain elevation differences and change directions between adjacent grids are analyzed. For example, the terrain elevation of grid G3 is 50 meters, and the terrain elevation of adjacent grid G4 is 48 meters. The terrain elevation correlation between them is expressed as an elevation difference of -2 meters, and the change direction is a decrease from G3 to G4. The urban building distribution feature is determined by detailed statistics of information such as the floor area of buildings and the proportion of building types in each grid. For example, the floor area of buildings in grid G5 is 0.5 square kilometers, among which commercial buildings account for 30% and residential buildings account for 70%. The water system coverage feature records the specific situation of the water system within the grid. For example, there is a river passing through grid G6, the river width is 50 meters, the water flow direction is from west to east, and the water system coverage area accounts for 15% of the grid area.

[0021] Step S123: Perform trend fitting processing on the spatio-temporally aligned meteorological grid data based on a sliding time window to generate meteorological change trend features including the cumulative change rate of meteorological elements, the sudden occurrence probability of meteorological events, and the meteorological state transition characteristics.

[0022] In this embodiment, the sliding time window is set to 6 hours. Taking the temperature element as an example, in the current spatio-temporally aligned meteorological grid data, within the sliding time window from 8 am to 2 pm, the temperature data is [20°C, 21°C, 22°C, 23°C, 24°C, 25°C]. Calculate the cumulative change rate of meteorological elements. The temperature cumulative change rate = (final temperature - initial temperature) ÷ time span = (25 - 20) ÷ 6 ≈ 0.83°C / hour.

[0023] For the sudden occurrence probability of meteorological events, by analyzing historical data and current meteorological grid data, the frequency of occurrence of a certain meteorological event (such as heavy rain) under similar meteorological conditions is statistically analyzed. For example, under the current temperature, humidity, wind speed, etc., combined with historical data, it is found that in the past 100 times of similar meteorological conditions, heavy rain occurred 10 times. Then the sudden occurrence probability of the current meteorological event (heavy rain) is 10 ÷ 100 = 10%.

[0024] In terms of meteorological state transition characteristics, the conversion situations between different meteorological states can be analyzed. For example, the current meteorological state is sunny. Through the analysis of historical data and current meteorological trends, it is found that within the next 6 hours, there is a 30% probability of changing to cloudy weather and a 10% probability of changing to light rain weather, etc., so as to determine the meteorological state transition characteristics.

[0025] Step S124: Combine the meteorological element correlation features, the geographical spatial distribution features, and the meteorological change trend features to generate the meteorological spatio-temporal feature set.

[0026] In this embodiment, the meteorological element correlation features, geographical spatial distribution features, and meteorological change trend features obtained above are integrated. For example, meteorological element correlation features such as temperature-humidity coupling features, air pressure gradient change features, and wind speed-precipitation cooperation features are combined together; geographical spatial distribution features such as terrain elevation correlation features, urban building distribution features, and water system coverage features are combined; meteorological change trend features such as meteorological element cumulative change rate, meteorological event sudden probability, and meteorological state transition features are combined. Then these three groups of features are spliced to form a meteorological spatio-temporal feature set. During the splicing process, those skilled in the art can standardize the dimensions and scales of different features according to the actual situation through existing technologies.

[0027] Step S130: Invoke the trained deep reinforcement learning recognition model to perform dynamic disaster pattern matching processing on the meteorological spatio-temporal feature set, and generate a meteorological disaster recognition result set for the target urban area. The meteorological disaster recognition result set includes disaster type identifiers, disaster impact ranges, and disaster evolution prediction paths.

[0028] Optionally, step 130 includes: Step S131: Input the meteorological element correlation features into the feature encoding layer of the deep reinforcement learning recognition model to generate meteorological element feature vectors.

[0029] In this embodiment, the meteorological element correlation features extracted previously, such as temperature-humidity coupling features, air pressure gradient change features, and wind speed-precipitation cooperation features, are input into the feature encoding layer of the deep reinforcement learning recognition model. The feature encoding layer processes these features. For example, for the temperature-humidity coupling feature, it is digitally encoded. For example, the temperature-humidity coupling feature is represented by a two-dimensional array. The first element of the array represents the linear relationship coefficient between temperature and humidity, and the second element represents the constant term. The feature encoding layer converts this two-dimensional array into a vector with a length of 10 through an existing encoding algorithm. Similar encoding processes are also performed on the air pressure gradient change feature and the wind speed-precipitation cooperation feature. Finally, these encoded feature vectors are spliced together to form meteorological element feature vectors.

[0030] Step S132: Input the geographical spatial distribution features into the spatial relationship analysis layer of the deep reinforcement learning recognition model to generate a spatial topological relationship matrix, and the spatial topological relationship matrix includes the meteorological influence weights between different geographical units.

[0031] In this embodiment, the geospatial distribution features are input into the spatial relationship analysis layer. Taking the terrain elevation correlation feature as an example, the spatial relationship analysis layer analyzes the impact of the terrain elevation difference between different grids on meteorology. For example, the terrain elevation difference between grid A and grid B is large. Through historical data analysis, it is found that this elevation difference will cause the air flow in grid A to flow more easily towards grid B, thus affecting the precipitation distribution. According to this degree of influence, the meteorological influence weight from grid A to grid B is assigned a value of 0.6. For the urban building distribution feature and the water system coverage feature, the impact of them on meteorological elements between different grids is also analyzed. For example, a large number of high-rise buildings in a certain grid will change the local wind speed and direction. According to the degree of influence, the meteorological influence weight between relevant grids is assigned, and finally a spatial topological relationship matrix representing the meteorological influence weight between different geographical units (grids) is formed.

[0032] Step S133: Input the meteorological change trend feature into the time series analysis layer of the deep reinforcement learning recognition model to generate a meteorological state transition probability map.

[0033] In this embodiment, the meteorological state transition feature in the meteorological change trend feature is input into the time series analysis layer. For example, the meteorological state transition feature records the probabilities of the current sunny state changing to other meteorological states (such as cloudy, light rain, etc.) in different time periods in the future. The time series analysis layer generates a meteorological state transition probability map based on this information, combined with historical meteorological data and the current meteorological trend. 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 probabilities of transitioning from one meteorological state to another. For example, the value corresponding to the probability of transitioning from sunny to cloudy in the matrix is 0.4, and the corresponding value for transitioning from sunny to light rain is 0.1, etc.

[0034] Step S134: Perform a joint policy evaluation on the meteorological element feature vector, the spatial topological relationship matrix, and the meteorological state transition probability map through the policy network of the deep reinforcement learning recognition model to generate a dynamic evaluation result including the matching degree of disaster trigger conditions, the similarity of disaster diffusion paths, and the scoring of disaster evolution stages.

[0035] In this embodiment, the policy network of the deep reinforcement learning recognition model comprehensively analyzes the previously generated meteorological element feature vector, spatial topological relationship matrix, and meteorological state transition probability map. The policy network first analyzes the meteorological element feature vector to determine whether the current combination of meteorological elements meets the trigger conditions of a certain disaster. For example, when the temperature-humidity coupling feature, the air pressure gradient change feature, etc. reach a certain threshold, it may trigger a rainstorm disaster. By comparing with a preset disaster trigger condition template, the matching degree of the disaster trigger conditions is calculated. For example, the matching degree between the current meteorological element feature vector and the rainstorm disaster trigger condition template is 0.7.

[0036] Next, the policy network combines the spatial topological relationship matrix and the meteorological state transition probability map to analyze the possible diffusion paths of the disaster. For example, according to the meteorological influence weights and meteorological state transition probabilities between different grids, it predicts which grids the rainwater may diffuse from and to if a rainstorm disaster occurs, and calculates the similarity between the current situation and the preset rainstorm disaster diffusion path template, such as 0.6.

[0037] Meanwhile, the policy network evaluates which evolution stage the disaster is in according to the current meteorological state and change trend, and gives a disaster evolution stage score. For example, it is determined that the current rainstorm disaster is in the initial stage, with a score of 0.5, and then these evaluation results are combined to form a dynamic evaluation result.

[0038] Step S135: Based on the value network of the deep reinforcement learning recognition model, conduct a disaster impact value prediction on the dynamic evaluation result, and generate a meteorological disaster recognition result set including the predicted value of disaster economic loss, the personnel safety risk index, and the infrastructure vulnerability level.

[0039] In this embodiment, the value network of the deep reinforcement learning recognition model conducts a disaster impact value prediction according to the dynamic evaluation result. For the predicted value of disaster economic loss, the value network analyzes by combining historical disaster data and the economic conditions of the current urban area. For example, when the historical disaster data is known and the disaster trigger condition matching degree, diffusion path similarity, and evolution stage score are similar, the average economic loss is 5 million yuan. According to factors such as the economic growth of the current urban area and the economic density of the affected area, through existing calculation methods (such as a linear regression model combined with relevant parameter adjustments), it is predicted that the economic loss of this rainstorm disaster is 6 million yuan.

[0040] For the personnel safety risk index, the value network considers factors such as the population density and building types in the affected area. For example, the population density in the affected area is relatively high, and there are many old buildings. By analyzing the probability of personnel casualties in similar situations in historical data and combining the current meteorological disaster situation, the calculated personnel safety risk index is 0.6 (the higher the index, the greater the risk).

[0041] For the infrastructure vulnerability level, the value network analyzes the infrastructure types in the affected area (such as roads, bridges, hydropower facilities, etc.) and their construction years, maintenance conditions, etc. For example, a certain area has bridges with an early construction year and average maintenance conditions. Combining the impact degree of the current meteorological disaster on the bridges, the infrastructure vulnerability level of this area is evaluated as medium level, and these prediction results are combined to form a meteorological disaster recognition result set.

[0042] Step S140: Generate a set of disaster response strategies based on the set of meteorological disaster identification results. The set of disaster response strategies includes an emergency resource scheduling strategy, a risk area division strategy, and a warning information release strategy.

[0043] Optionally, step S140 includes: Step S141: Match the preset emergency resource type library according to the disaster type identifier, and generate an emergency resource scheduling strategy including the material storage location, transportation route planning, and distribution priority.

[0044] In this embodiment, when the disaster type identifier in the set of meteorological disaster identification results is a rainstorm disaster, the preset emergency resource type library is matched. The emergency resource type library records various required emergency resources for rainstorm disasters, such as flood control sandbags, water pumps, etc. In terms of the material storage location, it is known that flood control sandbags are stored in three warehouses in the city, namely Warehouse A, Warehouse B, and Warehouse C. Warehouse A stores 5,000 flood control sandbags, Warehouse B stores 3,000, and Warehouse C stores 2,000; water pumps are stored in the city's maintenance center, with a total of 10 units. The transportation route planning is formulated according to the geographical spatial information and real-time traffic conditions. For example, if the affected area is mainly in the eastern region of the city, and Warehouse A is closer to the affected area and has good traffic conditions, it is preferred to transport flood control sandbags from Warehouse A. Plan the transportation route from Warehouse A to the affected area, passing through main roads X and Y. The distribution priority is determined according to the severity of the affected area and the urgency of the demand. The most severely affected areas are given priority for material distribution. For example, if the severity score of affected area A is 8 points (out of 10), and the severity score of affected area B is 6 points, then the distribution priority of flood control sandbags in affected area A is higher than that in affected area B. According to the above rules, an emergency resource scheduling strategy is generated.

[0045] Step S142: Overlay the urban population density distribution map based on the disaster impact range, and generate a risk area division strategy including a high-risk area blockade plan, a personnel evacuation route, and the setting of temporary resettlement points.

[0046] In this embodiment, the disaster impact scope in the known meteorological disaster identification result set is such that the impact scope of the rainstorm disaster covers multiple areas of the city. The disaster impact scope is analyzed by overlaying it with the urban population density distribution map. For example, the areas with relatively high population density within the disaster-affected areas are mainly concentrated in several blocks, such as Block M, Block N, and Block O. These blocks are prone to waterlogging during rainstorms due to their low terrain and pose a relatively high risk. For these high-risk areas, a lockdown plan is formulated. For example, traffic control is implemented in Block M, non-emergency vehicles are prohibited from entering, and warning signs are set up. The evacuation routes for personnel are formulated based on geospatial information and building distribution. For example, the residents within Block M evacuate to nearby highlands through Street P and Street Q, and obvious guiding signs are set up along the evacuation routes. The temporary resettlement sites are set up in areas with high terrain, safety, and certain infrastructure. For example, a temporary resettlement site is set up in a park near Block M. This park has enough space to accommodate the evacuated people and is equipped with temporary tents, drinking water supply facilities, etc. In the above manner, a risk area division strategy is generated.

[0047] Step S143: Generate an early warning information release strategy including the early warning information release level, the priority of the early warning coverage area, and the combination of information dissemination channels according to the time series characteristics of the disaster evolution prediction path.

[0048] In this embodiment, according to the time series characteristics of the disaster evolution prediction path in the meteorological disaster identification result set. For example, the disaster evolution prediction path of the rainstorm disaster shows that the disaster will gradually worsen within the next 3 hours. Based on this time series characteristic, the early warning information release level is determined. Since the disaster is gradually worsening, the early warning information release level is upgraded from a yellow warning (relatively severe) to an orange warning (severe). The priority of the early warning coverage area is determined according to the disaster impact scope and population density. The areas with high population density within the disaster-affected areas are given priority for early warning, such as Block M, Block N, and Block O mentioned above. The combination of information dissemination channels is selected in multiple ways to ensure that the information can be conveyed to relevant personnel in a timely manner. For example, orange warning information is released through television and radio, and at the same time, warning text messages are sent to mobile phone users within the disaster-affected areas using the mobile phone text message platform, and warning notices are posted on the community bulletin boards in the disaster-affected areas. In accordance with the above rules, an early warning information release strategy is generated.

[0049] Step S144: Perform strategy conflict detection on the emergency resource scheduling strategy, the risk area division strategy, and the early warning information release strategy, and eliminate the execution time conflict and resource allocation conflict between different strategies to obtain an optimized set of disaster response strategies.

[0050] In this embodiment, conflict detection is performed on the generated emergency resource scheduling strategy, risk area division strategy, and early warning information release strategy. For example, in the emergency resource scheduling strategy, it is planned to transport flood control sandbags from Warehouse A to the disaster area during a certain time period, but in the risk area division strategy, traffic control is implemented in this area during this time period, which results in a conflict in the execution time. By adjusting the transportation time and postponing it until after the traffic control is lifted, the conflict in the execution time is resolved.

[0051] In terms of resource allocation conflicts, the early warning information release strategy requires using some vehicles to post notices in the disaster area, but these vehicles are arranged to transport materials in the emergency resource scheduling strategy. By reallocating vehicle resources and arranging a part of the standby vehicles for posting notices, the resource allocation conflict is resolved. After detecting and adjusting these conflicts, the execution time conflicts and resource allocation conflicts between different strategies are eliminated, and an optimized set of disaster response strategies is obtained. This optimized set of strategies ensures that when dealing with meteorological disasters, various measures can be implemented coordinately and efficiently, minimizing the losses caused by the disasters to the greatest extent.

[0052] Step S145: Encode the optimized set of disaster response strategies into the format of instructions to be executed and push them to the corresponding emergency management terminal devices.

[0053] In this embodiment, the optimized set of disaster response strategies is encoded to conform to the format that can be recognized and executed by the emergency management terminal devices. For example, information such as the material storage location, transportation route planning, and allocation priority in the emergency resource scheduling strategy is converted according to specific encoding rules. For example, the material storage location is represented by a digital code, Warehouse A is represented by "01", Warehouse B is represented by "02", and Warehouse C is represented by "03"; the transportation route is digitally encoded through road numbers and node information; the allocation priority is 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", which means transporting materials from Warehouse A along route 101 - 102 - 103 with an allocation priority of 3, and transporting materials from Warehouse B along route 104 - 105 with an allocation priority of 2.

[0054] For the risk area division strategy, information such as the high-risk area blockade plan, personnel evacuation route, and temporary resettlement point setting is also encoded. For example, the high-risk area block M is represented by "M", the blockade plan is represented by the code "L1"; the personnel evacuation route is encoded through street numbers and direction information, such as "P - N" represents evacuating from street P to the north; the temporary resettlement point park is represented by "P1". These encoded information are combined into a string similar to "M, L1; P - N; P1".

[0055] The early warning information release strategy is also encoded accordingly. The early warning information release level is represented by specific characters. The orange warning is represented by "O". The priority of the warning coverage area is combined according to the area number and the priority number, such as "M, 1; N, 2; O, 3", indicating that the priority of block M is 1, the priority of block N is 2, and the priority of block O is 3. The combination of information dissemination channels is represented by codes. For example, television is represented by "T", radio is represented by "R", text message is represented by "S", and bulletin board is represented by "B", combined into a string of "O, M, 1; N, 2; O, 3, T-R-S-B".

[0056] Combine these strings of the encoded emergency resource scheduling strategy, risk area division strategy, and early warning information release strategy to form a complete optimized set of disaster response strategy codes. Then, through network communication technology, push the encoded instructions to the corresponding emergency management terminal devices, such as the computers in the emergency command center, the handheld terminals of rescue personnel, etc. After receiving the instructions, these terminal devices can automatically decode and execute the corresponding response measures according to the instructions.

[0057] Step S150: Perform dynamic policy optimization processing on the deep reinforcement learning recognition model based on the set of disaster response strategies to obtain an optimized deep reinforcement learning recognition model, and deploy the optimized deep reinforcement learning recognition model to the meteorological disaster monitoring system.

[0058] Optionally, step S150 includes: Step S151: Extract the actual material arrival time data from the emergency resource scheduling strategy, extract the personnel evacuation path execution efficiency data from the risk area division strategy, and generate a set of policy execution effectiveness indicators.

[0059] In this embodiment, during the implementation of the emergency resource scheduling strategy, by installing positioning devices and time recording devices on the material transportation vehicles, the actual material arrival time data is obtained in real time. For example, the flood control sandbags depart from warehouse A, and the planned arrival time at the affected area is 10 am, and the actual arrival time is 10:30 am. Record this time difference of 30 minutes as part of the actual material arrival time data. For different batches and different types of materials, record the corresponding actual arrival time data.

[0060] During the execution of the risk area division strategy, by setting sensors and statistical devices on the evacuation routes of personnel, data on the execution efficiency of the evacuation routes of personnel is obtained. For example, on streets P and Q of the evacuation route in block M, the number of people passing through per unit time is counted. For example, it is planned to evacuate 500 people per hour, but actually 450 people are evacuated per hour. By calculating the ratio of the actual number of evacuated people to the planned number of evacuated people (450÷500 = 0.9), as well as data such as the difference between the evacuation time and the expected evacuation time, data on the execution efficiency of the evacuation routes of personnel is comprehensively generated.

[0061] Combining these actual material arrival time data and the data on the execution efficiency of the evacuation routes of personnel together forms a set of policy execution effectiveness indicators, which reflects the effectiveness and efficiency of the disaster response strategy during the actual execution process.

[0062] Step S152: Input the set of policy execution effectiveness indicators into the policy network backpropagation interface of the deep reinforcement learning recognition model, and calculate the weight gradient deviation matrix of the disaster trigger condition matching degree.

[0063] In this embodiment, the generated set of policy execution effectiveness indicators is input into the policy network backpropagation interface of the deep reinforcement learning recognition model. The policy network backpropagation interface calculates the weight gradient deviation of the disaster trigger condition matching degree according to the input set of indicators and the current parameters of the model. For example, there is a deviation between the disaster trigger condition matching degree predicted by the current model and the actual situation. Actually, a rainstorm disaster should be triggered under a certain meteorological condition, but the matching degree predicted by the model is relatively low. Through the backpropagation algorithm, the gradient deviation of each weight parameter causing this deviation is calculated.

[0064] The specific calculation process is as follows: First, according to the data in the set of policy execution effectiveness indicators, determine the error between the actual disaster occurrence situation and the disaster trigger condition matching degree predicted by the model. For example, a rainstorm disaster actually occurred, but the matching degree predicted by the model is only 0.4, while the ideal matching degree should be close to 1. Then, according to this error, combined with the structure of the model and the current weight parameters, use the backpropagation algorithm to backpropagate the error from the output layer to the input layer, calculate the contribution of each weight parameter to the error, that is, the weight gradient. Finally, organize these weight gradients into a matrix to form a weight gradient deviation matrix. Each element in this matrix represents the gradient deviation of the corresponding weight parameter, reflecting the direction and amplitude of adjustment required for this weight parameter.

[0065] Step S153: Based on the weight gradient deviation matrix, adjust the meteorological influence weight allocation ratio between different geographical units in the spatial topology relationship matrix to generate an optimized spatial topology relationship matrix.

[0066] In this embodiment, the spatial topological relationship matrix is adjusted according to the weight gradient deviation matrix. For example, the weight gradient deviation matrix shows that the deviation contribution of the meteorological influence weight between a certain geographical unit (grid) and its adjacent grids to the matching degree of the disaster triggering condition is relatively large. For example, the current value of this weight is 0.5, and according to the direction and magnitude of the weight gradient deviation, it is decided to adjust it to 0.6.

[0067] For other weights in the spatial topological relationship matrix, they are adjusted in a similar manner. By traversing the weight gradient deviation matrix, according to the situation of each element, the corresponding meteorological influence weights in the spatial topological relationship matrix are adjusted. The adjustment principle is to reduce the weights that cause the deviation of the disaster triggering condition matching degree and increase the weights that help improve the matching degree. After the above adjustment, an optimized spatial topological relationship matrix is generated, and this optimized matrix can more accurately reflect the meteorological influence relationship between different geographical units and can better support the model to identify and predict disasters.

[0068] Step S154: Perform a 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 a correction coefficient for the value network loss function.

[0069] 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 early warning 1 hour before the disaster occurs, but the actual release time is delayed by 30 minutes, that is, the actual response delay data is 30 minutes. Perform a convolution alignment operation on this actual response delay data and the time series of the meteorological state transition probability map.

[0070] The time series of the meteorological state transition probability map records the probability situation of meteorological state transitions at different times. For example, using the existing convolution alignment processing, the actual response delay data is matched and calculated with the time series of the meteorological state transition probability map. For example, the time series of the meteorological state transition probability map records probability values at intervals of every 15 minutes. The 30 minutes of the actual response delay is corresponding to this time series and calculated with the corresponding probability values. Among them, the calculation method can be weighted summation and other methods. By calculating a value through weighted summation, after this value undergoes existing function transformations (such as normalization processing), a correction coefficient for the value network loss function is generated, and 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.

[0071] Step S155: Update the output parameters of the spatial relationship analysis layer using the optimized spatial topological relationship matrix, and at the same time adjust the error calculation weight of the predicted value of the disaster economic loss through the correction coefficient of the value network loss function.

[0072] In this embodiment, the optimized spatial topological relationship matrix is applied to the spatial relationship analysis layer of the deep reinforcement learning recognition model to update its output parameters. For example, the spatial relationship analysis layer originally generated output results based on the old spatial topological relationship matrix. Now, the new optimized matrix is used as the input to adjust the relevant calculation parameters and weights, so that the spatial relationship analysis layer can output results that more accurately reflect the meteorological impact relationship in the geographical space. At the same time, the error calculation weight of the disaster economic loss prediction value is adjusted by using the correction coefficient of the value network loss function. For example, when the value network calculates the error of the disaster economic loss prediction value, the weight of a certain factor (such as the economic growth situation in the disaster-affected area) was originally set to 0.3. According to the correction coefficient of the value network loss function and the analysis of the difference between the actual situation and the model prediction, this weight is adjusted to 0.4. Through the above adjustment, when the value network calculates the error of the disaster economic loss prediction value, it can more reasonably consider various factors and improve the accuracy of model prediction.

[0073] Step S156: Perform joint convergence verification on the updated parameters of the policy network and the value network. When the accuracy rate of the disaster evolution prediction path is increased to a preset threshold, output the optimized deep reinforcement learning recognition model.

[0074] In this embodiment, joint convergence verification is performed on the updated parameters of the policy network and the value network. By continuously inputting new meteorological spatio-temporal feature set data, the updated model is used for meteorological disaster recognition and related predictions. For example, meteorological data for different time periods are input multiple times, and the model is allowed to predict the disaster evolution prediction path. The degree of coincidence between the disaster evolution prediction path predicted by the model and the actual situation is statistically analyzed, and the accuracy rate is calculated. The preset threshold is set to 80%. When, after multiple verifications, the accuracy rate of the disaster evolution prediction path predicted by the model continuously and stably increases to 80% or more, it is considered that the parameter adjustment of the model has achieved a good effect. At this time, the optimized deep reinforcement learning recognition model is output.

[0075] In an alternative embodiment, the method further includes: Step S210: Real-time monitor the data recognition delay index and the disaster omission rate index of the meteorological disaster monitoring system.

[0076] In this embodiment, a monitoring module is set in the meteorological disaster monitoring system to collect and analyze data in real time to identify the data recognition delay index and the disaster omission rate index. For the data recognition delay index, record the time it takes from the meteorological monitoring data entering the system to the deep reinforcement learning recognition model outputting the meteorological disaster recognition result set. For example, at a certain moment, a set of meteorological data is monitored to enter the system at 9:00 am, and the model outputs the recognition result at 9:15 am. Then the data recognition delay time for this time is 15 minutes. Continuously record multiple above-mentioned data recognition delay times and calculate the average value as the current data recognition delay index.

[0077] For the disaster omission rate index, it is statistically compared with the actual occurrence of meteorological disasters. For example, within a certain period of time, 10 meteorological disasters actually occurred, while the meteorological disaster monitoring system only identified 8 times. Then the disaster omission rate = (actual disaster times - identified disaster times) ÷ actual disaster times = (10 - 8) ÷ 10 = 20%. By monitoring these indicators in real time, the performance status of the meteorological disaster monitoring system can be understood in a timely manner.

[0078] Step S220: When the data recognition delay index exceeds the preset response threshold, perform lightweight processing on the deep reinforcement learning recognition model: compress the number of neurons in the feature encoding layer, reduce the hidden layer dimension of the policy network, and quantize the parameter precision of the value network.

[0079] In this embodiment, the preset response threshold is set to 20 minutes. When the monitored data recognition delay index exceeds 20 minutes, start performing lightweight processing on the deep reinforcement learning recognition model. For the compression of the number of neurons in the feature encoding layer, for example, originally there are 100 neurons in the feature encoding layer. According to a certain compression ratio (such as 20%), the number of neurons is reduced to 80. In the process of reducing neurons, by analyzing the contribution of each neuron to the model performance, preferentially retain the neurons with greater contribution and remove the neurons with smaller contribution. For the reduction of the hidden layer dimension of the policy network, for example, the policy network originally has 3 hidden layers, and each hidden layer has 50 neurons. According to the lightweight requirement, the number of hidden layers is reduced to 2, and the number of neurons in each hidden layer is adjusted to 40. By readjusting the network structure and weight parameters, the policy network can still maintain a certain performance when the hidden layer dimension is reduced. For the quantization of the parameter precision of the value network, for example, the parameters of the original value network are represented by 32-bit floating-point numbers, and now they are quantized to 16-bit floating-point numbers. In the quantization process, through existing quantization algorithms, try to reduce the impact of quantization on the model performance and ensure that the model can still accurately predict the disaster impact value when the parameter precision is reduced.

[0080] Step S230: When the disaster underreporting rate index exceeds the preset safety threshold, activate the incremental learning mode, and input the latest obtained meteorological disaster case data into the deep reinforcement learning recognition model for local parameter fine-tuning.

[0081] In this embodiment, the preset safety threshold is set to 15%. When the monitored disaster underreporting rate index exceeds 15%, activate the incremental learning mode. For example, a new meteorological disaster case data is newly obtained, which has some unique combinations of meteorological elements and geospatial characteristics. Input this case data into the deep reinforcement learning recognition model. Then, the model performs local parameter fine-tuning according to this new data. For example, for the special meteorological element correlation characteristics in this new case, adjust the relevant encoding parameters in the feature encoding layer so that the model can better encode and recognize such features. For the spatial relationship analysis layer, according to the characteristics of the geospatial features in the new case, fine-tune the weight parameters in the spatial topology relationship matrix to more accurately reflect the impact of the geospatial on meteorological disasters. In terms of the value network, according to the disaster impact situation in the new case, adjust the calculation weights of relevant parameters such as the predicted value of disaster economic losses, the personnel safety risk index, and the infrastructure vulnerability level, so as to improve the model's recognition ability for new types or special meteorological disasters.

[0082] Step S240: Dynamically adjust the model compression ratio and the incremental learning frequency through the delay-accuracy balance algorithm, so that the generation efficiency and recognition accuracy of the meteorological disaster recognition result set maintain a preset balanced state.

[0083] In this embodiment, the delay-accuracy balance algorithm dynamically adjusts the model compression ratio and the incremental learning frequency according to the data recognition delay index and the disaster underreporting rate index monitored in real time, as well as the performance of the model after lightweight processing and incremental learning. For example, when the data recognition delay index has decreased but is still close to the preset response threshold, and the disaster underreporting rate index has increased but does not exceed the preset safety threshold, the algorithm appropriately reduces the model compression ratio, reduces the compression ratio of the number of neurons in the feature encoding layer and the dimension of the hidden layer of the policy network, and at the same time increases the incremental learning frequency, increases the input of new meteorological disaster case data and the number of parameter fine-tuning times, so as to improve the recognition accuracy while ensuring the generation efficiency. On the contrary, when the data recognition delay index is much lower than the preset response threshold and the disaster underreporting rate index is also low, the algorithm appropriately increases the model compression ratio to further reduce the computational resource consumption of the model, and at the same time reduces the incremental learning frequency to reduce unnecessary parameter fine-tuning, so as to achieve the preset balanced state of the generation efficiency and recognition accuracy. Through this dynamic adjustment mechanism, it is ensured that the meteorological disaster monitoring system can operate efficiently and accurately under different actual situations.

[0084] In yet another alternative embodiment, the method further includes: 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.

[0085] In this embodiment, the historical disaster case library collects various meteorological disaster cases that occurred in the urban area over the past many years. Each case records the meteorological monitoring data at the time of the disaster, including multi-dimensional meteorological time-series data and geospatial coding data, as well as corresponding information such as disaster type, affected area, and evolution path. For example, a heavy rain disaster case is recorded. At that time, the temperature time-series data was [25°C, 24°C, 23°C,...], the humidity time-series data was [70%, 72%, 75,...], the geospatial coding data showed that the affected area was mainly in the low-lying areas of the city, the disaster type was heavy rain, the affected area covered several blocks, and the disaster evolution path was spreading gradually from a certain area to the surrounding areas.

[0086] The set of simulated disaster scenarios is a variety of possible meteorological disaster scenarios generated by computer simulation. Using meteorological models and geographic information systems, different meteorological parameters and geospatial conditions are input to generate multiple simulated disaster scenarios. For example, to simulate a hurricane disaster scenario, initial meteorological parameters such as wind speed, wind direction, and air pressure, as well as geospatial parameters such as the terrain and building distribution in the urban area are set, and the formation, development, and impact process of the hurricane in this area are simulated, and the meteorological data and disaster-related information during the simulation are recorded.

[0087] The real-time monitoring and comparison data is the comparison information between the actual meteorological data collected in real time and the model prediction results during the operation of the meteorological disaster monitoring system. For example, when the current meteorological data is monitored in real time and the model predicts that a certain meteorological disaster may occur, the difference data between the actual meteorological situation and the disaster type, affected area, evolution path, etc. predicted by the model is recorded. Combining these historical disaster case libraries, sets of simulated disaster scenarios, and real-time monitoring and comparison data forms a meteorological disaster identification and verification matrix.

[0088] Step S320: Perform multi-dimensional comparison of the meteorological disaster identification result set output by the deep reinforcement learning identification model with the meteorological disaster identification and verification matrix. The multi-dimensional comparison includes: verification of disaster type matching degree, verification of affected area coincidence degree, and verification of evolution path similarity degree.

[0089] In this embodiment, when verifying the matching degree of disaster types, the disaster type identifiers output by the deep reinforcement learning recognition model are 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 "heavy rain", search for all cases marked as "heavy rain" in the historical disaster case library and the simulated disaster scenario set, and calculate the ratio of the number of times the model correctly identifies "heavy rain" to the total number of comparisons as the matching degree of the disaster type. For example, if a total of 50 comparisons are made and the model correctly identifies "heavy rain" 40 times, then the matching degree of the disaster type = 40 ÷ 50 = 0.8.

[0090] For the verification of the overlap degree of the affected range, compare the disaster affected range output by the model with the data in the verification matrix. For example, if the model predicts that the affected range of a certain heavy rain disaster covers 5 blocks, find the corresponding cases in the historical disaster case library and the simulated disaster scenario set, and calculate the ratio of the overlapping area between the predicted affected range by the model and the actual or simulated affected range to the total area of the two. For example, if the overlapping area is 3 blocks and the total area of the two is 8 blocks, then the overlap degree of the affected range = 3 ÷ 8 = 0.375.

[0091] In terms of the verification of the similarity of the evolution path, compare the predicted disaster evolution path output by the model with the data in the verification matrix. For the cases in the historical disaster case library, obtain the detailed records of their actual disaster evolution processes, including information such as the development locations and intensity changes of the disaster at different time points. For the simulated disaster scenario set, there are also simulated disaster evolution path data.

[0092] Parse the predicted disaster evolution path output by the model in detail according to the time series and spatial location information, and compare it with the evolution paths in the historical cases and simulated scenarios point by point and period by period. For example, compare the path predicted by the model that the disaster spreads from area A to area B within the next 1 hour with the spread path of a similar disaster in the historical case within the same time period. By calculating multiple factors such as the coincidence degree of key position points on the path, the consistency of the path direction, and the matching degree of the disaster development speed in different time periods, a similarity value is comprehensively obtained.

[0093] Through the verification and comparison of these three dimensions (disaster type matching degree, affected range overlap degree, evolution path similarity), the accuracy and reliability of the output results of the deep reinforcement learning recognition model can be comprehensively and accurately evaluated.

[0094] Step S330: When there are abnormal recognition results exceeding the preset error threshold in the comparison results, trigger the model interpretability analysis module to generate a visual analysis report including a feature weight distribution diagram, a trace of the policy decision path, and the basis for value evaluation.

[0095] In this embodiment, the preset error threshold is set according to the actual application requirements and the requirements for model accuracy. For example, for the disaster type matching degree, the preset error threshold is set to 0.7; for the overlap degree of the influence range, the preset error threshold is set to 0.4; for the similarity of the evolution path, the preset error threshold is set to 0.5.

[0096] 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 recognition result, and at this time, the model interpretability analysis module is triggered.

[0097] The generation process of the feature weight distribution diagram is as follows: The model interpretability analysis module deeply analyzes the weight contributions of each feature (such as temperature-humidity coupling features, air pressure gradient change features, terrain elevation correlation features, etc.) to the final disaster recognition result when the deep reinforcement learning recognition model processes the meteorological spatio-temporal feature set. The weight value of each feature can be determined based on existing algorithms or calculation methods and presented in a graphical way. For example, in the form of a bar chart, the abscissa represents different feature names, and the ordinate represents the feature weight values, intuitively showing which features play a key role in model decision-making and which features have relatively low weights.

[0098] Regarding the trace of the policy decision path, the module starts from the input layer of the model and analyzes the processing process and decision basis of each layer (such as the feature encoding layer, the spatial relationship parsing layer, the time series analysis layer, the policy network, the value network, etc.) along the data propagation path in the model. Record how the data is transformed, how the features are extracted and combined in each layer, and how the final decision is generated based on these processing results. For example, in the policy network, analyze how the joint policy evaluation is carried out according to the meteorological element feature vector, the spatial topological relationship matrix, and the meteorological state transition probability map, and specifically what calculation methods and rules are used to determine the disaster trigger condition matching degree, the disaster diffusion path similarity, and the disaster evolution stage score. Through this trace, the complete process of the model making a decision is clearly shown.

[0099] The generation of the basis for value assessment is based on various factors and calculation methods used by the module to analyze in detail the value network when generating the predicted values of disaster economic losses, the personnel safety risk index, and the infrastructure vulnerability level in the set of meteorological disaster identification results. For example, it shows how the value network combines historical data, current meteorological conditions, geospatial information, and the parameter settings of the model itself to calculate the predicted value of disaster economic losses. 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 through specific algorithms and model structures. For the assessment of the infrastructure vulnerability level, it elaborates on which attribute information of the infrastructure (such as construction age, material, maintenance status, etc.) is relied on and the impact mechanism of meteorological disasters on these infrastructures.

[0100] Integrate these characteristic weight distribution maps, strategy decision path tracing, and the basis for value assessment in a visual manner in a report to form a visual analysis report that combines intuitive and easy-to-understand graphics, charts, and text descriptions.

[0101] Step S340: Locate the root feature dimension of the model recognition deviation according to the visual analysis report, and perform targeted optimization on the preprocessing process of the meteorological spatio-temporal feature set based on the root feature dimension.

[0102] In this embodiment, by carefully studying the visual analysis report, the root feature dimension of the model recognition deviation is located. For example, it is found from the characteristic weight distribution map that the weight of the temperature-humidity coupling feature does not match the expectation in the model decision-making, resulting in a deviation in the matching of disaster types. Further analyzing the strategy decision path tracing information, it is found that there is a problem with the encoding method of the temperature-humidity coupling feature in the feature encoding layer, so that this feature is not correctly expressed and utilized in the subsequent model processing, which determines that the temperature-humidity coupling feature dimension is one of the roots of the model recognition deviation.

[0103] Based on this root feature dimension, perform targeted optimization on the preprocessing process of the meteorological spatio-temporal feature set. For the extraction process of the temperature-humidity coupling feature, re-examine the algorithms and parameter settings used. For example, originally, a simple linear relationship model was used to calculate the temperature-humidity coupling feature, but through analysis, it is found that the actual temperature-humidity relationship is more complex and may have a non-linear relationship. Therefore, adjust the calculation method and introduce a more complex non-linear model, such as a polynomial regression model or a neural network model, to fit the relationship between temperature and humidity.

[0104] During the data collection and processing phase, increase the sampling frequency and accuracy of temperature and humidity data. For example, originally, temperature and humidity data were recorded every half hour, but now it is increased to once 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 carried out on the collected data to eliminate outliers and incorrect data.

[0105] In terms of feature encoding, for the new temperature-humidity coupling feature calculation method, design a more suitable encoding scheme. For example, adopt an encoding method that can better retain feature information, such as encoding the temperature-humidity coupling feature as a multi-dimensional vector, where each dimension of the vector represents different feature attributes or feature relationships, enabling the model to more accurately understand and utilize this feature in subsequent processing.

[0106] Through the above targeted optimization of the preprocessing process of the meteorological spatio-temporal feature set based on the root feature dimension, the quality and accuracy of the features can be improved, thereby enhancing the recognition performance of the deep reinforcement learning recognition model, reducing the model recognition deviation, and improving the overall accuracy and reliability of the meteorological disaster monitoring system.

[0107] Optionally, the method further includes: Step S410: Deploy a distributed model update architecture including edge computing nodes and a cloud coordination center in the meteorological disaster monitoring system.

[0108] In this embodiment, the urban area covered by the meteorological disaster monitoring system is relatively large. To improve the processing efficiency of the system and meet the meteorological data processing requirements of different regions, a distributed model update architecture is deployed. Edge computing nodes are distributed in various regions of the city, close to meteorological monitoring stations. These edge computing nodes have certain computing power and storage capacity. For example, edge computing nodes are set up in the east, west, south, and north areas of the city respectively. Each edge computing node is equipped with a high-performance processor and a certain amount of memory and storage devices to meet the needs of real-time processing of meteorological data. The cloud coordination center is located in the data center and has powerful computing and storage resources. It is responsible for receiving the data uploaded by each edge computing node, conducting overall coordination and management, and updating and optimizing the model. The cloud coordination center is connected to each edge computing node through a high-speed network to ensure fast data transmission and interaction. The edge computing nodes and the cloud coordination center cooperate with each other to form an organic whole. The edge computing nodes are responsible for real-time processing of local meteorological data, reducing the burden on the cloud coordination center; the cloud coordination center analyzes and processes global data, updates the model, and timely sends the updated model parameters to the edge computing nodes to ensure the consistency and accuracy of the entire system.

[0109] Step S420: Deploy a lightweight recognition model at the edge computing node to process the meteorological spatio-temporal feature set in the area in real time to generate a preliminary recognition result.

[0110] In this embodiment, a lightweight recognition model is deployed on each edge computing node. These lightweight recognition models are simplified and optimized versions of the deep reinforcement learning recognition model, aiming to be able to run quickly on edge computing nodes with limited resources. The lightweight recognition model also receives the meteorological spatio-temporal feature set in the area as input. For example, the edge computing node located in the east district of the city obtains the meteorological monitoring data in this area in real time and generates a meteorological spatio-temporal feature set after preprocessing. The lightweight recognition model processes these feature sets and uses modules such as its internal feature encoding layer, spatial relationship analysis layer, and time series analysis layer for analysis and calculation. In the feature encoding layer, a simplified encoding algorithm is used to encode the meteorological element correlation features. For example, for the temperature-humidity coupling feature, a faster linear encoding method is used to encode it into a shorter vector. The spatial relationship analysis layer then uses the pre-computed simplified spatial topology relationship matrix to quickly analyze the impact of the geographical space on the meteorology. The time series analysis layer generates a preliminary estimate of the meteorological state transition probability through simple statistical methods and models. After the above processing, the lightweight recognition model generates preliminary recognition results, which include a preliminary judgment on the possible types of meteorological disasters and a rough estimate of the disaster impact range. For example, the preliminary recognition result may show that there is a high probability of a rainstorm disaster in the east district in the next few hours, and the affected area may cover several blocks. The edge computing node stores these preliminary recognition results in real time and uploads them to the cloud coordination center when necessary.

[0111] Step S430: Deploy a full-scale recognition model at the cloud coordination center to receive the abnormal feature data and high-risk recognition results uploaded by each edge computing node for review and analysis.

[0112] In this embodiment, the full-scale recognition model deployed by the cloud coordination center is a complete, complex, and highly accurate deep reinforcement learning recognition model. It has more powerful computing capabilities and richer parameters, enabling more in-depth and accurate analysis of meteorological data. After each edge computing node processes the meteorological spatio-temporal feature set to generate a preliminary recognition result, it will evaluate the result. If it is found that some feature data is abnormal, or the preliminary recognition result shows a high-risk situation (such as the possible occurrence of a severe meteorological disaster), then these abnormal feature data and high-risk recognition results will be uploaded to the cloud coordination center. For example, when processing meteorological data, the edge computing node in the west area of the city found that a set of meteorological element correlation features differed significantly from previous data and initially judged that it might be a new type of meteorological disaster pattern, so it uploaded this set of abnormal feature data to the cloud coordination center. At the same time, if the edge computing node initially identifies that a hurricane disaster may occur in a certain area, which belongs to a high-risk recognition result, it will also be promptly uploaded to the cloud coordination center.

[0113] After receiving these uploaded data, the full-scale recognition model of the cloud coordination center conducts a review and analysis on them. The full-scale recognition model uses its more comprehensive feature encoding method, more accurate spatial relationship analysis, and more complex time series analysis method to re-evaluate and analyze the abnormal feature data and high-risk recognition results. For example, for the uploaded abnormal meteorological element correlation features, the full-scale recognition model will analyze them from multiple perspectives, considering the more complex mutual relationships between different meteorological elements, as well as the influence of geographical space and time series factors on these features, to determine whether these abnormal features really represent a new type of meteorological disaster pattern or are caused by data errors. For high-risk recognition results, the full-scale recognition model will further accurately calculate the probability of the disaster occurring, the affected range, and the evolution path, providing a more accurate basis for subsequent decision-making.

[0114] Step S440: Establish an edge-cloud model parameter synchronization mechanism. When the cloud full-scale recognition model detects a new type of disaster pattern, the updated policy network parameters will be encrypted and transmitted to the corresponding edge computing node.

[0115] In this embodiment, in order to ensure that the models of the edge computing nodes and the cloud coordination center can work together and respond in a timely manner to newly emerging meteorological disaster situations, an edge-cloud model parameter synchronization mechanism is established. During the process of the cloud full-scale recognition model rechecking and analyzing the data uploaded by each edge computing node, if a new disaster mode is detected, it will adjust and update its own model parameters. For example, when a new combination mode of meteorological elements is found to be associated with a specific disaster type, the cloud full-scale recognition model will adjust the weight parameters of the corresponding features in the policy network to improve the recognition ability of this new disaster mode. The updated policy network parameters will be encrypted. Existing encryption algorithms, such as the AES encryption algorithm, are used to encrypt the parameters to ensure security and confidentiality during transmission. The encrypted parameters are transmitted to the corresponding edge computing nodes through a high-speed network. For example, when the cloud full-scale recognition model detects a new disaster mode in the southern area of the city, it encrypts the updated policy network parameters and sends them to the edge computing node located in the southern area through the network.

[0116] After receiving the encrypted parameters, the edge computing node first performs a decryption operation, decrypting the parameters using a pre-shared key. After successful decryption, the updated policy network parameters are applied to the local lightweight recognition model. The edge computing node will replace the old parameters with the new parameters according to a certain update process to ensure that the lightweight recognition model can timely acquire the ability to recognize new disaster modes. Through this edge-cloud model parameter synchronization mechanism, the entire meteorological disaster monitoring system can quickly adapt to new meteorological disaster situations and improve the overall performance and adaptability of the system.

[0117] Step S450: Optimize the spatial coverage influence weight and the new disaster adaptation influence weight of the meteorological disaster monitoring system through the coordinated operation of the local response of the edge computing node and the global policy optimization of the cloud.

[0118] In this embodiment, the local response of the edge computing node and the global policy optimization of the cloud cooperate with each other to jointly optimize the relevant weights of the meteorological disaster monitoring system. During the process of the edge computing node real-time processing the meteorological spatio-temporal feature set in its area and generating a preliminary recognition result, it will adjust the spatial coverage influence weight according to the local actual situation. For example, in the area where an edge computing node is located, due to the special geographical environment (such as topographical factors like mountains and rivers), the influence of certain meteorological elements on disasters is more significant. The edge computing node will appropriately increase the spatial coverage influence weight corresponding to these key meteorological elements according to these local features. For example, the edge computing node in the mountainous area finds that the terrain elevation has a great influence on precipitation, and when calculating the spatial coverage influence weight, it will increase the weight value of the meteorological elements related to the terrain elevation, so that the model can more accurately identify and predict meteorological disasters in this area.

[0119] Cloud - side global policy optimization considers the performance and adaptability of the system from an overall perspective. When the cloud - side full - scale recognition model detects a new disaster pattern and updates the policy network parameters, it will adjust the adaptation impact weights for the new disaster. For example, for a newly discovered meteorological disaster pattern, the cloud - side full - scale recognition model will analyze the occurrence probability and impact degree of this disaster in different geographical regions, and adjust the adaptation impact weights of different regions and different meteorological elements for the new disaster according to the analysis results. For regions prone to this new disaster, increase the weights of relevant meteorological elements and geographical spatial features in this region so that the model can better identify and respond to this new disaster.

[0120] Through the coordinated operation of the local response of edge - computing nodes and the cloud - side global policy optimization, continuously adjust and optimize the spatial coverage impact weights and new - disaster adaptation impact weights of the meteorological disaster monitoring system. This collaborative optimization mechanism enables the system to maintain a high recognition accuracy and adaptability in different geographical regions and in the face of various meteorological disaster situations, improving the overall efficiency of the meteorological disaster monitoring system.

[0121] Optionally, the method further includes: Step S510: Match the disaster type identifiers in the meteorological disaster recognition result set with the cause - chain nodes in the historical disaster case library to generate a disaster - cause association vector.

[0122] In this embodiment, the meteorological disaster recognition result set contains information such as the type identifiers of the current meteorological disaster. The historical disaster case library stores a large number of past meteorological disaster cases, and each case details the cause - chain node information of the disaster. For the disaster type identifier in the meteorological disaster recognition result set, such as "rainstorm disaster", search the historical disaster case library for all cases marked as "rainstorm disaster". For each matching case, analyze its cause - chain nodes. The cause - chain nodes may include factors in meteorological elements (such as abnormal atmospheric circulation, water vapor transportation, etc.), factors in geographical space (such as topography, urban layout, etc.), and other relevant factors.

[0123] Extract and organize the characteristics of these cause - chain nodes. For example, for a rainstorm disaster case, the characteristics of its cause - chain nodes may include: a specific atmospheric circulation pattern causes a large amount of water vapor to accumulate in this region, and the mountain terrain around the city blocks the diffusion of water vapor, resulting in concentrated precipitation in this region, etc. Digitalize these characteristics and convert them into vector form. For example, use a 10 - dimensional vector to represent these characteristics, and each dimension of the vector corresponds to different cause - chain node characteristics respectively. For example, the first dimension represents a certain characteristic value of the atmospheric circulation pattern, the second dimension represents the characteristic value of the water vapor transportation volume, the third dimension represents the characteristic value of the terrain altitude, etc.

[0124] By processing the cause-chain node features of multiple matching historical disaster cases and comprehensively calculating these vectors (such as calculating the average value or weighted average value), a disaster cause correlation vector is generated. This vector comprehensively reflects the historical disaster cause characteristics related to the current disaster type and provides an important basis for subsequent analysis and decision-making.

[0125] Step S520: Input the disaster cause correlation vector into the graph reasoning module of the pre-constructed meteorological disaster knowledge graph to trigger the case strategy nodes associated with the geospatial distribution characteristics.

[0126] In this embodiment, the pre-constructed meteorological disaster knowledge graph contains rich meteorological disaster-related knowledge and information, as well as the association relationships between these information. For example, the generated disaster cause correlation vector is input into the graph reasoning module of the meteorological disaster knowledge graph. The graph reasoning module searches and matches in the knowledge graph according to the disaster cause characteristics represented by the vector. For example, if the disaster cause correlation vector contains specific atmospheric circulation patterns and topographic and geomorphic feature information, the graph reasoning module will search for the nodes and relationships associated with these features in the knowledge graph.

[0127] In the knowledge graph, the case strategy nodes associated with the geospatial distribution characteristics store the response strategies and case information for meteorological disasters under different geospatial conditions. When the graph reasoning module finds the matching nodes and relationships according to the disaster cause correlation vector, these case strategy nodes associated with the geospatial distribution characteristics will be triggered. For example, if it is found that the current disaster cause is similar to a certain type of historical rainstorm disaster, and there are corresponding response strategies and case records for this type of rainstorm disaster under a specific geospatial distribution (such as low-lying areas in the city), these relevant case strategy nodes will be triggered to obtain relevant information and strategies.

[0128] Step S530: Extract the response strategy templates in the case strategy nodes that meet the meteorological change trend characteristics to generate a derivative strategy set including cross-case resource scheduling rules, multi-disaster chain priority sorting, and new disaster analogy thresholds.

[0129] In this embodiment, the triggered case strategy nodes contain multiple response strategy templates, which are formulated for different meteorological conditions and geospatial situations. Specifically, according to the current meteorological change trend characteristics, the response strategy templates that match them are screened out from the case strategy nodes. For example, if the current meteorological change trend shows that precipitation will continue to increase and the wind force will gradually increase, search for the response strategy templates for this meteorological change trend in the case strategy nodes.

[0130] For the selected response strategy templates, further organize and extract them. Regarding the cross-case resource scheduling rules, analyze the methods and principles of resource scheduling in different cases when dealing with similar meteorological disasters. For example, in multiple heavy rain disaster cases, summarize the rule that when the disaster impact scope is large, prioritize the allocation of flood control materials from warehouses that are closer to the disaster-stricken area and have sufficient material reserves.

[0131] Regarding the priority ranking of multi-disaster chains, considering that in actual situations, multiple meteorological disasters may be faced simultaneously or a disaster may trigger a chain reaction of other disasters, analyze the methods for determining the priority of different disaster chains in the cases. For example, when heavy rain may trigger floods and landslides, based on factors such as the population density in the disaster-stricken area and the importance of infrastructure, determine the priority of dealing with flood disasters first to ensure the safety of people's lives, and then the priority ranking principle for dealing with landslide disasters.

[0132] The determination of the new disaster analogy threshold is to set a standard by comparing the similarity between the current meteorological disaster and historical cases. For example, calculate the similarity scores of the current disaster in terms of meteorological elements, geographical spatial characteristics, disaster evolution trends, etc. compared with historical cases. When the similarity score reaches 80% (the set threshold) or more, it is considered that the current disaster has a high similarity with the historical case, and the response strategies of the historical case can be used for reference.

[0133] Organize and combine these cross-case resource scheduling rules, multi-disaster chain priority rankings, and new disaster analogy thresholds, etc. to generate a set of derivative strategies. This set of derivative strategies combines the experience and wisdom of historical cases and can provide more comprehensive and targeted strategy guidance for the response to current meteorological disasters.

[0134] Step S540: Fuse the set of derivative strategies and the emergency resource scheduling strategy in the set of disaster response strategies 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.

[0135] In this embodiment, deeply fuse the set of derivative strategies and the emergency resource scheduling strategy in the set of disaster response strategies. First, conduct a detailed analysis of each instruction and parameter in the emergency resource scheduling strategy, such as the location of material reserves, transportation route planning, allocation priorities, etc.

[0136] For the dynamic priority adjustment field, according to the multi-disaster chain priority sorting principle in the derived policy set, adjust the allocation priority in the emergency resource scheduling policy. For example, if the derived policy set indicates that under the current meteorological conditions, a certain disaster-stricken area requires a higher resource allocation priority due to the possible occurrence of more severe secondary disasters, then in the emergency resource scheduling policy, the allocation priority of this area is increased from the original level 3 to level 1 (for example, the priorities are divided into levels 1-5, with level 1 being the highest). At the same time, continuously monitor and dynamically adjust the priorities of each disaster-stricken area according to the real-time meteorological change trend and the development of disasters. For example, if it is found that the disaster situation in another area has worsened and may affect more important infrastructure and the safety of a large number of people, promptly increase the priority of this area.

[0137] In terms of the resource conflict resolution protocol, combine the cross-case resource scheduling rules in the derived policy set to formulate detailed resource conflict resolution methods. For example, when multiple disaster-stricken areas simultaneously request the same type of emergency resource (such as flood control sandbags), according to the cross-case resource scheduling rules, give priority to meeting the needs of the area closer to the disaster source and with a higher degree of disaster severity. If there are route conflicts during the transportation of different types of emergency resources, determine the priority transportation route of the resources according to the urgency and importance of the resources. For example, the transportation of medical rescue supplies takes precedence over general flood control supplies to ensure the safety of the affected people.

[0138] Through the above instruction-level integration, incorporate the effective information and rules in the derived policy set into the emergency resource scheduling policy to generate an enhanced emergency resource scheduling policy. This policy not only considers the basic situation of the current disaster but also combines the experience and response strategies of historical cases, and can more flexibly and effectively respond to complex and changeable meteorological disaster situations, improving the efficiency and rationality of emergency resource scheduling.

[0139] Step S550: Update the infrastructure vulnerability level evaluation parameters in the value network of the deep reinforcement learning recognition model using the instruction execution log of the enhanced emergency resource scheduling policy.

[0140] In this embodiment, the enhanced emergency resource scheduling policy will generate detailed instruction execution logs during the execution process. These logs record the execution situation of each scheduling instruction, including information such as the allocation time of materials, the arrival location, and the actual usage.

[0141] Analyze these instruction execution logs to extract information related to infrastructure vulnerabilities. For example, recording the deployment of a large number of power equipment in a disaster-stricken area for power restoration indicates that the power infrastructure in this area has been greatly affected during this meteorological disaster and may have vulnerability issues. By counting the demand and damage of different types of infrastructure (such as power, communication, transportation, etc.) during the emergency resource scheduling process, evaluate their vulnerability levels.

[0142] According to these evaluation results, update the infrastructure vulnerability level evaluation parameters in the value network of the deep reinforcement learning recognition model. For example, if it is found that a large amount of emergency resources are required for the repair and guarantee of the communication infrastructure in a certain area during multiple meteorological disaster responses, it indicates that the vulnerability of the communication infrastructure in this area is relatively high, and the evaluation parameters of the vulnerability level of the communication infrastructure in the corresponding area in the value network are adjusted upward. On the contrary, if an infrastructure shows strong stability during disaster response and requires fewer emergency resources, its vulnerability level evaluation parameters are appropriately adjusted downward.

[0143] By continuously updating the infrastructure vulnerability level evaluation parameters in the value network with the instruction execution logs of the enhanced emergency resource scheduling strategy, the deep reinforcement learning recognition model can more accurately reflect the vulnerability situation of the infrastructure in reality under meteorological disasters. Thus, in the process of future meteorological disaster recognition and response strategy generation, infrastructure factors can be more reasonably considered, improving the accuracy and practicality of the model and providing support for more effective meteorological disaster response.

[0144] In a non-limiting embodiment, the method further includes: Step S610: Input the geospatial coding data into the voxelization modeling module of the three-dimensional visualization sand table module to generate a three-dimensional base model including terrain elevation gradient and building outlines, and map the wind speed-precipitation collaborative feature in the meteorological spatio-temporal feature set to the atmospheric particle simulator of the three-dimensional base model to generate a three-dimensional dynamic scene with hydrodynamic attributes.

[0145] In this embodiment, the geospatial coding data is input into the voxelization modeling module of the three-dimensional visualization sand table module. The geospatial coding data contains detailed information such as the terrain elevation and urban building distribution of each grid within the urban area. The voxelization modeling module processes these data and divides the entire urban area into tiny voxels (similar to pixels in three-dimensional space).

[0146] For each voxel, its height value is determined according to the terrain elevation information in the geospatial encoding data, so as to construct a three-dimensional model of the terrain, reflecting the terrain elevation gradient. For example, if the grid terrain elevation corresponding to a voxel is 50 meters, the voxel is set at the corresponding height position in the three-dimensional model. At the same time, the outline of the building in the voxel is determined according to the urban building distribution information. If a voxel is located in the building area, the voxel is divided and processed accordingly according to the shape and boundary information of the building to form the three-dimensional outline of the building. By processing all voxels, a three-dimensional base model containing the terrain elevation gradient and the building outline is generated.

[0147] Next, the wind speed-precipitation synergistic features in the meteorological spatiotemporal feature set are mapped to the atmospheric particle simulator of the three-dimensional base model. The wind speed-precipitation synergistic features contain the correlation information of wind speed and precipitation at different locations and times. The atmospheric particle simulator simulates the movement and interaction of particles in the atmosphere. The wind speed information is mapped to the movement speed and direction of atmospheric particles. For example, in an area with a wind speed of 5m / s, the atmospheric particles are set to move at the corresponding speed and direction in the area. The 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 particles is set.

[0148] Through this mapping, the atmospheric particle simulator simulates a three-dimensional dynamic scene with fluid dynamics properties based on the three-dimensional base model. In this scene, you can intuitively see the interaction between wind speed and precipitation under different terrains and building contours, such as the change in wind speed and precipitation distribution near mountains, the disturbance of airflow and the convergence of precipitation in densely built areas in cities, and other phenomena.

[0149] Step S620: superimpose the disaster impact range boundary coordinates of the meteorological disaster identification result set in the three-dimensional dynamic scene, generate a multi-layer disaster impact voxel model with a 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, generate an interactive evacuation simulation scene including a path availability identifier and a congestion risk heat map, and update the urban building distribution feature weights in the spatial relationship analysis layer of the deep reinforcement learning recognition model in real time based on the path parameters modified in the interactive evacuation simulation scene.

[0150] In this embodiment, on the basis of the generated three-dimensional dynamic scene with fluid dynamic properties, the boundary coordinates of the disaster impact range of the meteorological disaster identification result set are superimposed. The meteorological disaster identification result set clarifies the boundaries of the areas that may be affected by the disaster, and these boundary coordinates are marked and displayed in the three-dimensional dynamic scene. For example, for a rainstorm disaster, the boundary coordinates of its impact range are drawn in the three-dimensional scene to form a closed area, representing the area that may be affected by the rainstorm. As time goes on, according to the evolution prediction path of the disaster, the spread of the disaster impact range is dynamically displayed, and a multi-layer disaster impact voxel model with a diffusion path growth animation is generated. This model not only shows the impact range of the current disaster, but also intuitively presents the development trend of the disaster.

[0151] Then, the evacuation path coordinates of the risk area division strategy are imported into the collision detection module of the multi-layer disaster impact voxel model. The evacuation path coordinates determine the route for the affected people to evacuate from the dangerous area to the safe area. The collision detection module detects the evacuation path with the disaster impact area, buildings and other objects. For example, it detects whether the evacuation path passes through an area that is flooding or collides with a building.

[0152] According to the results of the collision detection, a path availability mark is generated. If the evacuation path does not conflict with the dangerous area or building, the path availability mark is green, indicating that the path is available; if the path is partially or completely located in the dangerous area or conflicts with obstacles such as buildings, the path availability mark is red, indicating that the path is unavailable. At the same time, by analyzing factors such as the flow and speed of people on the evacuation path, a congestion risk heat map is generated. For example, in areas with dense populations and slow evacuation speeds, red or orange is used to indicate high congestion risks; in areas with smoother personnel flow, green or light blue is used to indicate low congestion risks. In this way, an interactive evacuation simulation scenario containing path availability marks and congestion risk heat maps is generated.

[0153] In the interactive evacuation simulation scenario, operators can modify the path parameters according to the actual situation, such as adjusting the evacuation route, changing the evacuation starting point and end point, etc. When the path parameters are modified, the system will feed back these changes 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 re-evaluated. For example, if the new evacuation path passes through more high-rise building areas, it means that these buildings have a greater impact on the evacuation, and the weights of the urban building distribution features related to these buildings in the spatial relationship analysis layer of the deep reinforcement learning recognition model are correspondingly increased. In this way, the deep reinforcement learning recognition model can adjust parameters in real time according to the actual evacuation simulation situation, improve the understanding and prediction capabilities of meteorological disasters and personnel evacuation in the urban environment, and further optimize the meteorological disaster response strategy.

[0154] In a non-limiting embodiment, the method further includes: 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 degree, transportation speed volatility, and node residence duration, and correct the road passage 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 a dynamic road network influence factor; In this embodiment, in order to accurately evaluate the execution effect of the emergency resource scheduling strategy and optimize the deep reinforcement learning recognition model, the actual path trajectory data of the material transportation terminal is collected in real time. By installing high-precision positioning systems and sensors on terminal devices such as material transportation vehicles and ships, the position information and operation status data of the devices can be obtained in real time.

[0155] For the collected actual path trajectory data, analysis and processing are performed to generate a set of scheduling efficiency indicators. The calculation of the route deviation degree is by comparing the actual transportation path with the pre-planned route. For example, the planned transportation route is from warehouse A through road X and road Y to the disaster area Z, and the actual transportation path deviates from road Y. By calculating the geometric deviation degree between the actual path and the planned path, the route deviation degree is obtained. For example, if the planned path length is 10 kilometers and the length of the part where the actual path deviates from the planned path is 2 kilometers, then the route deviation degree = 2÷10 = 0.2. The transportation speed volatility is to statistically analyze the speed changes during the transportation process. Record the speeds of the transportation terminal at different time periods and calculate the average speed and standard deviation of the speeds. For example, the speeds in different sections during the transportation process are 30 km / h, 40 km / h, and 35 km / h respectively, and the average speed is calculated to be 35 km / h. Through a certain algorithm, the standard deviation of the speed is calculated to measure the transportation speed volatility. The node residence duration is to statistically analyze the residence time of the transportation terminal at each intermediate node (such as warehouses, transfer stations, etc.). For example, the transportation vehicle stays at warehouse A for 30 minutes for loading and stays at transfer station B for 15 minutes for cargo allocation, and these residence times are recorded as part of the node residence duration. These indicators such as route deviation degree, transportation speed volatility, and node residence duration are combined into a set of scheduling efficiency indicators. According to the indicators in this set, the road passage weight coefficient in the spatial topology relationship matrix of the deep reinforcement learning recognition model is corrected. For example, if the route deviation degree of a certain road is relatively high, it indicates that there may be difficulties or other problems in passing through this road, and the passage weight coefficient of this road in the spatial topology relationship matrix is reduced; if the transportation speed volatility of a certain road is relatively small, it indicates that the passage condition of this road is relatively stable, and its passage weight coefficient is appropriately increased.

[0156] Through the above-mentioned correction, a dynamic road network impact factor is generated. This dynamic road network impact factor can reflect the actual situation of the current road network during the transportation of emergency resources in real time, providing more accurate road network information for the deep reinforcement learning recognition model in subsequent meteorological disaster recognition and response strategy generation, enabling the model to more reasonably consider the impact of road traffic conditions on disaster response.

[0157] Step S720: Perform a spatio-temporal superposition operation on the dynamic road network impact factor and the meteorological state transition probability map, update the diffusion direction probability distribution of the disaster evolution prediction path, adjust the high-risk area blockade boundary coordinates in the risk area division strategy based on the updated diffusion direction probability distribution, and transmit the adjusted blockade boundary coordinates back to the value network of the deep reinforcement learning recognition model for loss function compensation calculation.

[0158] In this embodiment, a spatio-temporal superposition operation is performed on the generated dynamic road network impact factor and the meteorological state transition probability map. The meteorological state transition probability map describes the transition possibilities of different meteorological states at different times and in different spaces. The dynamic road network impact factor reflects the impact of the road network on resource transportation and personnel movement during the response to meteorological disasters.

[0159] The process of the spatio-temporal superposition operation is to combine the road traffic information in the dynamic road network impact factor with the spatial and temporal information in the meteorological state transition probability map. For example, during a certain time period, due to poor road conditions (the dynamic road network impact factor shows a low traffic weight) on a certain road, the flow of people and materials between the affected areas is restricted, which will affect the diffusion speed and direction of the meteorological disaster. According to this impact, the disaster diffusion direction probability in the corresponding area and time in the meteorological state transition probability map is adjusted. For example, the original probability of the meteorological disaster spreading in a certain direction in a certain area is 0.6. Due to the road traffic problem, the diffusion probability in this direction is adjusted to 0.4, and at the same time, the diffusion probabilities in other feasible directions are increased.

[0160] Through the above spatio-temporal superposition operation, the diffusion direction probability distribution of the disaster evolution prediction path is updated. Based on the updated diffusion direction probability distribution, the high-risk area blockade boundary coordinates in the risk area division strategy are re-evaluated. If it is found that the diffusion probability of the disaster in a new direction increases, then the blockade boundary of the high-risk area in this direction is expanded accordingly. For example, originally, the blockade boundary of the high-risk area only considered the diffusion of the disaster in the northeast direction. Through the updated probability distribution, it is found that the diffusion probability in the southeast direction has increased significantly, so the blockade boundary of the high-risk area is extended in the southeast direction.

[0161] The adjusted coordinates of the lockdown boundary are sent back to the value network of the deep reinforcement learning recognition model for loss function compensation calculation. When calculating the losses brought by disasters (such as economic losses, casualties, etc.), the value network needs to consider the scope of high-risk areas. The adjusted coordinates of the lockdown boundary mean changes in the high-risk areas. Based on the new scope of high-risk areas, the value network recalculates factors such as the number of people and infrastructure that may be affected, and then adjusts the calculation results of the loss function. For example, if the high-risk area expands, it may involve more residents and important infrastructure, and the value network will correspondingly increase indicators such as the predicted value of disaster economic losses 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 optimization of response strategies.

[0162] In the embodiments of the present invention, by obtaining multi-dimensional meteorological time series data and geospatial coding data, the meteorological conditions of the target urban area can be comprehensively grasped; by mining the correlation features of meteorological elements, geospatial distribution features and meteorological change trend features through meteorological data preprocessing operations, rich and valuable information can be provided for subsequent analysis; by using the deep reinforcement learning recognition model for dynamic disaster pattern matching, a set of meteorological disaster recognition results including disaster type identification, influence scope and evolution prediction path can be accurately generated. The set of disaster response strategies generated based on this covers strategies such as emergency resource scheduling, risk area division and early warning information release, and can effectively respond to meteorological disasters. Further, dynamically optimizing the deep reinforcement learning recognition model and deploying it to the meteorological disaster monitoring system can, on the one hand, improve the model performance, and on the other hand, enhance the accuracy and real-time nature of meteorological disaster monitoring.

[0163] Based on the same inventive concept, the embodiments of the present invention also provide a system for identifying urban meteorological disaster data. Refer to Figure 2 shown, which is a schematic structural diagram of a possible system for identifying urban meteorological disaster data provided in the embodiments of the present invention. Figure 2 In it, the system 200 for identifying urban meteorological disaster data includes: a processor 210 and a memory 220. Among them, the memory 220 stores a computer program executable by the processor 210. By executing the instructions stored in the memory 220, the processor 210 can execute the steps of the above-mentioned method for identifying urban meteorological disaster data based on deep reinforcement learning.

[0164] Based on the same inventive concept, the embodiments of the present invention provide a computer-readable storage medium, which includes a computer program. When the computer program runs on the system for identifying urban meteorological disaster data, the computer program is used to make the system for identifying urban meteorological disaster data execute the steps of the above-mentioned method for identifying urban meteorological disaster data based on deep reinforcement learning.

[0165] In the technical solutions involved in the above embodiments of the present invention, whether it is for the comparison calculation of multi-dimensional features or the construction of composite parameters, when there are problems caused by significant differences in the number of dimensions, dimensional units, 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 meanings.

[0166] The formulas and calculation processes involved in the embodiments of the present invention, whether for multi-dimensional feature comparison or composite loss function construction, strictly follow the principle of dimensional correspondence. Each variable in the formula has a clear and definite physical meaning, and its operation logic fully conforms to basic mathematical and physical logics, and the operation result must be a reasonable result expected by the present invention. Those skilled in the art have the ability to effectively solve various problems brought about by the number of dimensions, dimensional differences, etc. in the multi-dimensional feature comparison calculation and composite loss function construction in the embodiments by comprehensively applying the above general technical means, ensuring the accuracy, reliability, and feasibility of the technical solutions of the present invention.

Claims

1. A method for identifying urban meteorological disaster data based on deep reinforcement learning, characterized in that, The method includes: Obtaining a set of meteorological monitoring data for a target urban area; Performing meteorological data preprocessing operations on the set of meteorological monitoring data to obtain a set of preprocessed meteorological spatio-temporal features; Invoking a trained deep reinforcement learning recognition model to perform dynamic disaster pattern matching processing on the set of meteorological spatio-temporal features to generate a set of meteorological disaster recognition results for the target urban area; Generating a set of disaster response strategies based on the set of meteorological disaster recognition results; Performing dynamic policy optimization processing on the deep reinforcement learning recognition model based on the set of disaster response strategies to obtain an optimized deep reinforcement learning recognition model, and deploying the optimized deep reinforcement learning recognition model to a meteorological disaster monitoring system.

2. The method according to claim 1, wherein The set of meteorological monitoring data includes multi-dimensional meteorological time series data and corresponding geospatial coding data. The set of meteorological spatio-temporal features includes meteorological element correlation features, geospatial distribution features, and meteorological change trend features. Performing meteorological data preprocessing operations on the set of meteorological monitoring data to obtain a set of preprocessed meteorological spatio-temporal features includes: Performing abnormal data cleaning processing on the multi-dimensional meteorological time series data, removing data segments that exceed a preset meteorological element fluctuation threshold to generate cleaned meteorological time series data, performing spatio-temporal alignment processing on the cleaned meteorological time series data to unify meteorological data with different time resolutions to a preset time benchmark, and performing spatial grid division according to the geospatial coding data to generate spatio-temporally aligned meteorological grid data, and extracting meteorological element correlation features in the spatio-temporally aligned meteorological grid data. The meteorological element correlation features include temperature-humidity coupling features, air pressure gradient change features, and wind speed-precipitation coordination features; Performing spatial topological relationship analysis processing on the geospatial coding data to generate geospatial distribution features including terrain elevation correlation features, urban building distribution features, and water system coverage features; Performing trend fitting processing on the spatio-temporally aligned meteorological grid data based on a sliding time window to generate meteorological change trend features including meteorological element cumulative change rate, meteorological event sudden probability, and meteorological state transition features; Combining the meteorological element correlation features, the geospatial distribution features, and the meteorological change trend features to generate the set of meteorological spatio-temporal features.

3. The method according to claim 2, wherein The invoking of the trained deep reinforcement learning recognition model to perform dynamic disaster pattern matching processing on the set of meteorological spatio-temporal features to generate the set of meteorological disaster recognition results for the target urban area includes: Inputting the meteorological element correlation features into the feature encoding layer of the deep reinforcement learning recognition model to generate meteorological element feature vectors; Inputting the geospatial distribution features into the spatial relationship analysis layer of the deep reinforcement learning recognition model to generate a spatial topological relationship matrix, where the spatial topological relationship matrix includes meteorological influence weights between different geographical units; Inputting 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; 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 a dynamic evaluation result including the matching degree of disaster trigger conditions, the similarity of disaster diffusion paths, and the scoring of disaster evolution stages; Based on the value network of the deep reinforcement learning recognition model, disaster impact value prediction is performed on the dynamic evaluation result, generating a meteorological disaster recognition result set including the predicted value of disaster economic losses, the personnel safety risk index, and the infrastructure vulnerability level.

4. The method according to claim 3, characterized in that, The disaster response strategy set includes emergency resource scheduling strategies, risk area division strategies, and early warning information release strategies. Based on the disaster response strategy set, dynamic strategy optimization processing is performed on the deep reinforcement learning recognition model, and the optimized deep reinforcement learning recognition model is obtained, including: Extract the actual material arrival time data from the emergency resource scheduling strategy, and extract the personnel evacuation path execution efficiency data from the risk area division strategy to generate a set of policy execution efficiency indicators; Input the set of policy execution efficiency indicators into the policy network backpropagation interface of the deep reinforcement learning recognition model to calculate the weight gradient deviation matrix of the disaster trigger condition matching degree; Based on the weight gradient deviation matrix, adjust the meteorological influence weight distribution ratio between different geographical units in the spatial topological relationship matrix to generate an optimized spatial topological relationship matrix; 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 a correction coefficient for the value network loss function; Use the optimized spatial topological relationship matrix to update the output parameters of the spatial relationship analysis layer, and at the same time adjust the error calculation weight of the predicted value of disaster economic losses through the correction coefficient of the value network loss function; Perform joint convergence verification on the updated parameters of the policy network and the value network. When the accuracy rate of the disaster evolution prediction path is increased to the preset threshold, output the optimized deep reinforcement learning recognition model.

5. The method according to claim 1, wherein The meteorological disaster recognition result set includes disaster type identification, disaster impact scope, and disaster evolution prediction path. Generate a disaster response strategy set according to the meteorological disaster recognition result set, including: Match the preset emergency resource type library according to the disaster type identification, and generate an emergency resource scheduling strategy including material reserve location, transportation route planning, and allocation priority; Based on the disaster impact scope, superimpose the urban population density distribution map to generate a risk area division strategy including high-risk area blockade plan, personnel evacuation path, and temporary resettlement point setting; According to the time series characteristics of the disaster evolution prediction path, generate an early warning information release strategy including early warning information release level, early warning coverage area priority, and information dissemination channel combination; Perform strategy conflict detection on the emergency resource scheduling strategy, risk area division strategy, and early warning information release strategy, and eliminate the execution time conflict and resource allocation conflict between different strategies to obtain an optimized disaster response strategy set; Encode the optimized set of disaster response strategies into an instruction format to be executed and push it to the corresponding emergency management terminal device.

6. The method according to claim 3, characterized in that, The method further includes: Real-time monitoring of the data recognition delay index and disaster omission rate index of the meteorological disaster monitoring system; When the data recognition delay index exceeds the preset response threshold, perform lightweight processing on the deep reinforcement learning recognition model: compress the number of neurons in the feature encoding layer, reduce the hidden layer dimension of the policy network, and quantize the parameter accuracy of the value network; When the disaster omission rate index exceeds the preset safety threshold, activate the incremental learning mode, and input the latest obtained meteorological disaster case data into the deep reinforcement learning recognition model for local parameter fine-tuning; Dynamically adjust the model compression rate and incremental learning frequency through the delay-accuracy balance algorithm, so that the generation efficiency and recognition accuracy of the meteorological disaster recognition result set maintain a preset balance state.

7. The method according to claim 1, wherein The method further includes: Construct a meteorological disaster recognition verification matrix, which includes a historical disaster case library, a simulated disaster scenario set, and real-time monitoring comparison data; Perform multi-dimensional comparison of the meteorological disaster recognition result set output by the deep reinforcement learning recognition model with the meteorological disaster recognition verification matrix. The multi-dimensional comparison includes: verification of disaster type matching degree, verification of overlapping degree of influence range, and verification of similarity of evolution path; When there are abnormal recognition results exceeding the preset error threshold in the comparison results, trigger the model interpretability analysis module to generate a visual analysis report including a feature weight distribution map, a trace of the policy decision path, and the basis for value evaluation; Locate the root feature dimension of the model recognition deviation according to the visual analysis report, and perform targeted optimization on the preprocessing process of the meteorological spatio-temporal feature set based on the root feature dimension.

8. The method according to claim 1, characterized in that, The method further includes: Deploy a distributed model update architecture including edge computing nodes and a cloud coordination center in the meteorological disaster monitoring system; Deploy a lightweight recognition model at the edge computing node to real-time process the meteorological spatio-temporal feature set in the local area to generate a preliminary recognition result; Deploy a full-scale recognition model at the cloud coordination center to receive abnormal feature data and high-risk recognition results uploaded by each edge computing node for review and analysis; Establish an edge-cloud model parameter synchronization mechanism. When the cloud full-scale recognition model detects a new disaster mode, encrypt and transmit the updated policy network parameters to the corresponding edge computing node; Optimize the spatial coverage impact weight and new disaster adaptation impact weight of the meteorological disaster monitoring system through the collaborative operation of local response of edge computing nodes and global policy optimization of the cloud.

9. The method according to claim 1, wherein The method further includes: Match the disaster type identifier in the meteorological disaster recognition result set with the cause chain nodes in the historical disaster case library to generate a disaster cause association vector; Input the disaster cause association vector into the graph reasoning module of the pre-constructed meteorological disaster knowledge graph to trigger the case strategy nodes associated with the geospatial distribution characteristics; Extract the response strategy templates in the case strategy nodes that meet the meteorological change trend characteristics, and generate a set of derivative strategies including cross-case resource scheduling rules, multi-disaster chain priority ranking, and new disaster analogy thresholds; Perform instruction-level fusion of the set of derivative strategies with the emergency resource scheduling strategies in the disaster response strategy set to generate an enhanced emergency resource scheduling strategy including a dynamic priority adjustment field and a resource conflict resolution protocol; Update the infrastructure vulnerability level evaluation parameters in the value network of the deep reinforcement learning recognition model using the instruction execution logs of the enhanced emergency resource scheduling strategy.

10. An urban meteorological disaster data recognition system, characterized in that, It includes a processor and a memory. Among them, the memory stores a computer program. When the computer program is executed by the processor, the processor is caused to execute the steps of any one of the methods recited in claims 1 to 9.

Images