Meteorological element-based multi-scene fire-fighting resource dynamic scheduling system

Abstract

The invention discloses a multi-scene fire-fighting resource dynamic scheduling system based on meteorological elements, belongs to the technical field of fire-fighting resource scheduling, and optimizes distribution of fire-fighting resources by using meteorological data and historical disaster situation data. By comprehensively analyzing meteorological information, resource conditions of the fire station and past disaster records, necessary fire resources can be deployed in advance so as to cope with potential disaster risks. Compared with the prior art, the system has the advantages that the system has the capability of dynamically adjusting resource allocation, can be flexibly adjusted according to specific disaster types and meteorological conditions, ensures that fire-fighting resources are timely and effectively distributed and utilized, adopts a multi-objective optimization technology, combines a machine algorithm and meteorological data, and improves the system reliability. Dynamic scheduling and intelligent configuration of fire-fighting resources are realized, global features of disaster prediction are captured through a shared feature layer, and resource demand prediction is further refined in a task-specific sub-network, so that a scheduling scheme is optimized.

Description

Technical Field

[0001] The present invention belongs to the technical field of fire fighting resource scheduling, and specifically relates to a multi-scenario dynamic scheduling system for fire fighting resources based on meteorological elements. Background Art

[0002] Fire fighting resource scheduling refers to the process of the fire department reasonably allocating and commanding various fire fighting resources in case of a fire or other emergencies. Its core goal is to quickly and effectively assemble the fire fighting and rescue forces to minimize casualties and property losses. Fire fighting resources mainly include fire fighters, fire trucks, fire extinguishing equipment, water sources, communication equipment, etc. The scheduling process usually relies on a professional command center. Through links such as receiving alarms, handling alarms, force deployment, and on-site command, and by using advanced communication technologies and information systems, real-time monitoring, dynamic allocation, and unified command of fire fighting resources are achieved. Efficient fire fighting resource scheduling requires a complete emergency plan, scientific scheduling strategies, professional command personnel, and excellent equipment support, and close cooperation with departments such as public security, medical, and transportation to form a linked emergency rescue system. With the acceleration of urbanization and the diversification of disaster accidents, fire fighting resource scheduling is developing towards the direction of intelligence, refinement, and collaboration, continuously improving the emergency response ability and rescue efficiency.

[0003] In the prior art, resource scheduling between different fire fighting management areas generally relies on manual command or traditional optimization methods. Traditional methods optimize the configuration and distribution of fire trucks by analyzing the geographical distance between the fire occurrence location and the fire stations, the number of fire trucks and their distribution, and using optimization algorithms such as linear programming and integer programming, or simply rely on the dispatcher's experience judgment to dispatch fire trucks and equipment.

[0004] However, the prior art has the following disadvantages: 1. Lack of real-time integration of meteorological elements: Existing systems fail to utilize meteorological data and historical disaster data, and cannot dynamically adjust the fire fighting resource configuration to cope with rescue tasks under different meteorological conditions.

[0005] 2. Static scheduling strategy: Traditional optimization algorithms mainly optimize resources based on static data and lack the ability to respond to environmental changes. Under sudden meteorological conditions, the fire fighting resource scheduling is not flexible enough.

[0006] 3. Single rescue scenario: The prior art mainly targets the fire scenario and lacks comprehensive consideration of other fire fighting rescue scenarios (such as geological disasters, typhoon scenarios, etc.), and cannot achieve dynamic scheduling of fire fighting resources in multiple scenarios. Summary of the Invention

[0007] The purpose of the present invention is to provide a multi-scenario dynamic scheduling system for fire fighting resources based on meteorological elements in order to solve the above-mentioned problems.

[0008] The technical solution adopted by the present invention is as follows: a multi-scenario fire resource dynamic scheduling system based on meteorological elements, the system comprising: a data acquisition module, a data processing module, a disaster prediction module, an optimal fire resource evaluation module, a regional scheduling model module and a user interaction interface module; The output end of the data acquisition module transmits real-time meteorological data, fire resource data and historical disaster data to the input end of the data processing module through a standardized interface; The output end of the data processing module transmits the structured data set to the input end of the disaster prediction module as input features of the random forest model; The output end of the disaster prediction module inputs the disaster probability data of each region into the input end of the optimal fire resource assessment module to drive the DNN-MTL multi-task learning model to predict the optimal resource configuration of a single station; The output end of the optimal fire resource assessment module inputs the single-station resource demand prediction result to the input end of the regional scheduling model module to execute the minimum cost flow algorithm and the genetic algorithm to optimize the cross-regional resource scheduling; The output end of the regional scheduling model module transmits the global scheduling plan to the input end of the user interaction interface module for visual display of the scheduling path, resource status and weather warning information; The monitoring feedback end of the user interaction interface module transmits the manual intervention instructions back to the control end of the regional scheduling model module through the API to support dynamic adjustment of the scheduling strategy.

[0009] In a preferred embodiment, the data acquisition module obtains fire resource data, meteorological data, historical disaster data and other information from the Municipal Meteorological Bureau, Municipal Fire Brigade and other departments through API and offline data sets for subsequent processing and model training, including the following data sources: 1. Fire station resource dataset: 2. Historical fire disaster data set: 3. Historical optimal fire resource scheduling data set: 4. Weather real-time data set: 5. Weather forecast dataset: 6. Historical meteorological data set.

[0010] In a preferred embodiment, the data processing module includes data association, data cleaning, missing value processing, and data standardization: Data association: Based on the geographical location information of the fire station, the system calculates the distance between each fire station and the surrounding weather stations through the geographical distance calculation formula, and selects the nearest weather station. The calculation formula is as follows: ; In the formula: is the distance between two points, is the radius of the Earth, , are the latitudes of the two points, , are the longitudes of the two points.

[0011] In a preferred embodiment, the data cleaning specifically includes: removing duplicate or abnormal records in the fire protection data, such as duplicate fire records, etc. Since the meteorological data of this system is obtained from official meteorological channels and has undergone data quality control, the meteorological data will not be cleaned again here.

[0012] The data standardization normalizes data with different dimensions so that they can be compared and calculated on the same scale. This solution uses the Z-score standardization method.

[0013] ; wherein, is the original data, is the mean of the data, is the standard deviation of the data.

[0014] In a preferred embodiment, the input features X of the disaster prediction module include: Meteorological actual situation features: precipitation duration days, cumulative precipitation during precipitation duration days, maximum temperature, humidity, maximum wind speed, air pressure; Meteorological forecast features: meteorological warning signals, meteorological forecasts for the next 10 days; Historical disaster situation features: disaster type, scale and impact of the disaster, longitude of the disaster event, latitude of the disaster event; Fire station features: fire station number, longitude and latitude of the fire station; Other features: whether it is a working day, season.

[0015] The output targets Y of the disaster prediction module include: Probability of being affected by high temperature disasters ; Probability of being affected by heavy rain disasters ; Probability of being affected by strong winds, etc. .

[0016] In a preferred embodiment, the model training steps of the disaster prediction module include: S1 Feature selection: To improve the performance of the model, we use feature importance analysis to screen key features: ; wherein, is the category in the data set proportion.

[0017] S2 Random Forest Parameter Settings: The following are the parameters based on the training set of this model. Other training sets should be adjusted according to the actual situation.

[0018] n_estimators = 200, max_depth = 15, min_samples_split = 5, min_samples_leaf = 3, max_features ='sqrt', min_impurity_decrease = 0.0, bootstrap = True, class_weight = 'balanced' S3 Random Forest Training Process: S3.1 Data Preparation: Construct a training set from historical data, where the data are features , and the target variable .

[0019] S3.2 Training Steps: (1) Generate the training dataset: Use historical meteorological data and disaster data , as input features, and y as the target disaster category; (2) Tree construction: Each tree divides the data by selecting a subset of features during training. Each node selects the optimal split using the Gini impurity; (3) Tree training: Each tree is trained by selecting the bootstrap sampling method, randomly selecting m samples from the data for training the th tree. Each tree selects the optimal feature for splitting using the Gini impurity: ; (4) Feature selection: When splitting each node, randomly select the number of features, set to .

[0020] S3.3 Prediction Steps: Each tree will give a predicted probability of a disaster occurring. If the model has N trees, then the final predicted probability of a disaster occurring is obtained by voting on the prediction results of all trees. Finally, we get: The probability of being affected by high temperature disasters ; The probability of being affected by heavy rain disasters ; Probability of being affected by disasters such as strong winds 。

[0021] Evaluation of the S4 model: Use indicators such as confusion matrix and ROC curve to test the performance of the evaluation model on the test set to ensure good generalization of the model on new data.

[0022] In a preferred embodiment, the DNN-MTL adopted by the optimal fire resource evaluation module extracts global features of disaster prediction through a shared feature layer, and predicts the demand for fire trucks, the demand for firefighters, and the dispatching response time in the task-specific sub-network. This model adopts a dynamic task weight adjustment method, enabling different tasks to adaptively adjust the loss weights according to their difficulty and data distribution. Specifically, it includes:

[0023] Shared feature layer: The calculation formula of the shared feature layer is as follows: ; represents the shared feature; represents the activation function, and the ReLU activation function is used here; represents the weight of the shared layer; represents the input feature X; represents the bias of the shared layer.

[0024] Task-specific sub-network layer: Take the output of the shared feature layer as the input of the specific sub-network layer and perform further processing through the task-specific sub-network: ; represents the specific feature of task i; represents the activation function, and the ReLU activation function is still used here; represents the weight of the shared layer; represents the bias of the shared layer; In addition, set the number of neurons in the hidden layer to 64 and the L2 regularization to λ = 0.01 to prevent overfitting and enhance the generalization ability of the model.

[0025] Training process: The overall loss function uses weighted mean squared error ,where is the loss of each task: ; Adopt a dynamic task weight adjustment method to enable different tasks to adaptively adjust the loss weight according to their difficulty and data distribution.

[0026] ; Among them, is the loss change of task . The greater the task loss fluctuation, the smaller its weight, ensuring that the model will not be overly biased towards a certain task. The optimizer uses Adam, and the learning rate is set to 0.001.

[0027] 5. Prediction output: The prediction result of each task is a continuous value. The activation function of the output layer uses a linear activation function, and the loss function uses MSE to obtain: ; is the prediction output of task i, is the parameter of the final output layer.

[0028] The output predicted by the final model is: .

[0029] In a preferred embodiment, the regional dispatching model module numbers 49 fire stations as , the current number of fire trucks in each fire station , the optimal fire resource prediction of each fire station , the minimum guarantee resource of each fire station The minimum guarantee resource of each fire station , the matrix of the number of allocated fire trucks , indicating the number of fire trucks allocated from fire station to , set the optimization goal as , where represents the comprehensive cost of the fire truck allocated from fire station to , represents the matrix of the number of dispatched fire trucks; The judgment of the fire resource status of the regional dispatching model module includes: S1. Calculate the surplus of resources: . When > 0, it means that the station has surplus fire trucks and can allocate them to other fire stations; when < 0, it means that the resources of the station are insufficient and need to be allocated from other fire stations.

[0030] S2. Calculate the allocable resources: . > 0 means that the station has allocable resources.

[0031] S3. Total fire truck dispatching constraint: Since the number of fire trucks in the whole city is fixed, this constraint is created to ensure that the total number of fire trucks remains unchanged.

[0032] The resource allocation matching method of the regional dispatching model module includes: Using minimum cost flow optimization to find the optimal resource allocation plan: ; Establish the following constraints: ; ; After allocation, the new resource quantity of the fire station is: ; The regional dispatching model module transforms the fire station resource dispatching problem into a cost flow network, regards the fire station as a node in the network, constructs edges between nodes, and the weight of the edge represents the dispatching cost of allocating resources from one fire station to another, and uses the minimum cost maximum flow algorithm to realize fire truck dispatching.

[0033] In a preferred embodiment, the regional dispatching model module reaches the optimal resource allocation plan by locally optimizing the fire truck allocation matrix M, and the specific steps include: S1 Initialize the population: Obtain multiple allocation matrices through cost flow network modeling, and each allocation matrix is a dispatching plan. During the dispatching process, give priority to the minimum resource reservation of each fire station and perform dynamic optimization on this basis to reasonably allocate the resource requirements of each station.

[0034] S2 Fitness function: Design a fitness function to measure the quality of the current dispatching plan. Our goal is to minimize the dispatching cost, that is, to minimize the cost of dispatching fire trucks from one station to another. The fitness function F(M) evaluates the quality of each individual by calculating the dispatching cost, and the formula is as follows: ; Where: is the dispatching cost, indicating the cost of dispatching from fire station to .

[0035] The second term is the penalty term for the minimum resource guarantee to ensure that the resource quantity of all fire stations is not lower than the minimum guarantee quantity.

[0036] is the weight factor of the minimum resource guarantee.

[0037] S3 Crossover and Mutation Operations: To simulate the natural selection process, crossover and mutation operations are performed on the individuals in the population to generate new individuals. This can ensure the diversity of the solution space and explore possible optimal solutions.

[0038] The goal of the crossover operation is to combine two relatively good individuals to produce a new individual. In this problem, we adopt the partially matched crossover algorithm.

[0039] The purpose of the mutation operation is to introduce diversity into the solution space and prevent the genetic algorithm from falling into local optimal solutions. In the optimization of fire-fighting resource scheduling, the mutation operation is achieved by randomly changing some elements in the allocation matrix, and the mutation rate is set to 5% here.

[0040] Finally, a new allocation matrix of fire trucks is obtained to ensure the maximization of resource allocation and the minimization of scheduling costs.

[0041] S4 Termination Conditions: The goal of the genetic algorithm is to find the optimal scheduling scheme rather than simply performing random searches. To avoid excessive calculation of the algorithm, certain termination conditions are set. In this problem, 100 rounds of iteration are set as the maximum training period, and at the same time, the convergence of the fitness value is judged. When the change in the fitness value is less than the set threshold or the number of iterations reaches the maximum limit, the algorithm stops.

[0042] In a preferred embodiment, the user interface module includes: a meteorological module, a resource management module, a scheduling result analysis module, and a user management module.

[0043] In summary, due to the adoption of the above technical solutions, the beneficial effects of the present invention are as follows: 1. In the present invention, meteorological data and historical disaster data are used to optimize the allocation of fire-fighting resources. By comprehensively analyzing meteorological information, the resource status of fire stations, and past disaster records, necessary fire-fighting resources can be deployed in advance to cope with potential disaster risks. Compared with traditional technologies, the advantage of this system lies in its ability to dynamically adjust resource allocation, which can be flexibly adjusted according to specific disaster types and meteorological conditions to ensure the timely and effective allocation and utilization of fire-fighting resources.

[0044] 2. In the present invention, multi-objective optimization technology is adopted and combined with machine algorithms and meteorological data to achieve dynamic scheduling and intelligent allocation of fire-fighting resources. The system uses the DNN-MTL multi-task learning model to predict multiple target tasks, including the number of fire trucks, the demand for firefighters, etc. This model captures the global features of disaster prediction through a shared feature layer and further refines the prediction of resource requirements in task-specific sub-networks, thereby optimizing the scheduling scheme.

[0045] 3. In the present invention, the limited fire resources can be intelligently adjusted according to meteorological data and historical disaster situations. Meanwhile, through the optimization of the minimum cost maximum flow algorithm and genetic algorithm, the scheduling is further optimized. This system makes full use of meteorological data and fire data, and can effectively improve the efficiency of fire resource allocation and the advance deployment volume of disaster response. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 It is the overall system block diagram of the present invention; Figure 2 It is the system block diagram of the data processing module of the present invention; Figure 3 It is the system block diagram of the disaster prediction module of the present invention; Figure 4 It is the system block diagram of the optimal fire resource evaluation module of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0047] In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention, and are not used to limit the present invention.

[0048] Embodiment: Refer to Figures 1-4 , A multi-scenario fire resource dynamic scheduling system based on meteorological elements, the system includes: a data acquisition module, a data processing module, a disaster prediction module, an optimal fire resource evaluation module, a regional scheduling model module and a user interaction interface module; I. Data Acquisition Module The data acquisition module is the basic part of the whole system, and obtains information such as fire resource data, meteorological data, historical disaster situation data, etc. from departments such as the municipal meteorological bureau and the municipal fire brigade through APIs and offline data sets for subsequent processing and model training. Specifically, it includes the following data sources: 1. Fire station resource data set: 1.1 Fire station geographical location information: The longitude and latitude information of each fire station is the basis for determining its jurisdiction area. The system obtains the location of each fire station in real time, as well as the specific area range (boundary data of towns or streets, saved in vector format, such as *.shp file) responsible for this fire station.

[0049] 1.2 Number and type of fire trucks: The number and type of fire trucks in each fire station. Different types of fire trucks (such as water tanker fire trucks, foam fire trucks, aerial platform fire trucks, special fire trucks, etc.) have different application scenarios.

[0050] 1.3 Fire truck status: Get real-time status information of each fire truck, including whether it is in maintenance, dispatched, idle, etc.

[0051] 2. Historical fire disaster (rescue case) dataset: Obtain the geographical location (latitude and longitude), time, type, scale, resources invested, rescue time, economic losses, casualties and other information of fire disasters (rescue cases) in Wenzhou from 2010 to 2024 to help the model identify the regularity of disasters. The historical fire disaster data set is very important, providing reference data based on historical records for subsequent training of the system, helping the system to identify the characteristics of disasters under different meteorological conditions.

[0052] Example data: On x month x day, 2024: A packaging factory (latitude and longitude), building fire, 4 hours, 6 fire trucks, 30 firefighters, extinguishing time 3 hours, economic loss of approximately 2 million.

[0053] 3. Historical optimal fire resource scheduling data set The optimal resource dispatch plan (manual calculation) corresponding to historical rescue cases is obtained through the fire rescue department. These data include the number of firefighters, fire trucks, and dispatch response time that should be deployed in each fire station.

[0054] 4. Weather real-time data set: Obtain current meteorological data in real time through the interface of the meteorological department, including wind speed, wind direction, temperature, humidity, atmospheric pressure, precipitation, consecutive days without precipitation and other information.

[0055] 5. Weather forecast dataset: Obtain future weather forecast information through the meteorological department's data interface, including daily forecasts for the next 10 days and weather warning signals.

[0056] 6. Meteorological historical data set: Through the meteorological department, we obtain information including temperature, humidity, wind speed, precipitation, etc. from 2010 to 2024, and use this dataset to systematically identify the relationship between meteorological factors and disaster occurrence.

[0057] 2. Data processing module: The function of this module is to preprocess the collected multi-source data, mainly including data association, data cleaning, missing value processing, and data standardization.

[0058] 1. Data association: The system first ensures that meteorological data can be effectively matched with the historical disaster data and optimal dispatching plans of each fire station. Based on the geographical location information (latitude and longitude coordinates) of the fire stations, the system calculates the distances between each fire station and the surrounding weather stations through a geographical distance calculation formula (Haversine formula), and selects the weather station with the shortest distance (since the density of weather stations in the application area of this system has reached one every 5 kilometers, the system can reliably match the optimal meteorological data, and there is no situation where the distance between the weather station and the fire station is too far, resulting in the data being unreferenceable). The calculation formula is as follows: ; In the formula: is the distance between two points (unit: kilometers), is the radius of the earth, , are the latitudes of the two points, , are the longitudes of the two points.

[0059] The format of the associated data is shown in the following list: ; 2. Data cleaning: Remove duplicate or abnormal records in the fire data, such as duplicate fire records, etc. Since the meteorological data of this system is obtained from the official meteorological channels and has undergone data quality control, the meteorological data will not be cleaned here again.

[0060] 3. Missing value processing: For the missing fire data: Use mean filling for processing, such as missing data on the number of fire trucks, etc.

[0061] For the missing meteorological data: Considering the relatively dense distribution of weather stations, use the data of the nearest station to the missing data station for filling. For example, the temperature value at a certain whole hour is missing.

[0062] 4. Data standardization: Normalize data with different dimensions so that they can be compared and calculated on the same scale. This solution uses the Z-score standardization method.

[0063] ; Among them, is the original data, is the mean of the data, is the standard deviation of the data.

[0064] III. Disaster prediction module: First, it is necessary to predict the probability of a disaster occurring in a certain area (the area responsible for by this fire station) within a specific time window. The random forest algorithm is used for multi-classification (the probability of being affected by high temperature, heavy rain, strong wind, etc.) prediction: 1. Input feature X: 1.1 Meteorological actual situation features Precipitation duration days, cumulative precipitation during precipitation duration days, maximum temperature, humidity, maximum wind speed, air pressure 1.2 Meteorological forecast features: Meteorological warning signals, meteorological forecasts for the next 10 days (including 5 forecast elements: temperature, precipitation, wind speed, air pressure, humidity) 1.3 Historical disaster situation features: Disaster type, scale and impact of the disaster, longitude of the disaster event, latitude of the disaster event 1.4 Fire station features: Fire station number, longitude and latitude of the fire station 1.5 Other features: Whether it is a working day, season 2. Output target Y: Probability of being affected by high temperature ; Probability of being affected by heavy rain ; Probability of being affected by strong wind, etc. ; 3. Model training steps: 3.1 Feature selection To improve the performance of the model, we use feature importance analysis (based on Gini impurity) to screen key features: ; Among them, is the category in the data set the proportion in.

[0065] 3.2 Random forest parameter setting The following are the parameters based on the training set of this model. Other training sets are adjusted according to the actual situation.

[0066] n_estimators = 200, max_depth = 15, min_samples_split = 5, min_samples_leaf = 3, max_features ='sqrt', min_impurity_decrease = 0.0, bootstrap = True, class_weight = 'balanced' 3.3 Random Forest Training Process: 3.3.1 Data Preparation Construct a training set from historical data, where the data are features (select the standardized data feature items listed in "1. Input Feature X" above), and the target variable ; 3.3.2 Training Steps (1) Generate the training data set: Use historical meteorological data and disaster data , as input features, and y as the target disaster category; (2) Tree construction: For each tree divide the data by selecting a subset of features during training . Each node selects the optimal split using the Gini impurity; (3) Tree training: The training of each tree is done by selecting the bootstrap sampling method, randomly selecting m samples from the data for training the th tree. Each tree selects the optimal feature for splitting using the Gini impurity: ; (4) Feature selection: When splitting each node, randomly select the number of features, set to (where m is the total number of features).

[0067] 3.3.3 Prediction Steps Each tree will give a predicted probability of a disaster occurring. If the model has N trees, then the final predicted probability of a disaster occurring is obtained by voting on the prediction results of all trees, and finally: The probability of being affected by high temperature disasters ; The probability of being affected by heavy rain disasters ; The probability of being affected by strong winds, etc. ; 3.4 Model Evaluation Use metrics such as confusion matrix and ROC curve to evaluate the performance of the model on the test set to ensure good generalization of the model on new data.

[0068] IV. Optimal Fire Resource Evaluation Module: Further evaluate the optimal fire resources required for a single fire station based on the disaster probability and meteorological data.

[0069] After obtaining the probability of a disaster occurring in the management area of a certain fire station, to continue obtaining the optimal resource allocation of this fire station, we need to transform this task into a multi-objective regression problem. Here, a DNN-MTL multi-task learning model is designed to predict multiple target tasks. Specifically, it includes: 1. Input feature X: 1.1 Disaster prediction probability: Probability of disaster caused by high temperature, P_temp Probability of disaster caused by heavy rain, P_rain Probability of disaster caused by strong wind, etc., P_wind 1.2 Meteorological features: Weather conditions in the past 30 days: (where represents temperature, represents humidity, represents precipitation, represents wind speed, represents air pressure). Weather forecast for the next 10 days: .

[0070] 1.3 Historical optimal fire fighting dispatch resources: Based on the dispatch information from 2011 to 2023: (where represents the past fire truck dispatch situation, represents the past fire fighter dispatch situation, represents the past response time).

[0071] 1.4 Fire station resources: Fire station resources: (Each parameter represents the total number of fire trucks, the number of water tank fire trucks, the number of foam fire trucks, the number of aerial ladder fire trucks, the number of special fire trucks, the number of fire fighters, and the fire station number).

[0072] 2. Target variable Y: ( : The number of water tank fire trucks required, : The number of foam fire trucks required, : The number of aerial ladder fire trucks required, : The number of special fire trucks required, : The number of fire fighters required, : Dispatch response time).

[0073] 3. Design of DNN-MTL multi-task learning model: DNN-MTL (Deep Neural Network - Multi-Task Learning) is a multi-task learning method based on deep neural networks, which can optimize multiple related prediction targets simultaneously. The DNN-MTL adopted by this model extracts global features for disaster prediction through a shared feature layer, and predicts the demand for fire trucks, the demand for firefighters, and the dispatching response time in the task-specific sub-networks. This model adopts a dynamic task weight adjustment method, enabling different tasks to adaptively adjust the loss weights according to their difficulty and data distribution.

[0074] 3.1 Shared Feature Layer: The calculation formula of the shared feature layer is as follows: ; represents the shared feature; represents the activation function, and here the ReLU activation function is used; represents the weight of the shared layer; represents the input feature X; represents the bias of the shared layer.

[0075] 3.2 Task-Specific Sub-Network Layer: Take the output of the shared feature layer as the input of the specific sub-network layer and perform further processing through the task-specific sub-network: ; represents the specific feature of task i (fire truck prediction, firefighter prediction); represents the activation function, and here the ReLU activation function is still used; represents the weight of the shared layer; represents the bias of the shared layer.

[0076] In addition, set the number of hidden layer neurons to 64 and the L2 regularization to λ = 0.01 to prevent overfitting and enhance the generalization ability of the model.

[0077] 4. Training Process: 4.1 The overall loss function uses weighted mean squared error , where is the loss of each task: ( actual value, predicted value).

[0078] 4.2 Adopt a dynamic task weight adjustment method so that different tasks adaptively adjust the loss weights according to their difficulty and data distribution.

[0079] ; Among them, is the loss change of task . The greater the task loss fluctuation, the smaller its weight, ensuring that the model will not be overly biased towards a certain task.

[0080] Here, the optimizer uses Adam and the learning rate is set to 0.001.

[0081] 5. Prediction output: The prediction results of each task (fire truck prediction, firefighter prediction, dispatching response time) are continuous numerical values. The activation function of the output layer uses a linear activation function, and the loss function uses MSE (mean squared error) to obtain: ; is the predicted output of task i, is the parameter of the final output layer.

[0082] The output predicted by the final model is: .

[0083] V. Regional dispatching model module: 1. Task objectives: 1.1 The total number of fire trucks remains unchanged, and resources are dynamically allocated among different regions.

[0084] 1.2 The response time is the shortest, and the dispatching considers geographical distance, driving time, and path optimization.

[0085] 1.3 Ensure the minimum resource guarantee for fire stations, ensuring that the resources of each fire station are not lower than the set safety threshold.

[0086] 1.4 Intelligent dispatching, based on "minimum cost flow optimization + genetic algorithm optimization", to achieve reasonable resource allocation.

[0087] 2. Regional fire resource dispatching optimization modeling: Number the 49 fire stations as , the current number of fire trucks at each fire station , the optimal fire resource prediction for each fire station (from the prediction of the previous step), the minimum guaranteed resources for each fire station, the minimum guaranteed resources for each fire station , the matrix of the number of transferred fire trucks , indicating from fire station transfer to The number of fire trucks, and the optimization objective is set as (where represents the comprehensive cost of dispatching a fire truck from the fire station to , and represents the matrix of the number of dispatching fire trucks.) 3. Judgment of fire-fighting resource status: 3.1 Calculate the surplus of resources: . When > 0, it means that there are surplus fire trucks at this station and they can be allocated to other fire stations; when < 0, it means that the resources at this station are insufficient and need to be allocated from other fire stations.

[0088] 3.2. Calculate the allocable resources: . > 0 indicates that there are allocable resources at this station.

[0089] 3.3. Constraint on the total amount of fire truck dispatching: , the number of fire trucks in the whole city is fixed. Create this constraint to ensure that the total number of fire trucks remains unchanged.

[0090] 4. Matching of resource allocation: Adopt the minimum cost flow optimization to find the optimal resource allocation plan: ; Establish the following constraints: ; ; After allocation, the new resource quantity of the fire station is: ; 5. Construct the cost flow network: Transform the fire station resource dispatching problem into a cost flow network. Regard the fire stations as nodes in the network and construct edges between the nodes. The weight of the edge represents the dispatching cost of allocating resources from one fire station to another. Use the minimum cost maximum flow algorithm (MCMF) to realize the fire truck dispatching.

[0091] 5.1 Cost flow network modeling: Transform the problem into a flow network. Each node represents a fire station, the capacity of each edge represents the maximum allocation quantity of resources, and the weight (i.e., cost) of the edge represents the cost of allocating resources.

[0092] Node : Each node represents a fire station. Edge : Each edge represents the allocation from the fire station Route to the fire station The route for allocating resources, where the capacity of the edge is the number of resources that can be allocated, and the weight (cost) is the scheduling cost.

[0093] 5.2. Names of each element in the network diagram: Source node and sink node: Select a virtual source node and a virtual sink node (The source node is responsible for allocating surplus fire truck resources to each fire station, and the sink node is responsible for receiving resources allocated to other fire stations).

[0094] Fire station node: Each fire station has an inflow capacity and an outflow capacity, indicating how many resources the station can receive and allocate. The fire station's resource demand and the adjustable resource quantity jointly affect the allocation plan.

[0095] Allocation cost (weight of the edge): The weight of each edge represents the allocation cost from the fire station to the fire station Allocation cost. The scheduling cost takes into account factors such as geographical distance, travel time, and scheduling difficulty, denoted as , the capacity of the edge is the maximum allocation quantity of the fire truck, denoted as .

[0096] 5.3 Minimum-cost maximum-flow algorithm, specifically including: Construct nodes and edges: For each fire station and each other fire station , we add an edge from to , the capacity of the edge is the adjustable resource quantity , and the weight is the scheduling cost .

[0097] Source node and sink node: The capacity of the source node is the total supply of fire trucks, that is .

[0098] Sink node has a capacity equal to the total demand of the target fire station, that is .

[0099] Flow balance of the network: The inflow of each fire station . should be equal to its outflow to ensure that the allocation quantity is within a reasonable range: ; Run the algorithm: Calculate the maximum flow from the source node to the sink node and minimize the cost required for the flow. At each step, transfer resources by finding the shortest path (i.e., the minimum-cost path) while satisfying all resource demands.

[0100] 6. User Interaction Interface Module: The user interaction interface module includes: a meteorological module, a resource management module, a scheduling result analysis module, and a user management module.

[0101] Scheduling result output: The final result is the scheduling plan for each fire station , that is, which sites each fire station needs to allocate resources from.

[0102] 6. Optimize through the genetic algorithm, specifically including: 6.1 Initialize the population: Obtain multiple allocation matrices through cost flow network modeling. Each allocation matrix is a scheduling plan. During the scheduling process, prioritize the minimum resource retention of each fire station and perform dynamic optimization on this basis to reasonably allocate the resource requirements of each site.

[0103] 6.2 Fitness function: Design a fitness function to measure the quality of the current scheduling plan. Our goal is to minimize the scheduling cost, that is, to minimize the cost of transferring fire trucks from one site to another. The fitness function F(M) evaluates the quality of each individual (allocation plan) by calculating the scheduling cost, and the formula is as follows: ; Where: is the scheduling cost, indicating the cost of transferring from fire station to .

[0104] The second term is a penalty term for the minimum resource guarantee to ensure that the resource volume of all fire stations is not lower than the minimum guarantee volume.

[0105] is the weight factor of the minimum resource guarantee.

[0106] 6.3 Crossover and mutation operations: To simulate the process of natural selection, perform crossover and mutation operations on the individuals in the population to generate new individuals. This can ensure the diversity of the solution space and explore possible optimal solutions.

[0107] The goal of the crossover operation is to combine two better individuals (parent generations) to produce a new individual (offspring generation). In this problem, we use the partially mapped crossover (PMX) algorithm.

[0108] The purpose of the mutation operation is to introduce the diversity of the solution space and prevent the genetic algorithm from falling into a local optimal solution. In the optimization of fire resource scheduling, the mutation operation is achieved by randomly changing some elements in the allocation matrix, and the mutation rate is set to 5% here.

[0109] Finally, a new fire truck allocation matrix is obtained to ensure the maximized allocation of resources while minimizing the scheduling cost.

[0110] 6.4 Termination condition: The goal of the genetic algorithm is to find the optimal scheduling plan rather than simply conducting random searches. To avoid excessive computation of the algorithm, certain termination conditions are set. In this problem, 100 rounds of iteration are set as the maximum training period, and at the same time, the convergence of the fitness value is judged. When the change in the fitness value is less than the set threshold, or the number of iterations reaches the maximum limit (100 rounds), the algorithm stops.

[0111] 7. Output the optimal scheduling conclusion: For each fire station, The specific scheduling tasks, including the number of fire trucks allocated from this station and the other stations to which they are dispatched.

[0112] VI. User interface module: The user interface module includes: a meteorological module, a resource management module, a scheduling result analysis module, and a user management module.

[0113] As can be seen from the above: In the present invention, meteorological data and historical disaster data are used to optimize the allocation of fire resources. By comprehensively analyzing meteorological information, the resource status of fire stations, and past disaster records, necessary fire resources can be deployed in advance to cope with potential disaster risks. Compared with traditional technologies, the advantage of this system lies in its ability to dynamically adjust resource allocation, which can be flexibly adjusted according to specific disaster types and meteorological conditions to ensure the timely and effective allocation and utilization of fire resources.

[0114] In the present invention, multi-objective optimization technology is adopted, combined with machine algorithms and meteorological data, to achieve the dynamic scheduling and intelligent allocation of fire resources. The system uses the DNN-MTL multi-task learning model to predict multiple target tasks, including the number of fire trucks, the demand for firefighters, etc. This model captures the global features of disaster prediction through a shared feature layer and further refines the resource demand prediction in task-specific sub-networks, thereby optimizing the scheduling plan.

[0115] In the present invention, the limited fire resources can be intelligently adjusted according to meteorological data and historical disasters. At the same time, through the optimization of the minimum cost maximum flow algorithm and the genetic algorithm, the scheduling is further optimized. This system makes full use of meteorological data and fire data, which can effectively improve the efficiency of fire resource allocation and the amount of advance deployment for disaster response.

[0116] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the term "comprise" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0117] The above description enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A multi-scenario fire resource dynamic scheduling system based on meteorological elements, characterized by: The system includes: a data acquisition module, a data processing module, a disaster prediction module, an optimal fire resource assessment module, a regional scheduling model module and a user interaction interface module; The output end of the data acquisition module transmits real-time meteorological data, fire resource data and historical disaster data to the input end of the data processing module through a standardized interface; The output end of the data processing module transmits the structured data set to the input end of the disaster prediction module as input features of the random forest model; The output end of the disaster prediction module inputs the disaster probability data of each region into the input end of the optimal fire resource assessment module to drive the DNN-MTL multi-task learning model to predict the optimal resource configuration of a single station; The output end of the optimal fire resource assessment module inputs the single-station resource demand prediction result to the input end of the regional scheduling model module to execute the minimum cost flow algorithm and the genetic algorithm to optimize the cross-regional resource scheduling; The output end of the regional scheduling model module transmits the global scheduling plan to the input end of the user interaction interface module for visual display of the scheduling path, resource status and weather warning information; The monitoring feedback end of the user interaction interface module transmits the manual intervention instructions back to the control end of the regional scheduling model module through the API to support dynamic adjustment of the scheduling strategy.

2. A multi-scenario fire resource dynamic scheduling system based on meteorological elements as claimed in claim 1, characterized in that: The data acquisition module obtains fire resource data, meteorological data, and historical disaster data information from the Municipal Meteorological Bureau and the Municipal Fire Brigade through API and offline data sets for subsequent processing and model training, including the following data sources:

1. Fire station resource data set, 2. Historical fire disaster data set, 3. Historical optimal fire resource dispatch data set, 4. Actual weather data set, 5. Weather forecast data set, 6. Weather history data set.

3. A multi-scenario fire resource dynamic scheduling system based on meteorological elements as claimed in claim 1, characterized in that: The data processing module includes data association, data cleaning, missing value processing, and data standardization: Data association is based on the geographical location information of the fire station. The system calculates the distance between each fire station and the surrounding weather stations through the geographical distance calculation formula and selects the nearest weather station; The calculation formula is as follows: ； In the formula: is the distance between two points, is the radius of the Earth, , are the latitudes of the two points, , is the longitude of the two points.

4. A multi-scenario fire resource dynamic scheduling system based on meteorological elements as claimed in claim 3, characterized in that: The data cleaning specifically includes: removing duplicate or abnormal records in firefighting data; The data standardization normalizes data of different dimensions so that they can be compared and calculated on the same scale, using the Z-score standardization method; ； in, is the original data, is the mean of the data, is the standard deviation of the data.

5. The multi-scenario firefighting resource dynamic scheduling system based on meteorological elements as claimed in claim 1, characterized in that: The input features X of the disaster prediction module include: Weather conditions: duration of precipitation, cumulative precipitation over the duration of precipitation, maximum temperature, humidity, maximum wind speed, and air pressure; Weather forecast features: weather warning signals, weather forecast for the next 10 days; Historical disaster characteristics: disaster type, disaster scale and impact, longitude of disaster event, latitude of disaster event; Fire station characteristics: fire station number, fire station latitude and longitude; Other features: whether it is a weekday, season; The output target Y of the disaster prediction module includes: Probability of being affected by high temperature disasters ; Probability of disaster caused by heavy rain ; Probability of being affected by strong winds .

6. A multi-scenario fire resource dynamic scheduling system based on meteorological elements as claimed in claim 1, characterized in that: The model training steps of the disaster prediction module include: S1 Feature Selection: In order to improve the performance of the model, feature importance analysis is used to screen key features: ； in, Yes Category In the dataset The proportion of S2 Random Forest Parameter Settings: The following are the parameters based on the training set of this model. Other training sets are adjusted according to actual conditions. S3 random forest training process: S3.1 Data preparation: Construct training sets from historical data, with data as features , the target variable ; S3.2 Training steps: (1) Generate training data set: Use historical meteorological data and disaster data , is the input feature, y is the target disaster category; (2) Tree construction: Each tree By selecting a subset of features during training To divide the data; each node selects the optimal partition through Gini impurity; (3) Tree training: Each tree is trained by randomly selecting m samples from the data using the bootstrap sampling method. trees; each tree selects the best feature to split by Gini impurity: ; (4) Feature selection: When each node is split, the number of features is randomly selected and set to ; S3.3 Prediction steps: Each tree A predicted probability of disaster occurrence will be given. If the model has N trees, the final probability of disaster occurrence can be obtained by voting on the predicted results of all trees. Evaluation of the S4 model.

7. A multi-scenario fire resource dynamic scheduling system based on meteorological elements as claimed in claim 1, characterized in that: The DNN-MTL adopted by the optimal fire resource assessment module extracts global features of disaster prediction through a shared feature layer, and predicts fire truck demand, firefighter demand, and dispatch response time in a task-specific subnetwork; The training process includes: The overall loss function of S1 uses weighted mean square error ,in is the loss for each task: ； S2 adopts a dynamic task weight adjustment method, which allows different tasks to adaptively adjust the loss weight according to their difficulty and data distribution; ； in, It's a task The greater the fluctuation of task loss, the smaller its weight is, ensuring that the model will not be overly biased towards a certain task; the optimizer uses Adam, and the learning rate is set to 0.001; Prediction output: The prediction result of each task is a continuous value. The output layer activation function uses a linear activation function, and the loss function uses MSE to obtain: ； is the predicted output of task i, is the final output layer parameter; The final model predicted output is: 。 8. The multi-scenario fire resource dynamic scheduling system based on meteorological elements as claimed in claim 1, characterized in that: The fire resource status judgment of the regional dispatch model module includes: S1. Redundancy of computing resources: ;when When >0, it means that the station has extra fire trucks that can be deployed to other fire stations; When <0, it means that the station has insufficient resources and needs to be allocated from other fire stations; S2. Calculate available resources: ; When >0, it means that the station has resources available for allocation; S3. Constraints on the total number of fire truck dispatches: , the number of fire trucks in the city is fixed, and this constraint is created to ensure that the total number of fire trucks remains unchanged; The resource allocation and matching method of the regional scheduling model module includes: Use minimum cost flow optimization to find the optimal resource allocation solution: ； Create the following constraints: ； ； After the transfer, the fire station The new resource quantities are: 。 9. The multi-scenario fire resource dynamic scheduling system based on meteorological elements as claimed in claim 1, characterized in that: The regional dispatch model module achieves the optimal resource allocation solution by locally optimizing the fire truck allocation matrix M. The specific steps include: S1 Initialize the population: multiple allocation matrices are obtained through cost flow network modeling, and each allocation matrix is ​​a scheduling scheme; in the scheduling process, the minimum resource reservation of each fire station is given priority, and dynamic optimization is performed on this basis, so that the resource demand of each station is reasonably allocated; S2 fitness function: The fitness function is designed to measure the quality of the current dispatch plan; that is, to minimize the cost of dispatching fire trucks from one station to another; the fitness function F(M) evaluates the quality of each individual by calculating the dispatch cost, the formula is as follows: ； in: is the dispatch cost, representing the dispatch cost from the fire station Transfer to Costs; The second item is a penalty item for the minimum resource guarantee, which ensures that the resource volume of all fire stations does not fall below the minimum guarantee volume; is the weight factor of the minimum resource guarantee; S3 Crossover and mutation operations: In order to simulate the natural selection process, crossover and mutation operations are performed on individuals in the population to generate new individuals; Combining two better individuals through crossover operation to produce a new individual; Introducing diversity in the solution space through mutation operations to prevent the genetic algorithm from falling into a local optimal solution; Finally, a new fire truck dispatch matrix is ​​obtained to ensure maximum resource allocation while minimizing dispatch costs. S4 termination condition: The goal of the genetic algorithm is to find the optimal scheduling solution rather than simply perform random search; in order to avoid excessive calculation of the algorithm, 100 iterations are set as the maximum training cycle, and the convergence of the fitness value is judged at the same time. When the fitness value change is less than the set threshold, or the number of iterations reaches the maximum limit, the algorithm stops.

10. The multi-scenario firefighting resource dynamic scheduling system based on meteorological elements as claimed in claim 1, characterized in that: The user interaction interface module includes: a meteorological module, a resource management module, a scheduling result analysis module and a user management module.

Images