A multi-disaster cascade evolution time window intelligent prediction method and system

Abstract

The application discloses a kind of multi-disaster kind cascade evolution time window intelligent prediction method and system, it is related to data processing technical field, including: acquisition multi-source historical disaster data set, carry out space-time alignment and multi-disaster kind cascade analysis, generate multi-disaster kind cascade topology graph;Multi-disaster kind cascade topology graph is carried out evolution trend prediction training based on multi-source historical disaster data set, constructs multi-disaster kind cascade evolution prediction model;Acquisition multi-source disaster real-time data, using multi-disaster kind cascade evolution prediction model to multi-source disaster real-time data is deduced prediction, obtain multi-disaster kind cascade evolution path and corresponding multi-disaster kind evolution prediction time window, to multi-disaster kind cascade evolution prediction model carries out closed loop prediction optimization.The application solves the technical problem that existing technology is difficult to effectively model and realize accurate time window prediction in multi-disaster kind cascade evolution process, reaches the technical effect of improving multi-disaster kind cascade evolution process prediction accuracy and time window determination precision.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, specifically to a method and system for intelligent prediction of multi-hazard cascade evolution time windows. Background Technology

[0002] In the context of complex coupling of multiple hazards, significant spatiotemporal correlations and triggering relationships often exist between different hazard types such as earthquakes, floods, landslides, and typhoons. The occurrence of a single hazard event may trigger or exacerbate a chain reaction of subsequent hazard responses of multiple types, forming a cascading evolution process with propagation and recursion. Simultaneously, hazard data comes from diverse sources, encompassing multiple systems such as meteorology, hydrology, geology, and emergency monitoring. These data vary considerably in temporal granularity, spatial scale, and representation, making it difficult to uniformly characterize the correlations between hazard events. Under these circumstances, systematically modeling the cascading evolution process of multiple hazards is challenging, and it is difficult to accurately characterize the development rhythm of the hazard chain and the timing of key nodes, thus affecting the effective prediction and judgment of hazard evolution time windows. Summary of the Invention

[0003] This application provides an intelligent prediction method and system for the time window of multi-hazard cascade evolution, which is used to address the technical problem that it is difficult to effectively model the multi-hazard cascade evolution process and achieve accurate time window prediction in the existing technology.

[0004] In view of the above problems, this application provides a method and system for intelligent prediction of multi-hazard cascade evolution time windows.

[0005] The first aspect of this application provides an intelligent prediction method for multi-hazard cascade evolution time windows, the method comprising: A multi-source historical disaster dataset is collected, and spatiotemporal alignment and multi-hazard cascade analysis are performed on the dataset to generate a multi-hazard cascade topology map. Based on the multi-source historical disaster dataset, the multi-hazard cascade topology map is trained for evolutionary trend prediction to construct a multi-hazard cascade evolution prediction model. Real-time multi-source disaster data is collected, and the multi-hazard cascade evolution prediction model is used to extrapolate and predict the real-time multi-source disaster data, obtaining the multi-hazard cascade evolution path and the corresponding multi-hazard evolution prediction time window. Based on the multi-hazard cascade evolution path and the corresponding multi-hazard evolution prediction time window, the multi-hazard cascade evolution prediction model is optimized through closed-loop prediction.

[0006] A second aspect of this application provides a multi-hazard cascade evolution time window intelligent prediction system, the system comprising: The data acquisition module is used to collect multi-source historical disaster datasets, perform spatiotemporal alignment and multi-hazard cascade analysis on the datasets, and generate a multi-hazard cascade topology map. The training module is used to train the multi-hazard cascade topology map based on the multi-source historical disaster datasets to predict its evolutionary trends, and construct a multi-hazard cascade evolution prediction model. The deduction and prediction module is used to collect real-time multi-source disaster data, and use the multi-hazard cascade evolution prediction model to deduce and predict the real-time multi-source disaster data, obtaining the multi-hazard cascade evolution path and the corresponding multi-hazard evolution prediction time window. The closed-loop prediction optimization module is used to optimize the multi-hazard cascade evolution prediction model based on the multi-hazard cascade evolution path and the corresponding multi-hazard evolution prediction time window.

[0007] One or more technical solutions provided in this application have at least the following technical effects or advantages: This application collects multi-source historical disaster datasets, performs spatiotemporal alignment and multi-hazard cascade analysis on these datasets, and generates a multi-hazard cascade topology map. Based on the multi-source historical disaster datasets, it trains the multi-hazard cascade topology map to predict its evolutionary trends, constructing a multi-hazard cascade evolution prediction model. It also collects real-time multi-source disaster data and uses the multi-hazard cascade evolution prediction model to extrapolate and predict the real-time multi-source disaster data, obtaining the multi-hazard cascade evolution path and corresponding multi-hazard evolution prediction time windows. Based on the multi-hazard cascade evolution path and corresponding multi-hazard evolution prediction time windows, it optimizes the multi-hazard cascade evolution prediction model through closed-loop prediction. This invention solves the technical problem in the prior art of effectively modeling the multi-hazard cascade evolution process and achieving accurate time window prediction. By constructing a multi-hazard cascade topology map and performing evolutionary trend modeling and prediction, it achieves the technical effect of improving the prediction accuracy of the multi-hazard cascade evolution process and the precision of time window determination. Attached Figure Description

[0008] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0009] Figure 1 A schematic diagram of a multi-hazard cascade evolution time window intelligent prediction method provided in this application embodiment; Figure 2 This is a schematic diagram of the structure of a multi-hazard cascade evolution time window intelligent prediction system provided in an embodiment of this application.

[0010] Figure labeling: Data acquisition module 11, training module 12, inference and prediction module 13, closed-loop prediction and optimization module 14. Detailed Implementation

[0011] This application provides an intelligent prediction method and system for the time window of multi-hazard cascade evolution, which addresses the technical problem of difficulty in effectively modeling the multi-hazard cascade evolution process and achieving accurate time window prediction in the existing technology. By constructing a multi-hazard cascade topology diagram and performing evolutionary situation modeling and prediction, the technical effect of improving the prediction accuracy of the multi-hazard cascade evolution process and the precision of time window determination is achieved.

[0012] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0013] It should be noted that any variation of the terms "comprising" and "having" is intended to cover non-exclusive inclusion, for example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or modules that are not explicitly listed or that are inherent to such processes, methods, products, or devices.

[0014] Example 1, as Figure 1 As shown, this application provides an intelligent prediction method for multi-hazard cascade evolution time windows, the method comprising: Step S100: Collect multi-source historical disaster datasets, perform spatiotemporal alignment and multi-hazard cascade analysis on the multi-source historical disaster datasets, and generate a multi-hazard cascade topology map.

[0015] In this embodiment, firstly, a multi-source historical disaster dataset is extracted from a preset database, and then cleaned and standardized according to data application standards to form a usable historical disaster dataset. Next, the usable historical disaster dataset is spatiotemporally aligned based on preset time and spatial granularity, and then reconstructed temporally according to disaster type to generate a historical disaster spatiotemporal data sequence set. Subsequently, based on the historical disaster spatiotemporal data sequence set, temporal correlation analysis and cascading relationship modeling are carried out between disaster types to construct a multi-disaster cascading topology graph with disaster type as nodes and disaster cascading relationship as directed edges.

[0016] Furthermore, the method provided in the application embodiments, which performs spatiotemporal alignment and multi-hazard cascade analysis on the multi-source historical disaster dataset to generate a multi-hazard cascade topology map, further includes: The multi-source historical disaster dataset is cleaned and standardized according to data application standards to obtain a usable historical disaster dataset. The usable historical disaster dataset is then aligned according to preset time and spatial granularity to obtain a standard historical disaster dataset. The standard historical disaster dataset is then arranged chronologically according to disaster type to generate a historical disaster spatiotemporal data sequence set. Based on the historical disaster spatiotemporal data sequence set, a multi-disaster cascade analysis is performed to generate a multi-disaster cascade topology map.

[0017] In this embodiment, the multi-source historical disaster dataset is first cleaned and standardized according to data application standards to establish a standard data field table. This table includes at least the following fields: disaster event identifier, data source identifier, disaster type, disaster occurrence time, longitude, latitude, spatial unit identifier, and disaster intensity. Each historical disaster record in the multi-source historical disaster dataset is read sequentially, and the original field names from different data sources are mapped according to the standard data field table. The disaster occurrence time is uniformly converted to a "year-month-day hour:minute:second" time format, and the longitude and latitude are uniformly converted to decimal coordinates in the same coordinate system. The disaster type is converted to a disaster type code according to a preset disaster type code table. A duplicate judgment field combination is constructed using the disaster type code, disaster occurrence time, longitude, latitude, and data source identifier. Historical disaster records with identical field combinations are retained as one record. Historical disaster records with empty disaster occurrence time, disaster type code, longitude, or latitude are written to an invalid data table. Finally, the historical disaster records after field mapping, format conversion, disaster type code conversion, and duplicate record processing are summarized to obtain a usable historical disaster dataset. Next, the available historical disaster dataset is aligned according to a preset time granularity and spatial granularity. The preset time granularity is set to 1 hour, and the preset spatial granularity is set to a 1 km × 1 km spatial grid. The disaster occurrence time of each historical disaster record in the available historical disaster dataset is read, and the disaster occurrence time is assigned to the corresponding time unit according to the 1-hour time granularity, and a time unit identifier is written for the historical disaster record. The longitude and latitude of the historical disaster record are read, and the 1 km × 1 km spatial grid that the longitude and latitude fall into is determined as the spatial unit, and a spatial unit identifier is written for the historical disaster record. The historical disaster records with time unit identifiers and spatial unit identifiers are merged according to the time unit identifier, spatial unit identifier, and disaster type code to form a standard historical disaster record expressed under a unified time unit and a unified spatial unit. All standard historical disaster records are summarized to obtain a standard historical disaster dataset. Subsequently, the standard historical disaster dataset is arranged chronologically according to disaster type. The disaster type code of each standard historical disaster record in the dataset is read, and disaster grouping is established according to the disaster type code. Within each disaster group, the standard historical disaster records are sorted from earliest to latest according to the time unit identifier. When standard historical disaster records with the same time unit identifier exist within the same disaster group, they are sorted in a secondary order according to the spatial unit identifier number. Each sorted disaster group record is written into the spatiotemporal data sequence of the corresponding disaster type. Each sequence element in the spatiotemporal data sequence includes a time unit identifier, a spatial unit identifier, a disaster type code, and a disaster intensity field. The spatiotemporal data sequences corresponding to each disaster type are collected and stored according to the disaster type code to generate a historical disaster spatiotemporal data sequence set.

[0018] Finally, a multi-hazard cascade analysis is performed based on the historical disaster spatiotemporal data sequence set. In this process, the spatiotemporal data sequences corresponding to each disaster type in the historical disaster spatiotemporal data sequence set are used as the analysis objects. Temporal correlation analysis is performed between different disaster types to obtain a set of disaster type temporal correlation coefficients used to characterize the degree of temporal correlation between different disaster types. Then, the disaster type correlation relationships in the historical disaster spatiotemporal data sequence set are identified and screened based on the disaster type temporal correlation coefficient set, and the disaster type correlation relationships that meet the saliency screening criteria constitute a candidate set of disaster type cascade relationships. Subsequently, multi-hazard cascade topology analysis is performed using disaster type as nodes and the candidate set of disaster type cascade relationships as directed edges, so that the cascade relationships between each disaster type form a directed graph structure, generating a multi-hazard cascade topology graph.

[0019] Furthermore, the method provided in the application embodiment, which performs multi-hazard cascade analysis based on the historical disaster spatiotemporal data sequence set to generate a multi-hazard cascade topology map, further includes: Based on the historical disaster spatiotemporal data sequence set, temporal correlation analysis is performed between various disaster types to obtain a disaster type temporal correlation coefficient set; the historical disaster type spatiotemporal data sequence set is saliency-based and screened according to the disaster type temporal correlation coefficient set to form a disaster type cascade relationship candidate set; multi-disaster type cascade topology analysis is performed with disaster type as node and the disaster type cascade relationship candidate set as directed edge to generate a multi-disaster cascade topology graph.

[0020] In this embodiment, when performing time-series correlation analysis between different disaster types based on historical disaster spatiotemporal data sequence sets, the spatiotemporal data sequences corresponding to each disaster type in the historical disaster spatiotemporal data sequence set are read. The spatiotemporal data sequences include time unit identifiers, spatial unit identifiers, disaster type codes, and disaster intensity fields. Two different disaster types are used to form disaster type pairs. The disaster type with the earlier time sequence is recorded as the preceding disaster type, and the disaster type with the later time sequence is recorded as the following disaster type. Using a preset time lag window statistical method, under the same spatial unit identifier, the number of times the disaster event corresponding to the following disaster type occurs within the preset time lag window after the disaster event corresponding to the preceding disaster type occurs is counted. The ratio of this number to the total number of disaster events corresponding to the preceding disaster type is calculated to obtain the disaster time-series correlation coefficient corresponding to the disaster type pair. The above statistical calculation is performed sequentially on all disaster type pairs, and each disaster type pair and its corresponding disaster time-series correlation coefficient are collected and recorded to obtain a disaster time-series correlation coefficient set.

[0021] Next, the historical disaster spatiotemporal data sequence set is saliency-based by using the disaster type temporal correlation coefficient set. This involves reading each disaster type pair, preceding disaster type, subsequent disaster type, and disaster type temporal correlation coefficient from the set. A fixed threshold filtering method is used to compare each disaster type temporal correlation coefficient with a preset saliency threshold. When the disaster type temporal correlation coefficient is greater than or equal to the preset saliency threshold, the disaster type pair is identified as a significant disaster type correlation, and written into the filtering results according to the field structure of preceding disaster type, subsequent disaster type, and disaster type temporal correlation coefficient. When the disaster type temporal correlation coefficient is less than the preset saliency threshold, writing to that disaster type pair is stopped. All significant disaster type correlations written into the filtering results are then aggregated and stored to form a candidate set of disaster type cascading relationships.

[0022] Finally, using disaster type as nodes and the candidate set of disaster type cascading relationships as directed edges, a multi-hazard cascading topology analysis is performed to generate a multi-hazard cascading topology graph. In this process, the significant disaster type associations in the candidate set of disaster type cascading relationships are read, and the preceding and subsequent disaster types contained therein are extracted. Each disaster type is established as a topology node. Using a directed weighted graph construction method, the topology node corresponding to the preceding disaster type is set as the starting point of the directed edge, and the topology node corresponding to the subsequent disaster type is set as the ending point of the directed edge. The temporal correlation coefficient of the corresponding disaster type is written into the edge weight field of the directed edge. Directed edge construction is performed sequentially on all significant disaster type associations in the candidate set of disaster type cascading relationships to obtain a multi-hazard cascading topology structure composed of topology nodes, directed edges, and edge weights. The multi-hazard cascading topology structure is stored according to the node set, edge set, and edge weight set to generate a multi-hazard cascading topology graph.

[0023] Step S200: Based on the multi-source historical disaster dataset, train the evolution trend prediction of the multi-hazard cascade topology map to construct a multi-hazard cascade evolution prediction model.

[0024] In this embodiment, when training the evolutionary trend prediction of a multi-hazard cascade topology based on a multi-source historical disaster dataset, the disaster node relationships and directed cascade relationships represented by the multi-hazard cascade topology are used as the basis. Combined with information on disaster occurrence time, disaster type, spatial location, and disaster intensity from the multi-source historical disaster dataset, disaster chain samples are extracted from historical disaster events that conform to the cascade topology, forming a disaster chain event sequence sample set arranged in the order of disaster occurrence. Subsequently, the disaster chain event sequence sample set is labeled as positive and negative. Disaster chain event sequences that satisfy the disaster cascade relationship and temporal sequence are labeled as positive disaster chain event samples, while those that do not satisfy the disaster cascade relationship and temporal sequence are labeled as negative disaster chain event samples. Finally, evolutionary trend prediction training is performed based on the positive and negative disaster chain event sample sets, enabling the model to learn the cascade path features and temporal evolution features in the disaster chain event sequences, thus constructing a multi-hazard cascade evolution prediction model.

[0025] Furthermore, in the method provided in the application embodiments, the method of training the evolutionary trend prediction of the multi-hazard cascade topology based on the multi-source historical disaster dataset to construct a multi-hazard cascade evolution prediction model further includes: Based on the multi-source historical disaster dataset, disaster chain samples are extracted from the multi-hazard cascade topology to obtain a disaster chain event sequence sample set; the disaster chain event sequence sample set is labeled with positive and negative samples to obtain a disaster chain event positive sample set and a disaster chain event negative sample set; based on the disaster chain event positive sample set and the disaster chain event negative sample set, evolution trend prediction training is performed to construct a multi-hazard cascade evolution prediction model.

[0026] In this embodiment, when extracting disaster chain samples from a multi-hazard cascaded topology graph based on a multi-source historical disaster dataset, the node set, edge set, and edge weight set in the multi-hazard cascaded topology graph are read. Each node in the node set corresponds to a disaster type, each directed edge in the edge set corresponds to a cascade relationship between a preceding disaster type and a subsequent disaster type, and the edge weight set records the disaster time-series correlation coefficients corresponding to each cascade relationship. A topology path sliding window extraction method is used to extract disaster type paths from the multi-hazard cascaded topology graph according to the directed edge connection order, and each disaster type path is used as a template for disaster chain sample extraction. Disaster events contained in the multi-source historical disaster dataset are read. Record the disaster events and arrange them according to spatial unit identifier and disaster occurrence time. Under the same spatial unit identifier, take the first disaster type in the disaster chain sample extraction template as the starting matching node, and match historical disaster event records item by item along the disaster type order in the disaster chain sample extraction template. When the historical disaster event records corresponding to two adjacent disaster types meet the directed edge connection relationship and the preset time interval condition, write the group of historical disaster event records into a disaster chain event sequence sample according to the disaster occurrence time order. Repeat the above extraction process for all disaster type paths in the multi-disaster cascade topology graph, and store the obtained multiple disaster chain event sequence samples in a set to obtain a disaster chain event sequence sample set.

[0027] When labeling a disaster chain event sequence sample set as positive or negative, each disaster chain event sequence sample in the set is read. Each disaster chain event sequence sample includes a disaster event identifier, disaster type, disaster occurrence time, spatial unit identifier, disaster intensity field, and time interval field between adjacent disaster events. A labeling rule method is used to evaluate each disaster chain event sequence sample item by item. If all adjacent disaster events in a disaster chain event sequence sample satisfy the directed edge connection relationship in the multi-hazard cascade topology graph, and the time interval between adjacent disaster events falls within a preset time interval range, then the disaster chain event is considered a positive or negative sample. A sequence sample is labeled with a positive sample to obtain a positive sample of a disaster chain event. When there are adjacent disaster events in the disaster chain event sequence sample that do not satisfy the directed edge connection relationship in the multi-hazard cascade topology graph, and the time interval between adjacent disaster events does not fall within the preset time interval range, the disaster chain event sequence sample is labeled with a negative sample to obtain a negative sample of a disaster chain event. After labeling all disaster chain event sequence samples in the disaster chain event sequence sample set, the disaster chain event sequence samples with positive sample labels are summarized into a positive sample set of disaster chain events, and the disaster chain event sequence samples with negative sample labels are summarized into a negative sample set of disaster chain events.

[0028] Finally, evolutionary trend prediction training is conducted based on the positive and negative sample sets of disaster chain events. In this process, a graph attention network is used to train feature encoding on the positive and negative sample sets of disaster chain events, extracting disaster type features, cascading relationship features, temporal interval features, and spatial unit features from the disaster chain event sequence, and generating disaster path prediction models and disaster time prediction models respectively. The disaster path prediction model and the disaster time prediction model are then jointly trained, allowing the path prediction task and the time prediction task to update parameters during the same training process, resulting in an initial disaster evolution prediction model. The initial disaster evolution prediction model is then validated and its parameters are tuned. Based on the validation results, the model parameters are adjusted to construct a multi-hazard cascading evolution prediction model for outputting multi-hazard cascading evolution paths and corresponding evolution prediction time windows.

[0029] Furthermore, in the method provided in the application embodiments, the evolutionary situation prediction training based on the positive sample set and negative sample set of disaster chain events to construct a multi-hazard cascade evolution prediction model further includes: A graph attention network is used to train the positive and negative sample sets of disaster chain events by feature encoding, generating a disaster path prediction model and a disaster time prediction model. The disaster path prediction model and the disaster time prediction model are jointly trained to obtain an initial disaster evolution prediction model. The performance of the initial disaster evolution prediction model is verified and the parameters are tuned to construct a multi-hazard cascade evolution prediction model.

[0030] In this embodiment, when training the feature encoding of the positive and negative sample sets of disaster chain events using a graph attention network, disaster chain event sequence samples from the positive and negative sample sets are read. Each disaster chain event sequence sample includes a disaster type code, disaster occurrence time, spatial unit identifier, disaster intensity field, time interval between adjacent disaster events, and a sample label. Each disaster event in the same disaster chain event sequence sample is set as a graph node, and the cascading relationship from one disaster event to the next is set as a directed edge. The disaster type code, relative time value, spatial unit identifier, and... The disaster intensity field is arranged in a fixed order to form the node feature vector of the graph node, where the relative time value is the time difference between the occurrence time of the current disaster event and the occurrence time of the first disaster event in the disaster chain event sequence sample. At the beginning of training, the graph attention network parameters, path classification layer parameters, and time regression layer parameters are initialized, and a number of disaster chain event sequence samples are read from the positive and negative disaster chain event sets according to a preset batch size as a training batch. The node feature vectors in the training batch are input into the graph attention network. The graph attention network first multiplies the node feature vectors with the weight matrix to obtain the intermediate features of the nodes, and then multiplies each node feature vector. The intermediate features of the nodes at both ends of a directed edge are concatenated, and the concatenation result is multiplied by the attention parameter to obtain the attention score corresponding to the directed edge. All attention scores connected to the same graph node are normalized until the sum of the normalized attention scores is one, resulting in the attention weight. The intermediate features of adjacent graph nodes are weighted and summed according to the attention weight to obtain the node encoding result. The average of all node encoding results in the same disaster chain event sequence sample is taken to obtain the disaster chain feature encoding of the disaster chain event sequence sample. The disaster chain feature encoding is input into the path classification layer, which outputs the probability of path establishment and sets the path... The probability of success and the sample labels are used to calculate the binary cross-entropy to obtain the path prediction loss. The parameters of the graph attention network and the path classification layer are updated based on the path prediction loss to train the disaster path prediction model. Disaster chain event sequence samples with effective adjacent disaster event time interval fields are read from the positive sample set of disaster chain events. The corresponding disaster chain feature encodings are input into the time regression layer. The time regression layer outputs the prediction time interval. The mean square error of the prediction time interval and the adjacent disaster event time interval fields is calculated to obtain the time prediction loss. The parameters of the graph attention network and the time regression layer are updated based on the time prediction loss to train the disaster time prediction model.

[0031] Next, the disaster path prediction model and the disaster time prediction model are jointly trained. In this process, a path prediction loss function is constructed based on the path prediction results of the disaster path prediction model, and a time window loss function is constructed based on the time window prediction results of the disaster time prediction model. These two loss functions together form the multi-task total loss function. A training gradient balancing strategy is constructed according to the model training task objectives to balance the gradient updates of the path prediction task and the time window prediction task during joint training. Based on the multi-task total loss function and the training gradient balancing strategy, the disaster path prediction model and the disaster time prediction model are trained synchronously, enabling the path prediction branch and the time prediction branch to complete joint optimization within the same training framework, resulting in the initial disaster evolution prediction model.

[0032] Finally, the initial disaster evolution prediction model was validated and its parameters were tuned to construct a multi-hazard cascade evolution prediction model. In this process, the positive and negative sample sets of disaster chain events were divided into training and validation sample sets according to a fixed ratio. The initial disaster evolution prediction model was trained using the training sample set, and then the validation sample set was input into the trained model to obtain the path prediction result and time window prediction result for each validation sample. The path prediction result was compared item by item with the sample labels in the validation samples. When the path prediction result matched the sample label, it was recorded as a correct prediction sample. The number of correct prediction samples was divided by the total number of validation samples to obtain the path prediction accuracy. Disaster chain event sequence samples with positive sample labels were read from the validation sample set, and their time window prediction results were compared with the actual time windows of adjacent disaster events. The difference is calculated at intervals, and the absolute values ​​of all differences are averaged to obtain the time window prediction error. The path prediction accuracy and the time window prediction error are used as the model prediction performance parameters. The learning rate, the dimension of the hidden layer of the graph attention network, and the number of training rounds of the initial disaster evolution prediction model are tuned using a grid search method. The model parameters are set one group at a time according to the preset parameter value table, and training and validation are re-executed under each group of model parameters. The corresponding path prediction accuracy and time window prediction error are recorded. The parameter combination that achieves the preset accuracy threshold for path prediction and the preset error threshold for time window prediction is selected as the target model parameters. The target model parameters are written into the initial disaster evolution prediction model to obtain the multi-hazard cascade evolution prediction model.

[0033] Furthermore, in the method provided in the application embodiments, the method of jointly training the disaster path prediction model and the disaster time prediction model to obtain an initial disaster evolution prediction model further includes: A multi-task total loss function is constructed, which includes a path prediction loss function and a time window loss function; a training gradient balancing strategy is constructed according to the model training task objective; the disaster path prediction model and the disaster time prediction model are jointly trained based on the multi-task total loss function and the training gradient balancing strategy to obtain an initial disaster evolution prediction model.

[0034] In this embodiment, a multi-task total loss function is first constructed, which includes a path prediction loss function and a time window loss function. Specifically, disaster chain event sequence samples from both the positive and negative disaster chain event sample sets are input into the disaster path prediction model to obtain the probability of path success for each disaster chain event sequence sample. The sample label of the positive disaster chain event sample set is set to 1, and the sample label of the negative disaster chain event sample set is set to 0. The path prediction loss function is calculated using the binary cross-entropy function, expressed as Lp=-1 / N×Σ[yi×log(pi)+(1-yi)×log(1-pi)], where Lp is the path prediction loss function, N is the number of disaster chain event sequence samples participating in path prediction training, yi is the sample label of the i-th disaster chain event sequence sample, and pi is the probability of path success for the i-th disaster chain event sequence sample. The disaster chain event sequence samples from the positive disaster chain event sample set are input into the disaster time prediction model to obtain the prediction time interval for each disaster chain event sequence sample. The corresponding actual adjacent disaster event time interval field is read, and the time window loss function is calculated using the mean squared error function, expressed as Lt=1 / M×Σ. Where Lt is the time window loss function, and M is the number of positive samples of disaster chain events participating in the time window training. Let ti be the predicted time interval corresponding to the positive sample of the i-th disaster chain event, and ti be the actual time interval between adjacent disaster events corresponding to the positive sample of the i-th disaster chain event. The path prediction loss function and the time window loss function are weighted and summed to obtain the multi-task total loss function, which is expressed as L = λp × Lp + λt × Lt, where L is the multi-task total loss function, λp is the weight of the path prediction loss function, and λt is the weight of the time window loss function. Next, based on the model training task objective, a training gradient balancing strategy is constructed. In each training batch, the gradient norm of the path prediction loss function Lp relative to the shared graph attention network parameters is first calculated, denoted as Gp. Then, the gradient norm of the time window loss function Lt relative to the shared graph attention network parameters is calculated, denoted as Gt. The gradient norm is calculated using the L2 norm. Gp represents the update magnitude of the shared graph attention network parameters by the path prediction task, and Gt represents the update magnitude of the shared graph attention network parameters by the time window prediction task. Based on Gp and Gt, the weights λp and λt of the path prediction loss function and the time window loss function are calculated, where λp = Gt / , λt=Gp / ), where ε is a preset minimum constant; when the gradient norm Gp of the path prediction task is large, λp decreases, and the proportion of the path prediction loss function in the total loss function of the multi-task task decreases; when the gradient norm Gt of the time window prediction task is large, λt decreases, and the proportion of the time window loss function in the total loss function of the multi-task task decreases; the total loss function of the multi-task task L is recalculated according to the updated λp and λt, so that the path prediction task and the time window prediction task participate in parameter updates according to the balanced gradient in the same training batch.

[0035] Finally, the disaster path prediction model and the disaster time prediction model are jointly trained based on the multi-task total loss function and the training gradient balancing strategy. In this process, the graph attention networks in the disaster path prediction model and the disaster time prediction model are merged into a shared graph attention network. The path classification layer in the disaster path prediction model is set as the path prediction branch, and the time regression layer in the disaster time prediction model is set as the time prediction branch. Disaster chain event sequence samples from the positive and negative sample sets of disaster chain events are read according to a preset batch size and input into the shared graph attention network to obtain disaster chain feature codes. The disaster chain feature codes are input into the path prediction branch to obtain the path establishment probability, and the path prediction loss function Lp is calculated based on the path establishment probability and sample labels. The disaster chain feature codes corresponding to the positive sample set of disaster chain events are input into the time prediction branch to obtain the prediction time interval, and the prediction time interval and actual adjacent disaster events are calculated accordingly. The time window loss function Lt is calculated at time intervals; the gradient norms Gp and Gt are calculated based on the path prediction loss function Lp and the time window loss function Lt, and then λp and λt are calculated according to the training gradient balancing strategy. The multi-task total loss function is obtained by L=λp×Lp+λt×Lt; backpropagation is performed based on the multi-task total loss function L to synchronously update the parameters of the shared graph attention network, path classification layer, and time regression layer; training is repeated in the order of training batch reading, disaster chain feature encoding calculation, path prediction loss function calculation, time window loss function calculation, training gradient balancing, multi-task total loss function calculation, and parameter update until the preset training rounds are reached. The trained shared graph attention network, path prediction branch, and time prediction branch are combined to obtain the initial disaster evolution prediction model.

[0036] Step S300: Collect real-time data of multi-source disasters, and use the multi-hazard cascade evolution prediction model to extrapolate and predict the real-time data of multi-source disasters to obtain the multi-hazard cascade evolution path and the corresponding multi-hazard evolution prediction time window.

[0037] In this embodiment, when collecting real-time data on multi-source disasters and using a multi-hazard cascade evolution prediction model to extrapolate and predict the real-time data, meteorological monitoring data, hydrological monitoring data, geological monitoring data, and disaster event reporting data are read in real time, and disaster type codes, disaster occurrence time, spatial unit identifiers, and disaster intensity fields are extracted. Disaster occurrence times are assigned to corresponding time units according to a preset time granularity, and disaster locations are assigned to corresponding spatial units according to a preset spatial granularity, resulting in real-time disaster spatiotemporal data. The real-time disaster spatiotemporal data within the same spatial unit are sorted according to disaster occurrence time to generate a real-time disaster event sequence. Each real-time disaster event in the sequence is set as a graph node, and the sequential relationship between adjacent real-time disaster events is set as a directed edge. The disaster type code, relative time value, spatial unit identifier, and disaster intensity field are combined to form a node feature vector. The node feature vector is input into the graph attention network in the multi-hazard cascade evolution prediction model for encoding to obtain real-time disaster chain feature codes.

[0038] Next, the real-time disaster chain feature encoding is input into the path prediction branch. The path prediction branch calculates the path success probability for candidate subsequent disaster types that are connected to the current disaster type by directed edges in the multi-hazard cascade topology graph, and determines the candidate subsequent disaster type with the highest path success probability that reaches the preset path probability threshold as the next disaster type. The deduction is repeated in the order from the current disaster type to the next disaster type until the path success probability is lower than the preset path probability threshold, thus obtaining the multi-hazard cascade evolution path. The real-time disaster chain feature encoding is input into the time prediction branch to obtain the prediction time interval from the current disaster type to the next disaster type. The prediction time interval is added to the prediction time interval, with the disaster occurrence time of the current disaster event as the starting time, to obtain the predicted occurrence time. According to the preset time window deviation value, the predicted occurrence time is subtracted from the preset time window deviation value to obtain the time window start time, and the predicted occurrence time is added to the preset time window deviation value to obtain the time window end time. The above calculation is repeated for each group of adjacent disaster types in the multi-hazard cascade evolution path to obtain the corresponding multi-hazard evolution prediction time window.

[0039] Step S400: Based on the multi-hazard cascade evolution path and the corresponding multi-hazard evolution prediction time window, perform closed-loop prediction optimization on the multi-hazard cascade evolution prediction model.

[0040] In this embodiment, when optimizing the multi-hazard cascade evolution prediction model based on the multi-hazard cascade evolution path and the corresponding multi-hazard evolution prediction time window, the multi-hazard cascade evolution path and the corresponding multi-hazard evolution prediction time window are compared and verified with the subsequently collected actual disaster evolution data. Based on the consistency between the path prediction results and the actual disaster chain events, as well as the deviation between the prediction time window and the actual disaster occurrence time, the model prediction performance parameters are obtained. Then, based on the model prediction performance parameters, the multi-hazard cascade evolution prediction model is optimized in a closed loop. Through model parameter adjustment, prediction error feedback, and real-time data updates, the multi-hazard cascade evolution prediction model is adaptively corrected during continuous prediction, and adaptive early warning decisions are executed based on the corrected prediction results.

[0041] Furthermore, in the method provided in the application embodiments, the closed-loop prediction optimization of the multi-hazard cascade evolution prediction model based on the multi-hazard cascade evolution path and the corresponding multi-hazard evolution prediction time window further includes: The multi-hazard cascade evolution prediction model is validated and evaluated based on the multi-hazard cascade evolution path and the corresponding multi-hazard evolution prediction time window to obtain model prediction performance parameters; based on the model prediction performance parameters, the multi-hazard cascade evolution prediction model is optimized through closed-loop prediction and adaptive early warning decision-making.

[0042] In this embodiment, when validating and evaluating the multi-hazard cascade evolution prediction model based on the multi-hazard cascade evolution path and the corresponding multi-hazard evolution prediction time window, after the multi-hazard cascade evolution prediction model outputs the multi-hazard cascade evolution path and the corresponding multi-hazard evolution prediction time window, the starting disaster event, disaster type code, spatial unit identifier, predicted subsequent disaster type, path success probability, prediction time window start time, and prediction time window end time corresponding to this prediction are recorded. After the prediction time window ends, actual disaster evolution data under the same spatial unit identifier are collected. The actual disaster evolution data includes the actual disaster type code, actual disaster occurrence time, actual spatial unit identifier, and actual disaster intensity field. The actual disaster evolution data are sorted from earliest to latest according to the actual disaster occurrence time to obtain the actual disaster chain event sequence. The predicted disaster type in the multi-hazard cascade evolution path is matched item by item with the actual disaster type in the actual disaster chain event sequence. When the predicted disaster type is consistent with the actual disaster type, and the actual disaster intensity is... When the time of a hazard occurrence falls within the corresponding multi-hazard evolution prediction time window, the prediction result is recorded as a hit sample. When the predicted hazard type is inconsistent with the actual hazard type, or the actual hazard occurrence time does not fall within the corresponding multi-hazard evolution prediction time window, the prediction result is recorded as a deviation sample. The number of hit samples, deviation samples, and the total number of prediction samples are counted. The number of hit samples divided by the total number of prediction samples yields the path-time joint hit rate. The number of samples whose predicted hazard type matches the actual hazard type is divided by the total number of prediction samples to obtain the path prediction consistency rate. The number of samples whose actual hazard occurrence time falls within the multi-hazard evolution prediction time window is divided by the total number of prediction samples to obtain the time window hit rate. The time difference between each actual hazard occurrence time and the center time of the corresponding prediction time window is calculated, and the absolute values ​​of all time differences are averaged to obtain the average time window deviation value. The path-time joint hit rate, path prediction consistency rate, time window hit rate, and average time window deviation value are summarized as model prediction performance parameters. Next, based on the model's prediction performance parameters, closed-loop prediction optimization and adaptive early warning decision-making are performed on the multi-hazard cascade evolution prediction model. In this process, the model's prediction performance parameters are read, and the path prediction consistency rate is compared with a preset path consistency rate threshold, the time window hit rate is compared with a preset time window hit rate threshold, and the average time window deviation is compared with a preset time deviation threshold. When the path prediction consistency rate is lower than the preset path consistency rate threshold, the actual hazard type order, actual hazard occurrence time, and actual spatial unit identifier in the deviation samples are written into the hazard chain event sequence sample set, and positive hazard chain event samples are regenerated according to the actual hazard type order. When the time window hit rate is lower than the preset time window hit rate threshold, and the average time window deviation is lower than the preset time window hit rate threshold, the positive time window deviation is calculated. When the deviation value exceeds the preset time deviation threshold, the actual time intervals between adjacent disaster events in the deviation samples are written into the positive sample set of disaster chain events as new training samples for the disaster time prediction model. The newly added positive samples of disaster chain events are merged with the original positive sample set of disaster chain events, and the prediction results that do not match are written into the negative sample set of disaster chain events according to the sample labeling rules. The merged positive and negative sample sets of disaster chain events are used to incrementally train the multi-hazard cascade evolution prediction model. During the incremental training process, the graph attention network structure, path classification layer structure, and time regression layer structure remain unchanged, and only the shared graph attention network is modified. The parameters, path classification layer parameters, and time regression layer parameters are updated. After the update, the new model parameters are written into the multi-hazard cascade evolution prediction model, and the updated multi-hazard cascade evolution prediction model is used to continue to extrapolate and predict the next batch of real-time multi-source disaster data, forming a closed-loop prediction optimization process consisting of prediction output, actual verification, performance parameter calculation, sample write-back, and model update. In adaptive early warning decision-making, the updated multi-hazard cascade evolution path, path establishment probability, and multi-hazard evolution prediction time window are read. When the path establishment probability is greater than or equal to the preset high-level early warning threshold, and the prediction time window start time is the same as the current... When the time interval between events is less than or equal to the preset short-term warning duration, a high-level warning decision is generated; when the probability of a path being established is greater than or equal to the preset medium-level warning threshold, and the interval between the start time of the prediction time window and the current time is less than or equal to the preset medium-term warning duration, a medium-level warning decision is generated; when the probability of a path being established is lower than the preset medium-level warning threshold, a low-level warning decision is generated; the high-level, medium-level, and low-level warning decisions are associated and stored with the corresponding disaster type, spatial unit identifier, multi-hazard cascade evolution path, and multi-hazard evolution prediction time window to complete closed-loop prediction optimization and adaptive warning decision-making.

[0043] In summary, the embodiments of this application have at least the following technical effects: This application collects multi-source historical disaster datasets, performs spatiotemporal alignment and multi-hazard cascade analysis on these datasets, and generates a multi-hazard cascade topology map. Based on the multi-source historical disaster datasets, it trains the multi-hazard cascade topology map to predict its evolutionary trends, constructing a multi-hazard cascade evolution prediction model. It also collects real-time multi-source disaster data and uses the multi-hazard cascade evolution prediction model to extrapolate and predict the real-time multi-source disaster data, obtaining the multi-hazard cascade evolution path and corresponding multi-hazard evolution prediction time windows. Based on the multi-hazard cascade evolution path and corresponding multi-hazard evolution prediction time windows, it optimizes the multi-hazard cascade evolution prediction model through closed-loop prediction. This invention solves the technical problem in the prior art of effectively modeling the multi-hazard cascade evolution process and achieving accurate time window prediction. By constructing a multi-hazard cascade topology map and performing evolutionary trend modeling and prediction, it achieves the technical effect of improving the prediction accuracy of the multi-hazard cascade evolution process and the precision of time window determination.

[0044] Example 2, based on the same inventive concept as the multi-hazard cascade evolution time window intelligent prediction method in the foregoing examples, such as... Figure 2 As shown, this application provides an intelligent prediction system for multi-hazard cascade evolution time windows. The system and method embodiments in this application are based on the same inventive concept. The system includes: The data acquisition module 11 is used to collect multi-source historical disaster datasets, perform spatiotemporal alignment and multi-hazard cascade analysis on the multi-source historical disaster datasets, and generate a multi-hazard cascade topology map; the training module 12 is used to train the multi-hazard cascade topology map based on the multi-source historical disaster datasets to predict the evolution trend, and construct a multi-hazard cascade evolution prediction model; the inference prediction module 13 is used to collect real-time multi-source disaster data, use the multi-hazard cascade evolution prediction model to infer and predict the real-time multi-source disaster data, and obtain the multi-hazard cascade evolution path and the corresponding multi-hazard evolution prediction time window; the closed-loop prediction optimization module 14 is used to perform closed-loop prediction optimization on the multi-hazard cascade evolution prediction model based on the multi-hazard cascade evolution path and the corresponding multi-hazard evolution prediction time window.

[0045] Furthermore, the system is also used to implement the following functions: The multi-source historical disaster dataset is cleaned and standardized according to data application standards to obtain a usable historical disaster dataset. The usable historical disaster dataset is then aligned according to preset time and spatial granularity to obtain a standard historical disaster dataset. The standard historical disaster dataset is then arranged chronologically according to disaster type to generate a historical disaster spatiotemporal data sequence set. Based on the historical disaster spatiotemporal data sequence set, a multi-disaster cascade analysis is performed to generate a multi-disaster cascade topology map.

[0046] Furthermore, the system is also used to implement the following functions: Based on the historical disaster spatiotemporal data sequence set, temporal correlation analysis is performed between various disaster types to obtain a disaster type temporal correlation coefficient set; the historical disaster type spatiotemporal data sequence set is saliency-based and screened according to the disaster type temporal correlation coefficient set to form a disaster type cascade relationship candidate set; multi-disaster type cascade topology analysis is performed with disaster type as node and the disaster type cascade relationship candidate set as directed edge to generate a multi-disaster cascade topology graph.

[0047] Furthermore, the system is also used to implement the following functions: Based on the multi-source historical disaster dataset, disaster chain samples are extracted from the multi-hazard cascade topology to obtain a disaster chain event sequence sample set; the disaster chain event sequence sample set is labeled with positive and negative samples to obtain a disaster chain event positive sample set and a disaster chain event negative sample set; based on the disaster chain event positive sample set and the disaster chain event negative sample set, evolution trend prediction training is performed to construct a multi-hazard cascade evolution prediction model.

[0048] Furthermore, the system is also used to implement the following functions: A graph attention network is used to train the positive and negative sample sets of disaster chain events by feature encoding, generating a disaster path prediction model and a disaster time prediction model. The disaster path prediction model and the disaster time prediction model are jointly trained to obtain an initial disaster evolution prediction model. The performance of the initial disaster evolution prediction model is verified and the parameters are tuned to construct a multi-hazard cascade evolution prediction model.

[0049] Furthermore, the system is also used to implement the following functions: A multi-task total loss function is constructed, which includes a path prediction loss function and a time window loss function; a training gradient balancing strategy is constructed according to the model training task objective; the disaster path prediction model and the disaster time prediction model are jointly trained based on the multi-task total loss function and the training gradient balancing strategy to obtain an initial disaster evolution prediction model.

[0050] Furthermore, the system is also used to implement the following functions: The multi-hazard cascade evolution prediction model is validated and evaluated based on the multi-hazard cascade evolution path and the corresponding multi-hazard evolution prediction time window to obtain model prediction performance parameters; based on the model prediction performance parameters, the multi-hazard cascade evolution prediction model is optimized through closed-loop prediction and adaptive early warning decision-making.

[0051] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0052] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A smart prediction method for multi-hazard cascade evolution time windows, characterized in that, The method includes: Collect multi-source historical disaster datasets, perform spatiotemporal alignment and multi-hazard cascade analysis on the multi-source historical disaster datasets, and generate a multi-hazard cascade topology map; Based on the multi-source historical disaster dataset, the evolution trend prediction training of the multi-hazard cascade topology map is carried out to construct a multi-hazard cascade evolution prediction model; Collect real-time data of multi-source disasters, and use the multi-hazard cascade evolution prediction model to extrapolate and predict the real-time data of multi-source disasters, so as to obtain the multi-hazard cascade evolution path and the corresponding multi-hazard evolution prediction time window; Based on the multi-hazard cascade evolution path and the corresponding multi-hazard evolution prediction time window, the multi-hazard cascade evolution prediction model is optimized through closed-loop prediction.

2. The intelligent prediction method for multi-hazard cascade evolution time window as described in claim 1, characterized in that, The multi-source historical disaster dataset is subjected to spatiotemporal alignment and multi-hazard cascade analysis to generate a multi-hazard cascade topology map, including: The multi-source historical disaster dataset was cleaned and standardized according to data application standards to obtain a usable historical disaster dataset. The available historical disaster dataset is aligned according to a preset time granularity and spatial granularity to obtain a standard historical disaster dataset; The standard historical disaster dataset is arranged chronologically according to disaster type to generate a historical disaster spatiotemporal data sequence set; Based on the historical disaster spatiotemporal data sequence set, a multi-hazard cascade analysis is performed to generate a multi-hazard cascade topology map.

3. The intelligent prediction method for multi-hazard cascade evolution time window as described in claim 2, characterized in that, Based on the aforementioned historical disaster spatiotemporal data sequence set, a multi-hazard cascade analysis is performed to generate a multi-hazard cascade topology map, including: Based on the historical disaster spatiotemporal data sequence set, conduct temporal correlation analysis between various disaster types to obtain a set of disaster type temporal correlation coefficients. The historical disaster spatiotemporal data sequence set is saliency-based and filtered according to the disaster type temporal correlation coefficient set to form a candidate set of disaster type cascade relationships; Using disaster type as nodes and the candidate set of disaster type cascade relationships as directed edges, multi-disaster cascade topology analysis is performed to generate a multi-disaster cascade topology graph.

4. The intelligent prediction method for multi-hazard cascade evolution time window as described in claim 1, characterized in that, Based on the aforementioned multi-source historical disaster dataset, the evolutionary trend prediction training of the multi-hazard cascade topology map is performed to construct a multi-hazard cascade evolution prediction model, including: Based on the multi-source historical disaster dataset, disaster chain samples are extracted from the multi-hazard cascade topology graph to obtain a disaster chain event sequence sample set. The disaster chain event sequence sample set is labeled with positive and negative samples to obtain a disaster chain event positive sample set and a disaster chain event negative sample set; Based on the positive and negative sample sets of disaster chain events, an evolutionary trend prediction training is conducted to construct a multi-hazard cascade evolution prediction model.

5. The intelligent prediction method for multi-hazard cascade evolution time window as described in claim 4, characterized in that, Based on the positive and negative sample sets of disaster chain events, an evolutionary trend prediction training is conducted to construct a multi-hazard cascade evolution prediction model, including: A graph attention network is used to train the feature encoding on the positive sample set and the negative sample set of disaster chain events to generate a disaster path prediction model and a disaster time prediction model. The disaster path prediction model and the disaster time prediction model are jointly trained to obtain an initial disaster evolution prediction model; The initial disaster evolution prediction model was validated and its parameters were optimized to construct a multi-hazard cascade evolution prediction model.

6. The intelligent prediction method for multi-hazard cascade evolution time window as described in claim 5, characterized in that, The disaster path prediction model and the disaster time prediction model are jointly trained to obtain an initial disaster evolution prediction model, including: Construct a multi-task total loss function, which includes a path prediction loss function and a time window loss function; Based on the model training task objective, construct a training gradient balancing strategy; The disaster path prediction model and the disaster time prediction model are jointly trained based on the multi-task total loss function and the training gradient balancing strategy to obtain the initial disaster evolution prediction model.

7. The intelligent prediction method for multi-hazard cascade evolution time window as described in claim 1, characterized in that, Based on the multi-hazard cascade evolution path and the corresponding multi-hazard evolution prediction time window, the multi-hazard cascade evolution prediction model is optimized through closed-loop prediction, including: The multi-hazard cascade evolution prediction model is validated and evaluated based on the multi-hazard cascade evolution path and the corresponding multi-hazard evolution prediction time window to obtain the model prediction performance parameters. Based on the model's prediction performance parameters, the multi-hazard cascade evolution prediction model is subjected to closed-loop prediction optimization and adaptive early warning decision-making.

8. A multi-hazard cascade evolution time window intelligent prediction system, characterized in that, The system is used to execute the intelligent prediction method for multi-hazard cascade evolution time window as described in any one of claims 1-7, and the system includes: The data acquisition module is used to collect multi-source historical disaster datasets, perform spatiotemporal alignment and multi-hazard cascade analysis on the multi-source historical disaster datasets, and generate a multi-hazard cascade topology map; The training module is used to train the evolution trend prediction of the multi-hazard cascade topology based on the multi-source historical disaster dataset, and to construct a multi-hazard cascade evolution prediction model. The extrapolation and prediction module is used to collect real-time data of multi-source disasters, and to extrapolate and predict the real-time data of multi-source disasters using the multi-hazard cascade evolution prediction model to obtain the multi-hazard cascade evolution path and the corresponding multi-hazard evolution prediction time window. The closed-loop prediction optimization module is used to perform closed-loop prediction optimization on the multi-hazard cascade evolution prediction model based on the multi-hazard cascade evolution path and the corresponding multi-hazard evolution prediction time window.

Images