Multi-agent collaborative cross-platform flood emergency decision support system and method

Abstract

The application discloses a multi-agent cooperative cross-platform flood emergency decision support system and method, comprising a data acquisition and processing module, which is used for collecting multi-source sensor data and pre-processing; a terrain confluence construction module, which is used for constructing a directed terrain confluence graph and generating a weighted adjacency matrix; a prediction processing module, which is used for establishing a terrain confluence graph neural ODE model, outputting a short-term prediction sequence and a long-term prediction sequence; a consistency correction module, which is used for consistency discrimination, outputting a prediction result and a dangerous area map; a causal relationship construction module, which is used for forming an event-resource causal relationship graph; a causal game scheduling module, which is used for constructing a game problem and generating an emergency resource scheduling scheme; and a result issuing and displaying module, which is used for issuing the emergency resource scheduling scheme to an emergency response terminal. Through terrain confluence graph neural ODE and causal game equilibrium, the application realizes accurate flood prediction and efficient emergency resource scheduling.

Description

Technical Field

[0001] The present invention relates to the technical field of flood prevention, disaster reduction and emergency management, and particularly to a cross-platform flood emergency decision-making support system and method for multi-agent collaboration. Background Art

[0002] As one of the most major natural disasters, floods often result in casualties, infrastructure damage and economic losses. Existing flood emergency decision-making systems mainly rely on data from hydrological monitoring stations and meteorological prediction models, and predict water level changes and risk area distributions through static calculations or numerical simulations. Traditional methods usually can only provide overall trend information, and it is difficult to finely depict the confluence characteristics under local complex terrain. Traditional methods mostly center around a single platform or single department, lacking real-time linkage capabilities across regions and platforms, and there are problems of response lag and information islands in sudden flood events.

[0003] With the development of artificial intelligence and big data technologies, some research has attempted to introduce machine learning models or graph neural networks into flood prediction and emergency dispatching to improve prediction accuracy and decision-making efficiency. However, there are still limitations in the design of existing model structures. For example, graph neural networks often use a fixed adjacency matrix when dealing with the relationship between terrain and water flow, and cannot flexibly reflect the dynamic effects of terrain slope and confluence direction, easily resulting in inconsistent prediction results with actual water flow evolution. Physical constraints such as water quantity conservation and terrain monotonicity have not been effectively embedded in the model, leading to deviations in prediction results in the long-term trend.

[0004] In terms of emergency resource dispatching, existing research mostly based on heuristic rules or static optimization algorithms, and it is difficult to handle the causal coupling relationship between multi-agents, often ignoring the differences in time, space and capabilities of different resources, resulting in unbalanced resource allocation or rescue delays. Most existing balancing algorithms do not consider the constraint conditions of causal relationships and lack a priority response mechanism for high-risk areas, making it difficult to achieve truly efficient multi-agent collaborative dispatching in complex flood scenarios.

[0005] Therefore, how to provide a cross-platform flood emergency decision-making support system and method for multi-agent collaboration is an urgent problem to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a multi-agent collaborative cross-platform flood emergency decision support system and method. This invention fully utilizes the topographic confluence map neural ODE model, cross-temporal consistency correction mechanism, and distributed causal constraint Nash equilibrium algorithm. It details the complete process from multi-source sensor data acquisition and preprocessing, topographic confluence map construction, short-term and long-term prediction result generation, cross-temporal consistency correction, to event-resource causal relationship modeling and multi-agent game scheduling. This invention possesses advantages such as strong physical consistency of prediction results, accurate identification of hazardous areas, reasonable emergency resource scheduling, and high cross-platform collaborative efficiency.

[0007] A multi-agent collaborative cross-platform flood emergency decision support system according to an embodiment of the present invention includes the following modules:

[0008] The data acquisition and processing module is used to acquire multi-source sensor data, preprocess it, and generate input datasets.

[0009] The terrain confluence construction module is used to calculate the slope and confluence direction, construct a directed terrain confluence map, and generate a weighted adjacency matrix.

[0010] The prediction processing module is used to build a topographic confluence map neural ODE model and output short-term and long-term prediction sequences.

[0011] The consistency correction module is used to determine the consistency between short-term and long-term prediction sequences and output the prediction results and the danger area map.

[0012] The causal relationship construction module is used to generate an event-resource causal relationship diagram based on the danger zone map;

[0013] The causal game scheduling module is used to construct game problems based on event-resource causal relationship graphs and generate emergency resource scheduling solutions.

[0014] The results distribution and display module is used to distribute emergency resource dispatch plans to emergency response terminals for visual display.

[0015] Optionally, modules can be integrated using the following methods:

[0016] Collect multi-source sensor data within the flood monitoring area, process the multi-source sensor data, and generate an input dataset;

[0017] Using geospatial data, the terrain slope and confluence direction between monitoring points are calculated, a directed terrain confluence map is constructed, and a weighted adjacency matrix is ​​generated.

[0018] The input dataset and weighted adjacency matrix are input into the topographic confluence map neural ODE model. A differentiable routing kernel composed of topographic slope and confluence direction is introduced into the adjacency propagation kernel of the topographic confluence map neural ODE model. Water conservation constraints and topographic monotonicity constraints are introduced in the prediction process. The prediction results are output, including short-term prediction sequences and long-term prediction sequences.

[0019] The short-term prediction sequence and the long-term trend prediction sequence are input into the consistency discriminator, the difference between the two is calculated and a consistency index is generated. When the difference exceeds the preset threshold, the prediction result is corrected by the adversarial distillation mechanism to obtain the corrected prediction result and the danger area map.

[0020] Based on the dangerous area map, dangerous areas that need to be focused on prevention and control are identified. The dangerous areas are regarded as risk event objects in the causal relationship diagram, and emergency rescue teams, rescue equipment and drone units are regarded as resource objects in the causal relationship diagram. Causal relationship edges are established between risk event objects and resource objects to form an event-resource causal relationship diagram.

[0021] Based on the event-resource causal relationship graph, a causal game equilibrium mechanism is used for multi-agent scheduling. Each agent constructs a payoff function based on the causal mitigation effect, solves the equilibrium solution through a distributed causal constraint Nash equilibrium algorithm, generates an emergency resource scheduling plan, distributes the emergency resource scheduling plan to the emergency response terminal, and displays it visually on the decision support interface.

[0022] Optionally, the multi-source sensor data includes water level data, rainfall data, flow velocity data, temperature data, humidity data, wind speed data, and air pressure data.

[0023] Optionally, the processing of multi-source sensor data includes time synchronization processing of multi-source sensor data, noise reduction, missing value imputation and normalization processing of multi-source sensor data, and constructing an input dataset of the processed multi-source sensor data according to the timestamp order.

[0024] Optionally, generating the weighted adjacency matrix includes:

[0025] Acquire geospatial data and perform preprocessing. The geospatial data includes digital elevation data, river network data, land cover data, rainfall radar echo data, and latitude and longitude data of monitoring points. Perform invalid value filling, depression filling, and multi-source fusion reconstruction on the digital elevation data to generate a unified raster and quality mask aligned with the river network.

[0026] Based on the unified grid and quality mask, the adaptive neighborhood of the catchment area of ​​each monitoring point is determined according to the catchment boundary. Within the adaptive neighborhood of the catchment area, the slope and aspect of the terrain are calculated by robustly fitted local terrain surfaces.

[0027] Combining slope aspect, the water accumulation direction based on the river network line, the movement vector of rainfall radar echo, and the historical water level-rainfall time lag relationship of the monitoring point, the dominant confluence direction of each monitoring point is determined according to the monotonic elevation reduction constraint and the upstream-downstream coherence constraint.

[0028] Based on the dominant confluence direction, candidate edges are established between monitoring points that meet the conditions for passage of hydraulic obstacles. The edge weight is calculated for each candidate edge, and the edges are filtered according to the direction consistency threshold, the minimum elevation reduction threshold, and the minimum contribution area threshold to obtain a directed topographic confluence map.

[0029] The directed edges and their corresponding weights are organized into a weighted adjacency matrix, and the outgoing edge weights of each monitoring point are normalized to output the weighted adjacency matrix.

[0030] Optionally, the output prediction result includes:

[0031] Receive the weighted adjacency matrix and input dataset, and uniformly calibrate the monitoring point identifiers, time indexes, and external control inputs;

[0032] A graph structure is constructed using monitoring points as graph nodes and weighted adjacency matrices as directed edges. The node status is defined as the water level time series, the external control input is defined as the rainfall intensity time series, and short-term and long-term prediction time windows are set.

[0033] The structure of a topographic confluence map neural ODE model is established, which consists of a message passing unit, a state update unit, and a continuous-time evolution unit, and uses the graph structure, node states, and external control inputs as inputs.

[0034] A differentiable routing kernel is introduced into the adjacency propagation kernel of the topographic confluence map neural ODE model. For each directed edge, continuous differentiable weight modulation and gating are performed based on topographic slope, dominant confluence direction and relative orientation consistency to generate a set of post-routing adjacency weights.

[0035] The message passing unit performs neighborhood information aggregation for each node in each time step, specifically as follows:

[0036] Read the current state of adjacent nodes and their corresponding external control input values ​​according to the set of adjacent weights after routing, and perform gating suppression on the edges that do not meet the direction consistency requirement;

[0037] Weights are assigned to the edges that meet the accessibility conditions, and intermediate messages for each node are calculated and cached as the neighborhood aggregation results of the node in the time-progression substep.

[0038] Within the same time-progressing substep, the state update unit updates the state of each node, combines the intermediate message with the current state of the node and the external control input, and sequentially performs numerical normalization, gating fusion, residual superposition and boundary pruning to obtain the candidate state of the node at the next time step, which serves as the current state input for the next substep.

[0039] By using continuous time evolution units, the node states are gradually updated with a set time step. At each time step, water conservation constraints and terrain monotonicity constraints are applied to each node state candidate. Short-term prediction sequences are generated within the short-term prediction time window, and long-term prediction sequences are generated within the long-term prediction time window. The prediction results are then output.

[0040] Optionally, obtaining the corrected prediction results and the danger zone map includes:

[0041] Receive short-term and long-term prediction sequences, align the two types of sequences with monitoring points according to time steps, determine the range of overlapping time steps and the corresponding set of monitoring points, and establish the input pair of the consistency discriminator;

[0042] A consistency discriminant is constructed. The consistency discriminant is an analysis module used to compare the numerical differences between the short-term and long-term prediction sequences at each time step. The consistency discriminant processes the data based on a set of difference features, which includes numerical difference features, consistency features of change direction, local fluctuation difference features, and outlier marking features. The consistency discriminant performs weighted summation to generate a consistency score and a consistency label.

[0043] Consistency scores are aggregated within the overlapping time steps, normalized according to a preset scale to generate a cross-time domain consistency index, and a consistency threshold is set.

[0044] When the cross-temporal consistency index is lower than the consistency threshold, the adversarial distillation correction process is initiated:

[0045] The long-term forecast sequence is used as the reference sequence, and the short-term forecast sequence is used as the sequence to be corrected.

[0046] The numerical distribution of the short-term forecast sequence is gradually adjusted according to the time step so that the short-term forecast sequence approximates the trend of the long-term forecast sequence.

[0047] Continue until the preset iteration limit is reached or the cross-time domain consistency index no longer decreases, then output the converged correction sequence.

[0048] When the cross-temporal consistency index is not lower than the consistency threshold, the short-term prediction sequence is directly used as the correction sequence;

[0049] The corrected sequence is used as the corrected prediction result, and two types of hazard judgment thresholds are set according to the elevation values ​​of the monitoring points. The fixed threshold is five meters, and the relative threshold is the elevation of the monitoring point minus 0.5 meters. When the water level value of the corrected sequence exceeds any threshold at a time step, the corresponding hazard mark is generated. The hazard marks of each time step are summarized to form a hazard area map.

[0050] Optionally, the construction of the event-resource causal graph based on the danger zone map includes:

[0051] Receive the hazardous area map, identify all key hazardous areas, and decompose each hazardous area into multiple risk evolution segments according to the time series to form a set of risk event objects;

[0052] Collect an emergency resource list, organize all emergency rescue teams, rescue equipment and drone units into a set of resource objects, and construct a multi-dimensional capability vector for each resource object;

[0053] Establish causal relationship edges between the risk event object set and the resource object set. Each edge represents the potential effect of a resource object on a risk event object at a specific time segment.

[0054] A weight value is assigned to each causal relationship edge. The weight value is calculated by an adaptive weight learning function trained based on historical emergency response data. The calculation process takes into account the multidimensional capability vector of the resource object and the risk intensity and time evolution characteristics of the risk event object.

[0055] The risk event object set, resource object set, and all causal relationship edges with adaptive weight values ​​are organized in a unified manner to form an event-resource causal relationship graph.

[0056] Optionally, the generation of the emergency resource scheduling scheme includes:

[0057] Receive the event-resource causal relationship graph, identify all risk event objects and all resource objects, and read the risk mitigation effect weights marked on the causal relationship edges;

[0058] Set a set of optional dispatch actions for each resource object, establish allocation variables to represent the allocation relationship between resource objects and risk event objects, and set constraints;

[0059] The revenue function for resource allocation is defined based on the weights of causal relationship edges and resource allocation variables. The revenue function comprehensively considers the risk mitigation effect and resource cost, which includes travel time, task duration and penalty for task switching.

[0060] The allocation variables of each resource object are adjusted sequentially, while keeping the allocation of other resource objects unchanged in each update. Only the current resource object is allowed to select a new dispatch action if the constraints are met.

[0061] After completing multiple rounds of distributed updates, the resource allocation results are balanced. When all resource objects cannot increase their own benefits beyond a preset threshold by unilateral adjustment while keeping the allocation of other resource objects unchanged, it is determined that a causal game equilibrium has been reached. After the determination converges to equilibrium, an emergency resource scheduling scheme in the equilibrium state is output.

[0062] The beneficial effects of this invention are:

[0063] This invention effectively addresses the shortcomings of existing technologies, such as fixed adjacency matrices and the inability to dynamically reflect the effects of terrain slope and confluence direction, by introducing a topographic confluence map neural ODE model during the prediction stage. By embedding a differentiable routing kernel during adjacency propagation and combining water conservation constraints with terrain monotonicity constraints, this invention can generate short-term and long-term prediction results that better conform to hydrophysical laws in complex terrain environments, improving the accuracy and physical consistency of flood evolution prediction.

[0064] This invention proposes a cross-temporal consistency correction mechanism based on the prediction results. Utilizing consistency discrimination and adversarial distillation correction, it effectively overcomes the lack of connection between short-term and long-term trend predictions in existing methods. This invention can dynamically correct inconsistencies in the prediction sequence, ensuring the continuity and stability of the prediction results across time scales. Simultaneously, it generates a hazardous area map, providing a reliable basis for emergency dispatch.

[0065] This invention introduces an event-resource causal relationship graph and a distributed causal constraint Nash equilibrium algorithm into the emergency resource scheduling process, overcoming the shortcomings of existing methods that rely on heuristic rules for resource allocation and struggle to handle the coupling relationships between multiple agents. Through causal relationship modeling and distributed game equilibrium solving, this invention achieves the rational allocation of various types of resources in the spatiotemporal dimensions, prioritizing coverage of high-risk areas. The generated emergency resource scheduling scheme not only improves rescue efficiency but also enhances the operability and scientific rigor of cross-platform and multi-department collaboration. Attached Figure Description

[0066] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0067] Figure 1 This is a schematic diagram of the structure of the multi-agent collaborative cross-platform flood emergency decision support system proposed in this invention;

[0068] Figure 2This is a flowchart illustrating the multi-agent collaborative cross-platform flood emergency decision support method proposed in this invention. Detailed Implementation

[0069] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0070] refer to Figure 1 A multi-agent collaborative cross-platform flood emergency decision support system includes the following modules:

[0071] The data acquisition and processing module is used to acquire multi-source sensor data, preprocess it, and generate input datasets.

[0072] The terrain confluence construction module is used to calculate the slope and confluence direction, construct a directed terrain confluence map, and generate a weighted adjacency matrix.

[0073] The prediction processing module is used to build a topographic confluence map neural ODE model and output short-term and long-term prediction sequences.

[0074] The consistency correction module is used to determine the consistency between short-term and long-term prediction sequences and output the prediction results and the danger area map.

[0075] The causal relationship construction module is used to generate an event-resource causal relationship diagram based on the danger zone map;

[0076] The causal game scheduling module is used to construct game problems based on event-resource causal relationship graphs and generate emergency resource scheduling solutions.

[0077] The results distribution and display module is used to distribute emergency resource dispatch plans to emergency response terminals for visual display.

[0078] refer to Figure 2 A multi-agent collaborative cross-platform flood emergency decision support method, including:

[0079] Collect multi-source sensor data within the flood monitoring area, process the multi-source sensor data, and generate an input dataset;

[0080] Using geospatial data, the terrain slope and confluence direction between monitoring points are calculated, a directed terrain confluence map is constructed, and a weighted adjacency matrix is ​​generated.

[0081] The input dataset and weighted adjacency matrix are input into the topographic confluence map neural ODE model. A differentiable routing kernel composed of topographic slope and confluence direction is introduced into the adjacency propagation kernel of the topographic confluence map neural ODE model. Water conservation constraints and topographic monotonicity constraints are introduced in the prediction process. The prediction results are output, including short-term prediction sequences and long-term prediction sequences.

[0082] The short-term prediction sequence and the long-term trend prediction sequence are input into the consistency discriminator, the difference between the two is calculated and a consistency index is generated. When the difference exceeds the preset threshold, the prediction result is corrected by the adversarial distillation mechanism to obtain the corrected prediction result and the danger area map.

[0083] Based on the dangerous area map, dangerous areas that need to be focused on prevention and control are identified. The dangerous areas are regarded as risk event objects in the causal relationship diagram, and emergency rescue teams, rescue equipment and drone units are regarded as resource objects in the causal relationship diagram. Causal relationship edges are established between risk event objects and resource objects to form an event-resource causal relationship diagram.

[0084] Based on the event-resource causal relationship graph, a causal game equilibrium mechanism is used for multi-agent scheduling. Each agent constructs a payoff function based on the causal mitigation effect, solves the equilibrium solution through a distributed causal constraint Nash equilibrium algorithm, generates an emergency resource scheduling plan, distributes the emergency resource scheduling plan to the emergency response terminal, and displays it visually on the decision support interface.

[0085] In this embodiment, the multi-source sensor data includes water level data, rainfall data, flow velocity data, temperature data, humidity data, wind speed data, and air pressure data.

[0086] In this embodiment, the processing of multi-source sensor data includes time synchronization processing of multi-source sensor data, noise reduction, missing value imputation and normalization processing of multi-source sensor data, and constructing an input dataset of the processed multi-source sensor data according to the timestamp order.

[0087] In this embodiment, generating the weighted adjacency matrix includes:

[0088] Acquire geospatial data and perform preprocessing. The geospatial data includes digital elevation data, river network data, land cover data, rainfall radar echo data, and latitude and longitude data of monitoring points. Perform invalid value filling, depression filling, and multi-source fusion reconstruction on the digital elevation data to generate a unified raster and quality mask aligned with the river network.

[0089] Based on the unified grid and quality mask, an adaptive neighborhood of the catchment area for each monitoring point is determined according to the catchment boundary. Within the adaptive neighborhood of the catchment area, the slope and aspect of the terrain are calculated using a robustly fitted local terrain surface. Specifically, the calculation of the slope and aspect using a robustly fitted local terrain surface involves:

[0090] Extract neighborhood raster elevation values ​​and plane coordinates within the adaptive neighborhood of the catchment area of ​​each monitoring point to construct a local elevation sample set;

[0091] A robust least squares regression method is used to perform plane fitting on the local elevation sample set to obtain the local terrain fitting plane parameters.

[0092] The slope and aspect are derived from the parameters of the fitted plane based on the local terrain. The slope is determined by the angle of inclination of the plane, and the aspect is determined by the direction of inclination of the plane.

[0093] Combining slope aspect, the water accumulation direction based on the river network line, the movement vector of rainfall radar echoes, and the historical water level-rainfall time lag relationship at monitoring points, the dominant flow direction at each monitoring point is determined according to the monotonic elevation reduction constraint and the upstream-downstream coherence constraint, where:

[0094] The monotonic elevation reduction constraint means that the elevation value along the confluence path from the upstream monitoring point to the downstream monitoring point must gradually decrease, and it is not allowed that the downstream elevation is greater than the upstream elevation, so as to ensure that the water flow direction conforms to the physical law of flowing from high to low.

[0095] Upstream and downstream continuity constraints refer to the requirement that the confluence path must be consistent with the existing river network when determining the dominant confluence direction, to ensure the hydraulic continuity of upstream inflow and downstream outflow, and to avoid discontinuities or situations that do not conform to the actual river network direction.

[0096] Based on the dominant confluence direction, candidate edges are established between monitoring points that meet the conditions for passage through hydraulic barriers. The edge weight is calculated for each candidate edge, and edges are filtered according to a directional consistency threshold, a minimum elevation reduction threshold, and a minimum contribution area threshold to obtain a directed topographic confluence map. Specifically, the calculation of the edge weight for each candidate edge involves:

[0097] Based on the elevation difference between monitoring points, the elevation reduction of water flow from upstream to downstream is calculated, and the elevation reduction is used as the slope factor for candidate connections.

[0098] Based on the angle difference between the dominant confluence direction and the candidate connection direction, a direction consistency index is calculated as the direction factor of the candidate connection.

[0099] Based on the catchment area covered by the candidate connection, the corresponding contribution area index is calculated and used as the area factor of the candidate connection.

[0100] The slope factor, direction factor, and area factor are combined according to preset weights to obtain the comprehensive weight value of the candidate connection.

[0101] The directed edges and their corresponding weights are organized into a weighted adjacency matrix, and the outgoing edge weights of each monitoring point are normalized to output the weighted adjacency matrix.

[0102] In this embodiment, the output prediction result includes:

[0103] Receive the weighted adjacency matrix and input dataset, and uniformly calibrate the monitoring point identifiers, time indexes, and external control inputs;

[0104] A graph structure is constructed using monitoring points as graph nodes and weighted adjacency matrices as directed edges. The node status is defined as the water level time series, the external control input is defined as the rainfall intensity time series, and short-term and long-term prediction time windows are set.

[0105] The structure of a topographic confluence map neural ODE model is established, which consists of a message passing unit, a state update unit, and a continuous-time evolution unit, and uses the graph structure, node states, and external control inputs as inputs.

[0106] A differentiable routing kernel is introduced into the adjacency propagation kernel of the topographic confluence map neural ODE model. For each directed edge, continuously differentiable weight modulation and gating are performed based on topographic slope, dominant confluence direction, and relative orientation consistency to generate a set of post-routing adjacency weights. Specifically, the continuously differentiable weight modulation and gating are performed as follows:

[0107] The weight of the connection is continuously scaled based on the slope of the terrain. The greater the slope, the greater the increase in the weight of the connection; the smaller the slope, the less the weight of the connection.

[0108] The connection weight is smoothly modulated based on the consistency between the connection direction and the dominant confluence direction. When the two are consistent, the weight is enhanced, and when the deviation increases, the weight gradually weakens.

[0109] Based on the consistency between the connection direction and the relative orientation, a differentiable gate function is set to gate the weights. When the relative orientation condition is met, the gate value is close to 1, and otherwise it gradually approaches 0.

[0110] The slope modulation factor, direction consistency modulation factor and gating factor are combined in a continuously differentiable weighted manner to generate the final route adjacency weight of each edge.

[0111] The message passing unit performs neighborhood information aggregation for each node in each time step, specifically as follows:

[0112] Read the current state of adjacent nodes and their corresponding external control input values ​​according to the set of adjacent weights after routing, and perform gating suppression on the edges that do not meet the direction consistency requirement;

[0113] Weights are assigned to edges that meet the accessibility criteria. The intermediate messages for each node are calculated and cached as the neighborhood aggregation results for each node in the time-progression sub-step. Specifically, the calculation and caching of the intermediate messages for each node involves:

[0114] For each edge, the state information of the upstream node is multiplied by the edge weight according to the recalibrated weight to obtain the weighted transfer value;

[0115] All weighted transit values ​​entering the same target node are summed or aggregated in a continuously differentiable manner to form the neighborhood aggregate value of the target node;

[0116] The neighborhood aggregation value is fused with the target node's own state vector to generate the node's intermediate message, and the generated intermediate message is written to the cache.

[0117] Within the same time-progressing substep, the state update unit updates the state of each node, combining intermediate messages with the node's current state and external control inputs, and sequentially performing numerical normalization, gating fusion, residual superposition, and boundary pruning to obtain the node's next-time state candidate, which serves as the current state input for the next substep.

[0118] Numerical normalization refers to applying a uniform scaling transformation to intermediate messages, the current state of nodes, and external control inputs, mapping each input variable to the same numerical range.

[0119] Gated fusion refers to the introduction of continuously differentiable gating factors to weight and combine intermediate messages, node states, and external control inputs according to the relative importance of different inputs.

[0120] Residual superposition refers to retaining the historical state components of a node during the update process and adding them to the merged new state components;

[0121] Boundary pruning refers to applying physical rationality constraints to the updated state candidates, forcibly truncating or compressing values ​​that exceed the feasible range into a preset interval;

[0122] Through continuous-time evolution units, the node states are progressively updated at set time steps. At each time step, water conservation constraints and terrain monotonicity constraints are applied to each node state candidate. Short-term prediction sequences are generated within the short-term prediction window, and long-term prediction sequences are generated within the long-term prediction window. The prediction results are then output, including:

[0123] Water conservation constraint refers to maintaining a computational balance between the amount of water entering a node and the amount of water flowing out of a node during the node state update process, ensuring that the total water volume of the entire watershed does not increase or decrease in a non-physical way.

[0124] The topographic monotonicity constraint means that the water level or elevation of the downstream node must not be higher than that of the upstream node on the confluence path, so as to ensure that the water flow always follows a monotonically decreasing law from high to low in the spatial evolution.

[0125] In this embodiment, obtaining the corrected prediction results and the danger zone map includes:

[0126] Receive short-term and long-term prediction sequences, align the two types of sequences with monitoring points according to time steps, determine the range of overlapping time steps and the corresponding set of monitoring points, and establish the input pair of the consistency discriminator;

[0127] A consistency discriminant is constructed, which is an analysis module used to compare the numerical differences between short-term and long-term prediction sequences at each time step. The consistency discriminant processes data based on a set of difference features, including numerical difference features, consistency features in the direction of change, local fluctuation difference features, and outlier marker features. These features are weighted and aggregated by the consistency discriminant to generate a consistency score and a consistency label.

[0128] The numerical difference feature is obtained by calculating the numerical difference between the short-term and long-term prediction sequences at each time step.

[0129] The consistency characteristic of change direction is obtained by comparing whether the upward or downward trends of the short-term forecast sequence and the long-term forecast sequence are consistent at adjacent time steps;

[0130] The local fluctuation difference characteristics are obtained by statistically analyzing and calculating the differences between the local fluctuation amplitudes of the short-term and long-term prediction sequences using a sliding window.

[0131] Anomaly marker features are obtained by detecting numerical abrupt changes, deviations from trends, or exceeding thresholds in short-term or long-term prediction sequences at time steps, and anomalies are used as feature markers.

[0132] Consistency scores are aggregated within the overlapping time steps, normalized according to a preset scale to generate a cross-time domain consistency index, and a consistency threshold is set.

[0133] When the cross-temporal consistency index is lower than the consistency threshold, the adversarial distillation correction process is initiated:

[0134] The long-term forecast sequence is used as the reference sequence, and the short-term forecast sequence is used as the sequence to be corrected.

[0135] The numerical distribution of the short-term forecast sequence is gradually adjusted according to the time step so that the short-term forecast sequence approximates the trend of the long-term forecast sequence.

[0136] Continue until the preset iteration limit is reached or the cross-time domain consistency index no longer decreases, then output the converged correction sequence.

[0137] When the cross-temporal consistency index is not lower than the consistency threshold, the short-term prediction sequence is directly used as the correction sequence;

[0138] The corrected sequence is used as the corrected prediction result, and two types of hazard judgment thresholds are set according to the elevation values ​​of the monitoring points. The fixed threshold is five meters, and the relative threshold is the elevation of the monitoring point minus 0.5 meters. When the water level value of the corrected sequence exceeds any threshold at a time step, the corresponding hazard mark is generated. The hazard marks of each time step are summarized to form a hazard area map.

[0139] In this embodiment, the construction of an event-resource cause-effect graph based on a hazardous area map includes:

[0140] Receive the hazardous area map, identify all key hazardous areas, and decompose each hazardous area into multiple risk evolution segments according to the time series to form a set of risk event objects;

[0141] Collect an emergency resource list, organize all emergency rescue teams, rescue equipment and drone units into a set of resource objects, and construct a multi-dimensional capability vector for each resource object;

[0142] Establish causal relationship edges between the risk event object set and the resource object set. Each edge represents the potential effect of a resource object on a risk event object at a specific time segment.

[0143] Each causal relationship edge is assigned a weight value, which is calculated by an adaptive weight learning function trained based on historical emergency response data. The calculation process considers both the multidimensional capability vector of the resource object and the risk intensity and temporal evolution characteristics of the risk event object. The adaptive weight learning function is a continuously differentiable function constructed based on supervised learning and feature embedding methods. By inputting the capability vector of the resource object and the feature vector of the risk event object, the interaction features between the two in the spatiotemporal dimension are extracted. The function is trained using historical emergency response data and the parameters are continuously adjusted so that the output weight value can dynamically reflect the relative contribution of resources to the risk mitigation effect.

[0144] The risk event object set, resource object set, and all causal relationship edges with adaptive weight values ​​are organized in a unified manner to form an event-resource causal relationship graph.

[0145] In this embodiment, generating an emergency resource scheduling plan includes:

[0146] Receive the event-resource causal relationship graph, identify all risk event objects and all resource objects, and read the risk mitigation effect weights marked on the causal relationship edges;

[0147] For each resource object, a set of optional dispatch actions is defined; an allocation variable is established to represent the allocation relationship between the resource object and the risk event object; and constraints are set, including:

[0148] Each resource object can only serve one risk event object at any given time.

[0149] Each risk event object should meet the minimum coverage requirement within the same time segment;

[0150] The travel time and task switching time of each resource object must not exceed the reachable range;

[0151] The load capacity of each resource object must not be lower than the task requirements;

[0152] A reward function for resource allocation is defined based on the weights of causal edges and resource allocation variables. This reward function comprehensively considers the risk mitigation effect and resource costs, including travel time, task duration, and penalties for task switching. The reward function is a function that quantifies the effect of a resource object on a specific risk event object under a given resource allocation scheme. The calculation process is as follows:

[0153] The contribution of resource objects to mitigating the intensity of risk events is determined based on the weight of the causal relationship edges;

[0154] The costs incurred by resource objects during execution are deducted, including travel time to the location of the risk event, duration of execution of the handling task, and delay costs incurred when switching between different tasks;

[0155] By comprehensively balancing the contribution of risk mitigation with the cost of resources, the benefit value of resource allocation can be obtained;

[0156] The allocation variables of each resource object are adjusted sequentially, while keeping the allocation of other resource objects unchanged in each update. Only the current resource object is allowed to select a new dispatch action if the constraints are met.

[0157] After completing multiple rounds of distributed updates, the resource allocation results are balanced. When all resource objects cannot increase their own benefits by more than a preset threshold by unilateral adjustment while keeping the allocation of other resource objects unchanged, it is determined that a causal game equilibrium has been reached. After the determination converges to equilibrium, an emergency resource scheduling scheme in the equilibrium state is output. The emergency resource scheduling scheme includes resource object identifier, dispatch target area, dispatch time segment, task execution duration and return path. Example

[0158] To verify the feasibility of this invention in practice, it was applied to the risk prevention and control of a river basin flood triggered by a period of continuous heavy rainfall. In this case, the water level in a certain river basin rose rapidly within a short period. The flood evolution was influenced by complex terrain and the combined effects of multiple tributaries. Traditional prediction methods based on static adjacency matrices showed significant errors in local low-lying areas, failing to accurately capture the dynamic changes in the confluence path. The dispatch of rescue resources still relied on human experience, resulting in problems such as delayed arrival of rescue teams, uneven deployment of pumping equipment, and insufficient monitoring coverage by drones. This led to a failure to respond promptly to some high-risk areas, severely impacting emergency response efficiency.

[0159] In the application of this invention, the data acquisition and processing module first collects and processes multi-source sensor data deployed within the watershed, including water level, rainfall, flow velocity, and meteorological information, generating a standardized input dataset. Using geospatial data, the slope and confluence direction between monitoring points are calculated to construct a directed topographic confluence map and generate a weighted adjacency matrix. In the topographic confluence map neural ODE prediction module, a differentiable routing kernel composed of slope and confluence direction is introduced. Simultaneously, water conservation constraints and topographic monotonicity constraints are applied during the prediction process to obtain short-term and long-term prediction sequences. Compared with traditional methods, the prediction results better conform to hydrophysical laws, and the error is significantly reduced in low-lying areas and slope transition zones.

[0160] After prediction, this invention further compares short-term and long-term results through a cross-temporal consistency correction module. When the consistency index falls below a set threshold, the system automatically executes an adversarial distillation correction mechanism to dynamically correct the short-term prediction results, ensuring the continuity and stability of the prediction results over time. In this scenario, the deviation between the corrected peak water level and the actual observed value is controlled within 2%, significantly better than the approximately 8% deviation level of traditional methods. Based on the correction results, the system generates a hazardous area map, accurately identifying several key prevention and control areas and providing input support for resource scheduling.

[0161] During the resource scheduling phase, the system constructs an event-resource causal relationship graph based on a danger zone map. Danger zones are treated as risk events, while rescue teams, water pumps, and drone units are treated as resources. An adaptive weight learning function assigns weights to the edges between risk events and resources, with the weight reflecting the effectiveness of each resource in mitigating a specific risk event. Subsequently, the causal game scheduling module employs a distributed causal constraint Nash equilibrium algorithm to progressively update the resource dispatch strategy until convergence to an equilibrium solution. High-risk areas, due to their higher weights, are prioritized for coverage, effectively improving overall rescue efficiency and resource utilization.

[0162] Ultimately, the system's output scheduling plan showed that, in situations where the risk in dangerous areas rapidly increased, rescue teams arrived an average of 30 minutes ahead of schedule, water pumps were more concentrated in areas with the highest risk of flooding, and drones achieved full coverage of all key monitoring points. Compared with traditional manual scheduling methods, this invention demonstrates significant advantages in prediction accuracy, rescue response time, and coverage, effectively solving the problems of large prediction errors, uneven scheduling, and insufficient multi-agent collaboration mentioned in the background technology.

[0163] Table 1 Comparison of Flood Forecasting and Resource Allocation Effects

[0164]

[0165] As shown in Table 1, in terms of water level prediction, traditional methods have an average error of about 8% in peak prediction. However, this invention, by introducing a topographic confluence map neural ODE model and cross-temporal consistency correction, controls the error to within 2%, a reduction of over 75%, thus improving the consistency between the prediction results and actual observations. Regarding hazard area identification, traditional methods have an accuracy rate of 72%, sometimes missing high-risk areas. This invention, however, achieves an accuracy rate of 94%, an improvement of 22 percentage points, providing a more reliable basis for resource allocation.

[0166] Regarding resource allocation efficiency, this invention optimizes the collaborative scheduling of multiple agents through a causal game equilibrium mechanism. The average arrival time of rescue teams is reduced from the traditional 2.3 hours to 1.8 hours, achieving a rapid response approximately 30 minutes earlier, thus saving valuable time for emergency response. The deployment coverage of water pumps has increased from the traditional 65% to 92%, an improvement of 27 percentage points, indicating that in areas with high risk of flooding, this invention can significantly improve equipment utilization and deployment rationality. The monitoring coverage of drones has also increased from 70% using traditional methods to 100%, achieving full coverage of key areas and providing sufficient data support for disaster monitoring and situation assessment.

[0167] In terms of overall prevention and control effectiveness, the dangerous area coverage rate generated by this invention has increased from the traditional 68% to 93%, an improvement of 25 percentage points. This means that most high-risk areas can be monitored and intervened in a timely manner, avoiding the resource gap problem existing in traditional methods. This invention not only improves prediction accuracy and dangerous area identification level, but also improves the response speed and coverage of rescue resources through a distributed causal constraint Nash equilibrium algorithm, effectively solving the defects of inaccurate prediction, unbalanced scheduling, and insufficient coordination in existing technologies.

[0168] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A cross-platform flood emergency decision support system with multi-agent collaboration, characterized in that, Includes the following modules: The data acquisition and processing module is used to acquire multi-source sensor data, preprocess it, and generate input datasets. The terrain confluence construction module is used to calculate the slope and confluence direction, construct a directed terrain confluence map, and generate a weighted adjacency matrix. The prediction processing module is used to build a topographic confluence map neural ODE model and output short-term and long-term prediction sequences. The consistency correction module is used to determine the consistency between short-term and long-term prediction sequences and output the prediction results and the danger area map. The causal relationship construction module is used to generate an event-resource causal relationship diagram based on the danger zone map; The causal game scheduling module is used to construct game problems based on event-resource causal relationship graphs and generate emergency resource scheduling solutions. The results distribution and display module is used to distribute emergency resource dispatch plans to emergency response terminals for visual display. include: Collect multi-source sensor data within the flood monitoring area, process the multi-source sensor data, and generate an input dataset; Using geospatial data, the terrain slope and confluence direction between monitoring points are calculated, a directed terrain confluence map is constructed, and a weighted adjacency matrix is ​​generated. The input dataset and weighted adjacency matrix are input into the topographic confluence map neural ODE model. A differentiable routing kernel composed of topographic slope and confluence direction is introduced into the adjacency propagation kernel of the topographic confluence map neural ODE model. Water conservation constraints and topographic monotonicity constraints are introduced in the prediction process. The prediction results are output, including short-term prediction sequences and long-term prediction sequences. The short-term prediction sequence and the long-term trend prediction sequence are input into the consistency discriminator, the difference between the two is calculated and a consistency index is generated. When the difference exceeds the preset threshold, the prediction result is corrected by the adversarial distillation mechanism to obtain the corrected prediction result and the danger area map. Based on the dangerous area map, dangerous areas that need to be focused on prevention and control are identified. The dangerous areas are regarded as risk event objects in the causal relationship diagram, and emergency rescue teams, rescue equipment and drone units are regarded as resource objects in the causal relationship diagram. Causal relationship edges are established between risk event objects and resource objects to form an event-resource causal relationship diagram. Based on the event-resource causal relationship graph, a causal game equilibrium mechanism is used for multi-agent scheduling. Each agent constructs a payoff function based on the causal mitigation effect, solves the equilibrium solution through a distributed causal constraint Nash equilibrium algorithm, generates an emergency resource scheduling plan, distributes the emergency resource scheduling plan to the emergency response terminal, and displays it visually on the decision support interface.

2. The multi-agent collaborative cross-platform flood emergency decision support system according to claim 1, characterized in that, The multi-source sensor data includes water level data, rainfall data, flow velocity data, temperature data, humidity data, wind speed data, and air pressure data.

3. The multi-agent collaborative cross-platform flood emergency decision support system according to claim 1, characterized in that, The processing of multi-source sensor data includes time synchronization processing, noise reduction, missing value imputation and normalization processing, and constructing an input dataset from the processed multi-source sensor data according to the timestamp order.

4. The multi-agent collaborative cross-platform flood emergency decision support system according to claim 1, characterized in that, The generation of the weighted adjacency matrix includes: Acquire geospatial data and perform preprocessing. The geospatial data includes digital elevation data, river network data, land cover data, rainfall radar echo data, and latitude and longitude data of monitoring points. Perform invalid value filling, depression filling, and multi-source fusion reconstruction on the digital elevation data to generate a unified raster and quality mask aligned with the river network. Based on the unified grid and quality mask, the adaptive neighborhood of the catchment area of ​​each monitoring point is determined according to the catchment boundary. Within the adaptive neighborhood of the catchment area, the slope and aspect of the terrain are calculated by robustly fitted local terrain surfaces. Combining slope aspect, the water accumulation direction based on the river network line, the movement vector of rainfall radar echo, and the historical water level-rainfall time lag relationship of the monitoring point, the dominant confluence direction of each monitoring point is determined according to the monotonic elevation reduction constraint and the upstream-downstream coherence constraint. Based on the dominant confluence direction, candidate edges are established between monitoring points that meet the conditions for passage of hydraulic obstacles. The edge weight is calculated for each candidate edge, and the edges are filtered according to the direction consistency threshold, the minimum elevation reduction threshold, and the minimum contribution area threshold to obtain a directed topographic confluence map. The directed edges and their corresponding weights are organized into a weighted adjacency matrix, and the outgoing edge weights of each monitoring point are normalized to output the weighted adjacency matrix.

5. The multi-agent collaborative cross-platform flood emergency decision support system according to claim 1, characterized in that, The output prediction results include: Receive the weighted adjacency matrix and input dataset, and uniformly calibrate the monitoring point identifiers, time indexes, and external control inputs; A graph structure is constructed using monitoring points as graph nodes and weighted adjacency matrices as directed edges. The node status is defined as the water level time series, the external control input is defined as the rainfall intensity time series, and short-term and long-term prediction time windows are set. The structure of a topographic confluence map neural ODE model is established, which consists of a message passing unit, a state update unit, and a continuous-time evolution unit, and uses the graph structure, node states, and external control inputs as inputs. A differentiable routing kernel is introduced into the adjacency propagation kernel of the topographic confluence map neural ODE model. For each directed edge, continuously differentiable weight modulation and gating are performed based on topographic slope, dominant confluence direction, and relative orientation consistency to generate a set of post-routing adjacency weights. Specifically, the continuously differentiable weight modulation and gating are performed as follows: The weight of the connection is continuously scaled based on the slope of the terrain. The greater the slope, the greater the increase in the weight of the connection; the smaller the slope, the less the weight of the connection. The connection weight is smoothly modulated based on the consistency between the connection direction and the dominant confluence direction. When the two are consistent, the weight is enhanced, and when the deviation increases, the weight gradually weakens. Based on the consistency between the connection direction and the relative orientation, a differentiable gate function is set to gate the weights. When the relative orientation condition is met, the gate value is close to 1, and otherwise it gradually approaches 0. The slope modulation factor, direction consistency modulation factor and gating factor are combined in a continuously differentiable weighted manner to generate the final route adjacency weight of each edge. The message passing unit performs neighborhood information aggregation for each node in each time step, specifically as follows: Read the current state of adjacent nodes and their corresponding external control input values ​​according to the set of adjacent weights after routing, and perform gating suppression on the edges that do not meet the direction consistency requirement; Weights are assigned to the edges that meet the accessibility conditions, and intermediate messages for each node are calculated and cached as the neighborhood aggregation results of the node in the time-progression substep. Within the same time-progressing substep, the state update unit updates the state of each node, combines the intermediate message with the current state of the node and the external control input, and sequentially performs numerical normalization, gating fusion, residual superposition and boundary pruning to obtain the candidate state of the node at the next time step, which serves as the current state input for the next substep. By using continuous time evolution units, the node states are gradually updated with a set time step. At each time step, water conservation constraints and terrain monotonicity constraints are applied to each node state candidate. Short-term prediction sequences are generated within the short-term prediction time window, and long-term prediction sequences are generated within the long-term prediction time window. The prediction results are then output.

6. The multi-agent collaborative cross-platform flood emergency decision support system according to claim 1, characterized in that, The obtained corrected prediction results and danger zone map include: Receive short-term and long-term prediction sequences, align the two types of sequences with monitoring points according to time steps, determine the range of overlapping time steps and the corresponding set of monitoring points, and establish the input pair of the consistency discriminator; A consistency discriminant is constructed. The consistency discriminant is an analysis module used to compare the numerical differences between the short-term and long-term prediction sequences at each time step. The consistency discriminant processes the data based on a set of difference features, which includes numerical difference features, consistency features of change direction, local fluctuation difference features, and outlier marking features. The consistency discriminant performs weighted summation to generate a consistency score and a consistency label. Consistency scores are aggregated within the overlapping time steps, normalized according to a preset scale to generate a cross-time domain consistency index, and a consistency threshold is set. When the cross-temporal consistency index is lower than the consistency threshold, the adversarial distillation correction process is initiated: The long-term forecast sequence is used as the reference sequence, and the short-term forecast sequence is used as the sequence to be corrected. The numerical distribution of the short-term forecast sequence is gradually adjusted according to the time step so that the short-term forecast sequence approximates the trend of the long-term forecast sequence. Continue until the preset iteration limit is reached or the cross-time domain consistency index no longer decreases, then output the converged correction sequence. When the cross-temporal consistency index is not lower than the consistency threshold, the short-term prediction sequence is directly used as the correction sequence; The corrected sequence is used as the corrected prediction result, and two types of hazard judgment thresholds are set according to the elevation values ​​of the monitoring points. The fixed threshold is five meters, and the relative threshold is the elevation of the monitoring point minus 0.5 meters. When the water level value of the corrected sequence exceeds any threshold at a time step, the corresponding hazard mark is generated. The hazard marks of each time step are summarized to form a hazard area map.

7. The multi-agent collaborative cross-platform flood emergency decision support system according to claim 1, characterized in that, The construction of the event-resource cause-effect graph based on the danger zone map includes: Receive the hazardous area map, identify all key hazardous areas, and decompose each hazardous area into multiple risk evolution segments according to the time series to form a set of risk event objects; Collect an emergency resource list, organize all emergency rescue teams, rescue equipment and drone units into a set of resource objects, and construct a multi-dimensional capability vector for each resource object; Establish causal relationship edges between the risk event object set and the resource object set. Each edge represents the potential effect of a resource object on a risk event object at a specific time segment. A weight value is assigned to each causal relationship edge. The weight value is calculated by an adaptive weight learning function trained based on historical emergency response data. The calculation process takes into account the multidimensional capability vector of the resource object and the risk intensity and time evolution characteristics of the risk event object. The risk event object set, resource object set, and all causal relationship edges with adaptive weight values ​​are organized in a unified manner to form an event-resource causal relationship graph.

8. The multi-agent collaborative cross-platform flood emergency decision support system according to claim 1, characterized in that, The generation of the emergency resource scheduling scheme includes: Receive the event-resource causal relationship graph, identify all risk event objects and all resource objects, and read the risk mitigation effect weights marked on the causal relationship edges; Set a set of optional dispatch actions for each resource object, establish allocation variables to represent the allocation relationship between resource objects and risk event objects, and set constraints; The revenue function for resource allocation is defined based on the weights of causal relationship edges and resource allocation variables. The revenue function comprehensively considers the risk mitigation effect and resource cost, which includes travel time, task duration and penalty for task switching. The allocation variables of each resource object are adjusted sequentially, while keeping the allocation of other resource objects unchanged in each update. Only the current resource object is allowed to select a new dispatch action if the constraints are met. After completing multiple rounds of distributed updates, the resource allocation results are balanced. When all resource objects cannot increase their own benefits beyond a preset threshold by unilateral adjustment while keeping the allocation of other resource objects unchanged, it is determined that a causal game equilibrium has been reached. After the determination converges to equilibrium, an emergency resource scheduling scheme in the equilibrium state is output.

Images