Flood priority scheduling intelligent evaluation method and system for multi-mode flood early warning

Abstract

The invention discloses a flood priority scheduling intelligent evaluation method for multi-modal flood early warning. The method comprises the following steps: step 1, rainfall, water level and flow velocity data are collected and preprocessed to generate standardized multi-modal input; 2, constructing a river network graph structure, and extracting node and edge features to generate graph structure input data; step 3, inputting the graph structure into an EvolveGCN evolutionary graph convolutional network, embedding scheduling path accessibility, resource bottleneck and execution constraint features, and outputting a state propagation feature set; step 4, generating a node state prediction vector according to the state propagation characteristics; 5, calculating three types of risk scores, and applying nonlinear constraints to construct a priority score set; step 6, generating a scheduling action sequence and collecting feedback data; and step 7, updating the convolution structure by using the feedback updating score and the scheduling sequence, and outputting a final instruction. The flood scheduling intelligent prediction closed loop is realized, and the response efficiency and the scheduling feasibility are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent scheduling evaluation, in particular to a flood priority scheduling intelligent evaluation method and system for multi-modal flood warning. BACKGROUND

[0002] The existing flood warning and scheduling system mainly relies on a rule model driven by hydrological monitoring data for risk evaluation and scheduling decision. Most systems use a static threshold judgment method to make a hierarchical response by setting the trigger conditions of rainfall, water level or flow rate. This method is relatively fixed in response speed and rule configuration, and is difficult to adapt to the dynamic changes of the flood evolution process in complex basin structures, especially in multi-source data access, multi-node linkage and control executability evaluation.

[0003] Some technical solutions attempt to introduce a time series prediction model or a multi-source data fusion method to make a more fine-grained judgment of flood risk. However, these methods often lack the modeling capability of river network structure and cannot effectively represent the upstream and downstream dependency relationship and control path structure between nodes. In the scheduling decision link, the traditional strategy performs sequentially based on the preset instruction priority, lacks dynamic scoring and feasibility evaluation mechanism, and causes some scheduling actions to be unable to truly land under the conditions of resource limitation or path interruption, thereby affecting the overall response effect.

[0004] In terms of graph structure modeling, the existing few methods use static graph neural networks to predict node states, but lack support for time series evolution, control reachability and feedback update process. The execution result of the instruction is disconnected from the priority model, the system cannot adjust the scoring logic and scheduling order according to the feedback, and lacks effective closed-loop scheduling capability. The existing technology cannot realize a flood priority scheduling evaluation method integrating multi-modal input, structure dynamic propagation, scheduling feasibility evaluation and feedback-driven update.

[0005] Therefore, how to provide a flood priority scheduling intelligent evaluation method and system for multi-modal flood warning is a problem that those skilled in the art need to solve. SUMMARY

[0006] One object of the present application is to provide a flood priority scheduling intelligent evaluation method and system for multi-modal flood warning. The present application uses an evolutionary graph convolution network and a feasibility-enhanced convolution structure to construct a scheduling strategy based on scoring-driven, structure-aware and feedback-updated, which has the advantages of strong decision-making closed loop, high response accuracy and good control landing performance.

[0007] The flood priority scheduling intelligent evaluation method and system for multi-modal flood warning according to the embodiments of the present application include the following steps: Step one, collect rainfall, water level and flow rate data, and preprocess to generate standardized multi-modal input data; Step two, construct the river network graph structure, represent the observation points as nodes and the river connections as edges, extract node and edge features, and generate graph structure input data; Step three, input the graph structure input data into the evolved graph convolutional network constructed based on EvolveGCN, propagate the node state feature using the feasible enhanced convolution structure, embed the reachability of the scheduling path, resource bottleneck, and execution constraint features during the convolution process, and output the state propagation feature set; Step four, generate a node state prediction vector based on the state propagation feature set; Step five, calculate three types of indexes: basic risk, dynamic threat, and regional vulnerability based on the node state prediction vector, apply a nonlinear constraint mechanism including an upper limit suppression function and a normalization function to construct a priority score feature set, and generate a node priority score set; Step six, generate a scheduling action sequence based on the node priority score set, send the scheduling action sequence to the control object for execution, collect water level change data, flow rate change data, and response state information for the corresponding nodes during execution, and construct a scheduling execution feedback data set; Step seven, trigger dynamic updating and reordering of the node priority score set and the scheduling action sequence based on the scheduling execution feedback data set, update the feasibility-enhanced convolution structure based on the feedback data set, and generate a final scheduling action instruction set.

[0008] Optionally, the preprocessing in step one includes performing time alignment, missing completion, anomaly removal, and normalization processing on rainfall, water level, and flow rate data to generate standardized multi-modal input data.

[0009] Optionally, the evolved graph convolutional network based on EvolveGCN in step three includes a graph convolution parameter initialization unit, a graph convolution parameter evolution unit, a gated recurrent unit, a node state update unit, a historical state fusion unit, a state storage unit, and an output scheduling unit, specifically: In the graph convolution initialization phase, the graph convolution parameter initialization unit receives the graph structure input data at the input end of the evolved graph convolutional network, divides the graph structure input data into several time-step graph structure snapshots, and initializes the node state propagation weight matrix and bias term for each layer of graph convolution structure based on the time-step graph structure snapshot; At the beginning of each time step, the graph convolution parameter evolution unit calls the convolution weight set of the previous time step in the parameter evolution path, combines the node state change information of the current time-step graph structure snapshot, performs parameter update operations through the gated recurrent unit, and generates the convolution weight set of the current time step; In each time step convolution propagation stage, the historical state propagation feature set is called by the node state update unit for each graph structure node, and the graph structure convolution operation is performed according to the convolution weight set and the adjacent edge weight information of the current time step to calculate the node state propagation feature, and the operation is completed in parallel in all nodes of the graph structure; After state propagation, the historical state fusion unit calls the state propagation features of the past time steps of the node in the state propagation path of the current time step, and adopts splicing, linear fusion or residual connection to complete historical information fusion, and generates the fused node state propagation feature representation; After each time step is completed, the state storage unit records the state propagation features of all nodes of the current time step in the time sequence feature cache area as the input basis for the next time step graph convolution parameter evolution and state fusion; After all time steps are executed, the output scheduling unit integrates the node state propagation feature sets generated in each time step in sequence, and outputs the final state propagation feature set.

[0010] Optionally, the feasibility enhancement convolution structure in step three includes a path reachability regulation unit, a resource bottleneck weight regulation unit and an execution constraint activation unit, specifically: The path reachability regulation unit calls the scheduling path reachability feature in the graph structure input data before the node state update unit performs the graph structure convolution operation, performs path reachability discrimination processing on the adjacent nodes of each graph structure node, removes the unreachable adjacent nodes from the feature aggregation path, and retains the reachable adjacent nodes to form the state propagation input set; The resource bottleneck weight regulation unit calls the field representing the resource capacity in the graph structure input data during the graph structure convolution operation, performs propagation channel weight adjustment operation, maps the resource bottleneck feature value corresponding to the edge to the propagation weight adjustment factor, and adjusts the weight value of the corresponding edge in the convolution weight set according to the edge; The execution constraint activation unit calls the execution constraint feature field in the graph structure input data after the state propagation feature is generated, compares the execution constraint state of each node with the set threshold, marks the nodes that do not meet the condition and stops the output of the state propagation feature of the node in the current time step, and retains the state propagation feature of the node that meets the condition.

[0011] Optionally, the non-linear constraint mechanism in step five is executed by a score construction unit, specifically including: The score construction unit receives the node state prediction vector, extracts the basic risk index for describing the flood intensity, the dynamic threat index for describing the state change rate and the regional vulnerability index for describing the population exposure degree to form a three-dimensional score index set after receiving the node state prediction vector; The score construction unit respectively applies nonlinear numerical suppression processing to various indexes in the three-dimensional score index set, interval stretching is performed on the basic risk index through a normalization function, gradient suppression is performed on the dynamic threat index through an exponential transformation, and saturation value limitation is performed on the regional vulnerability index through an upper limit suppression function. After the score construction unit performs the nonlinear processing, a set of weight factors is called to perform weighted combination calculation on the three processed indexes to construct a node priority score feature set.

[0012] Optionally, the control objects in the step six include gates, pump stations, dams and emergency response devices, and the control objects perform opening and closing, pumping and draining, water releasing and defense deploying operations according to the dispatch action sequence, and collect corresponding state information to generate a dispatch execution feedback data set.

[0013] Optionally, the dynamic updating and reordering operation in the step seven includes: After receiving the dispatch execution feedback data set, the water level change increment, the flow rate change amplitude and the execution response state identifier of each node are extracted, the maximum value of the basic risk index of each node in the node priority score set is updated, the difference value of the dynamic threat index is replaced, and the regional vulnerability index is corrected in proportion to generate an updated node priority score set; The score value in the updated node priority score set is matched with the target node number bound to the action in the original dispatch action sequence, a new score field is added to each action in the dispatch action sequence, the dispatch action sequence is arranged in descending order according to the score value, and the dispatch action sequence is regenerated; For the nodes marked as execution failure in the dispatch execution feedback data set, the record with the response state identifier as "failure" is extracted, the dispatch path accessibility feature field in the edge feature vector on the corresponding path is set to 0, the resource bottleneck feature field of the corresponding edge is decremented, the execution constraint feature field of the target node is increased, an updated feasibility enhanced convolution structure is formed, the reordered dispatch action sequence and the updated feasibility enhanced convolution structure are jointly input, and a final dispatch action instruction set is generated.

[0014] The flood priority dispatch intelligent evaluation system for multi-modal flood warning according to the embodiment of the application includes the following modules: The multi-modal data preprocessing module is used to collect rainfall, water level and flow rate data, and complete time alignment, abnormality elimination and normalization processing; The graph structure construction module is used to construct the preprocessed data into a river network graph structure, extract node features and edge features, and generate graph structure input data; An evolutionary graph convolution processing module is configured to receive graph structure input data, perform an EvolveGCN-based graph convolution operation, and output a state propagation feature set; A state prediction generation module is configured to generate a node state prediction vector according to the state propagation feature set, representing a future hydrological state of the node; A priority score construction module is configured to calculate three types of indexes and impose a nonlinear constraint to generate a node priority score set; A scheduling action generation module is configured to generate a scheduling action sequence according to the score set and issue the scheduling action sequence to each control object for execution; A feedback processing and reordering module is configured to analyze feedback data, update the score set and the scheduling sequence, and output a final scheduling action instruction set.

[0015] The present application has the following advantages: (1) The evolutionary graph convolution network structure is introduced, the node and edge features of the river network graph structure are combined, the expression ability for the multi-source hydrological state evolution process is enhanced on the basis of time continuity modeling, the dynamic prediction of the node water level, flow and flood peak arrival time can be realized, and the timeliness and accuracy of the flood evolution simulation are improved.

[0016] (2) The design of the enhanced convolution structure embeds the scheduling path accessibility feature, the resource bottleneck feature and the execution constraint feature in the state propagation process, avoids the problem that the scheduling action is invalid on the unexecutable path, realizes the structural constraint filtering and path executable guarantee of the scheduling action, and improves the landing rate of the control instruction.

[0017] (3) A three-dimensional priority score mechanism including a basic risk, a dynamic threat and a regional vulnerability is constructed, the nonlinear function processing and the dynamic reordering logic driven by the feedback data are combined, a complete score-scheduling-feedback closed loop chain is formed, and the system has the real-time updating, intelligent response and adaptive scheduling capability. BRIEF DESCRIPTION OF DRAWINGS

[0018] The accompanying drawings are included to provide a further understanding of the application and constitute a part of the specification, illustrate the application, and are used together with the embodiments of the application to explain the application, and do not constitute a limitation on the application. In the drawings: Figure 1 The overall flowchart of the flood priority scheduling intelligent evaluation method for multi-modal flood warning proposed by the present application; Figure 2 The module connection diagram of the flood priority scheduling intelligent evaluation system for multi-modal flood warning proposed by the present application. DETAILED DESCRIPTION

[0019] The application will now be described in further detail with reference to the drawings. These drawings are simplified schematic illustrations of the basic structure of the application and therefore only show what is relevant to the present application.

[0020] Reference Figure 1 and Figure 2 The flood priority scheduling intelligent evaluation method and system for multi-modal flood warning include the following steps: Step one, collect rainfall, water level and flow rate data, and preprocess to generate standardized multi-modal input data; Step two, build a river network graph structure, represent observation points as nodes and river connections as edges, extract node and edge features, and generate graph structure input data; Step three, input the graph structure input data into the evolved graph convolution network based on EvolveGCN, propagate node state features using the feasibility-enhanced convolution structure, embed scheduling path reachability, resource bottleneck and execution constraint features during convolution, and output a set of state propagation features; Step four, generate a node state prediction vector based on the set of state propagation features; in the specific implementation process, the set of state propagation features is output by the evolved graph convolution network after propagation in a multi-time step structure, and the state propagation features corresponding to each node are used as the graph space embedding representation of the node at the current time sequence. The system extracts and analyzes the node feature vectors in the set of state propagation features, maps and predicts the water level value, flow rate value and flood peak arrival time through a multi-layer perception unit, and outputs a node state prediction vector. The node state prediction vector contains multi-dimensional hydrological response information within a continuous time window, providing a basis for subsequent risk scoring and scheduling priority determination. The entire prediction process has time sensitivity and structural dependence, and can dynamically adapt to graph structure evolution and upstream and downstream influence changes.

[0021] Step five, calculate three types of indexes: basic risk, dynamic threat and regional vulnerability based on the node state prediction vector, apply a nonlinear constraint mechanism including an upper limit suppression function and a normalization function to build a priority score feature set, and generate a node priority score set; Step six, generate a scheduling action sequence based on the node priority score set, send the scheduling action sequence to the control object for execution, and collect water level change data, flow rate change data and response state information of the corresponding node during execution to build a scheduling execution feedback data set; Step seven, trigger dynamic update and reordering operations of the node priority score set and the scheduling action sequence based on the scheduling execution feedback data set, update the feasibility-enhanced convolution structure based on the feedback data set, and generate a final scheduling action instruction set.

[0022] In this embodiment, the preprocessing in step one includes performing time alignment, missing data completion, anomaly removal and normalization processing on rainfall, water level and flow rate data to generate standardized multi-modal input data.

[0023] In this embodiment, the evolved graph convolutional network constructed based on EvolveGCN in step three includes a graph convolutional parameter initialization unit, a graph convolutional parameter evolution unit, a gated recurrent unit, a node state update unit, a historical state fusion unit, a state storage unit and an output scheduling unit, specifically: In the graph convolution initialization stage, the graph structure input data is received by the graph convolutional parameter initialization unit at the input end of the evolved graph convolutional network, the graph structure input data is divided into several time step graph structure snapshots, and the node state propagation weight matrix and bias term of each layer of graph convolution structure are initialized based on the time step graph structure snapshot; At the beginning of each time step, the convolution weight set of the last time step is called in the parameter evolution path by the graph convolutional parameter evolution unit, combined with the node state change information of the current time step graph structure snapshot, and the parameter update operation is performed through the gated recurrent unit to generate the convolution weight set of the current time step; In specific implementation, the gated recurrent unit receives the convolution weight set from the last time step as the historical state input, and encodes the state change information of each node in the current time step graph structure snapshot into the current input state, and the unit is based on the internal structure of the reset gate and the update gate to control the reservation ratio of the historical weight and the activation degree of the current input respectively; In each update, the gated recurrent unit performs weighted combination and state update operation on the parameter matrix of each convolution layer to generate a convolution weight set matching the structure characteristics of the current time step, so that the graph convolutional network has the ability to adapt to the time sequence, and the update process is executed in time sequence in the graph convolutional parameter evolution path to ensure that the convolution parameters continuously respond to the evolution of the graph structure state.

[0024] In the node state update unit, the historical state propagation feature set of each graph structure node is called, the graph structure convolution operation is performed according to the convolution weight set of the current time step and the adjacent edge weight information, and the node state propagation feature is calculated, and the operation is completed in parallel on all nodes in the graph structure; In specific implementation, the graph structure convolution operation takes the convolution weight set of the current time step as the kernel function, and performs weighted aggregation on the adjacent node state propagation features of each node, and in the aggregation process, the adjacent edge weight information is combined to perform weighted accumulation calculation in the edge direction, and the state information of the adjacent node is mapped to the target node. After receiving the aggregation result, the target node performs fusion transformation with its historical state propagation feature to form the node state propagation feature of the current time step, and the convolution operation is expanded synchronously on all nodes in the graph structure to ensure that the overall state propagation process has spatial consistency and structure constraint.

[0025] After state propagation, the historical state fusion unit calls the state propagation features of the past time steps of the node in the state propagation path of the current time step, and completes the historical information fusion by splicing, linear fusion or residual connection to generate the fused node state propagation feature representation; in specific implementation, in the historical information fusion process, the historical state fusion unit first arranges the state propagation features of the target node at multiple time steps in time sequence; when splicing is used, the feature vectors at different times are directly connected in the channel dimension to expand the feature expression dimension; when linear fusion is used, the feature vectors at each time step are weighted and summed according to the set weight coefficient to generate a smooth transition fusion vector; when residual connection is used, the state propagation feature of the current time step is added to the feature vector of the previous time step at the element level to retain the change residual in the evolution direction. The above operations can be flexibly configured to adapt to different structural evolution rates and information decay degrees, ensuring that the node fused features retain historical information and have timeliness.

[0026] After each time step ends, the state storage unit records the state propagation features of all nodes at the current time step in the time sequence feature cache area as the input basis for the next time step graph convolution parameter evolution and state fusion; After all time steps are executed, the output scheduling unit integrates the node state propagation feature sets generated at each time step in sequence to output the final state propagation feature set.

[0027] In this embodiment, the feasibility enhancement convolution structure in step three includes a path reachability regulation unit, a resource bottleneck weight regulation unit and an execution constraint activation unit, specifically: The path reachability regulation unit calls the scheduling path reachability feature in the graph structure input data before the node state update unit performs the graph structure convolution operation, performs path reachability discrimination processing on the adjacent nodes of each graph structure node, removes the unreachable adjacent nodes from the feature aggregation path, and retains the reachable adjacent nodes to form a state propagation input set; in specific implementation, the reachability discrimination processing is performed by reading and analyzing the scheduling path reachability feature field associated with each edge in the graph structure input data. The field records whether the nodes have an effective control path connection relationship, and the value domain is usually represented by a Boolean value or a continuous score. In the discrimination process, the path reachability regulation unit iteratively traverses the adjacent edges of each node, performs threshold judgment or logical judgment operation on the reachability field on the edge, removes the corresponding adjacent node from the feature aggregation path of the current node if the judgment result is unreachable, and retains it for subsequent state propagation if it is reachable. This processing step ensures that information transmission is only performed between nodes that have effective connections in physics or control, improving the structural rationality and scheduling executability of the propagation path.

[0028] The resource bottleneck weight adjustment unit calls the field representing the resource capacity in the graph structure input data in the graph structure convolution operation process, performs a propagation channel weight adjustment operation, maps the resource bottleneck feature value corresponding to the edge to a propagation weight adjustment factor, and adjusts the weight value of the corresponding edge in the convolution weight set by edge; in the propagation channel weight adjustment operation process, the resource bottleneck weight adjustment unit extracts the resource bottleneck feature value from the resource capacity field of each edge in the graph structure input data, the feature reflects the limitation degree of the transmission capacity of the edge, and the system converts the resource bottleneck value into a propagation weight adjustment factor through a preset mapping function, such as a normalization or inverse ratio function. The propagation weight adjustment factor acts on the convolution weight of each edge in the graph structure convolution process, performs edge-by-edge multiplication adjustment, and the adjusted edge weight directly affects the contribution degree of the adjacent node feature in the convolution aggregation, thereby suppressing the information transmission intensity on the bottleneck edge, preferentially guiding the state information to propagate through the resource-rich path, and improving the scheduling rationality and stability of the overall propagation path.

[0029] The execution constraint activation unit compares the execution constraint state of each node with the set threshold after generating the state propagation feature, marks the nodes that do not meet the condition and stops the output of the state propagation feature of the nodes at the current time step, and retains the state propagation feature of the nodes that meet the condition. The comparison operation between the execution constraint state and the set threshold is triggered by the execution constraint activation unit after the state propagation is completed, the system reads the execution constraint feature field corresponding to each node in the graph structure input data, the field quantifies the controllability level of the node in the current scheduling period, and is represented by an integer score or a Boolean identifier. The activation unit compares the value with the preset threshold node by node, when the execution constraint state is lower than the threshold, the node is determined to not have a scheduling execution condition, is marked as a disabled state, and the output of the state propagation feature at the current time step is stopped; when the threshold condition is met, the node state feature is normally retained and used for subsequent processing. This comparison mechanism ensures that only nodes with controllable resources participate in scheduling decisions, improving the landability and reliability of the scheduling result.

[0030] In the embodiment, the nonlinear constraint mechanism in step five is performed by the score construction unit, specifically including: After receiving the node state prediction vector, the score construction unit extracts the basic risk index for describing the flood intensity, the dynamic threat index for describing the state change rate, and the regional vulnerability index for describing the population exposure degree, to form a three-dimensional score index set. The score construction unit respectively applies nonlinear numerical suppression processing to each type of index in the three-dimensional score index set. The basic risk index performs interval stretching through a normalization function. The dynamic threat index performs gradient suppression through an exponential transformation. The regional vulnerability index performs saturation value limitation through an upper limit suppression function. In the processing of the basic risk index, the score construction unit adopts a minimum-maximum normalization method to linearly map the index values to a set interval range, which is [0, 1], to eliminate the numerical scale differences between different nodes and highlight the risk proportion differences. The dynamic threat index is processed through an exponential transformation, which is wherein is the state change rate, is the adjustment coefficient. This way gives higher response sensitivity to nodes with rapid changes and suppresses the score influence of nodes with low fluctuations. The regional vulnerability index is clipped by setting a fixed upper threshold. When the original index value exceeds the upper limit, it is directly assigned to the threshold value, preventing high population exposure or high infrastructure density nodes from being infinitely amplified in the score and avoiding imbalance in scheduling priority. The three types of nonlinear processing methods respectively address the characteristics of different types of indexes to enhance the stability and discriminability of the scoring system.

[0031] After performing nonlinear processing, the score construction unit calls the set weight factor set to perform weighted combination calculation on the three processed indexes to construct the node priority score feature set.

[0032] In this embodiment, the control objects in step six include gates, pump stations, dams, and emergency response equipment. The control objects perform opening and closing, pumping and draining, water release, and defense deployment operations according to the scheduling action sequence, and collect corresponding state information to generate a scheduling execution feedback data set.

[0033] In this embodiment, the dynamic updating and reordering operation in step seven includes: After receiving the scheduling execution feedback data set, the water level change increment, flow rate change amplitude, and execution response state identifier of each node are extracted. The maximum value of the basic risk index of each node in the node priority score set is updated. The difference value of the dynamic threat index is replaced. The regional vulnerability index is corrected by scaling. An updated node priority score set is generated. The score value in the updated node priority score set is matched with the target node number bound to the action in the original scheduling action sequence. A new score field is added to each action in the scheduling action sequence. The actions are arranged in descending order according to the score values, and a new scheduling action sequence is generated. For the nodes marked as failed execution in the scheduling execution feedback data set, extract the records with the response status identified as "failure", set the scheduling path reachability feature field in the edge feature vector on the corresponding path to 0, perform a decremental update operation on the resource bottleneck feature field of the corresponding edge, and perform an upward operation on the execution constraint feature field of the target node to form an updated feasibility-enhanced convolution structure. The reordered scheduling action sequence is jointly input with the updated feasibility-enhanced convolution structure to generate the final scheduling action instruction set.

[0034] The flood priority scheduling intelligent assessment system for multimodal flood warning according to an embodiment of the present invention includes the following modules: Multimodal data preprocessing module, used to collect rainfall, water level and flow rate data, and complete time alignment, anomaly removal and normalization processing; The graph structure construction module is used to construct the preprocessed data into a river network graph structure, extract node features and edge features, and generate graph structure input data; The evolutionary graph convolution processing module is used to receive graph structure input data, perform graph convolution operations based on EvolveGCN, and output a set of state propagation features; The state prediction generation module is used to generate a node state prediction vector based on the state propagation feature set to represent the future hydrological state of the node; The priority scoring building module is used to calculate three types of indicators and impose nonlinear constraints to generate a set of node priority scores; The scheduling action generation module is used to generate a scheduling action sequence based on the score set and send it to each control object for execution; The feedback processing and reordering module is used to parse feedback data, update the scoring set and scheduling sequence, and output the final scheduling action instruction set.

[0035] Example 1: In order to verify the feasibility of the present invention in implementation, the present invention is applied to flood warning and linkage scheduling tasks in a multi-basin intersection area in a certain place. The terrain in this area changes dramatically, covering multiple upstream medium-sized tributaries and downstream main drainage channels. There is a risk of superposition of water from multiple sources during the flood season all year round. There are problems such as untimely peak warnings, inaccessible scheduling paths, failed pump station responses, and delayed command execution, which restricts the operational stability of the regional flood control system.

[0036] In the deployment process, the multi-modal sensing system synchronously collects the data of rainfall monitoring points, water level monitoring points and flow rate measurement sections in the region, all the data are uniformly subjected to time alignment, abnormality elimination and normalization operation in the system to form standardized input data, the river network graph structure is constructed by taking the standardized data as input, the nodes in the graph correspond to the monitoring points and control nodes, and the edges represent the connectivity relationship of the river channels, and the feature vectors are assigned to each node and edge, including real-time observation values, terrain attributes, historical evolution trend and control attributes.

[0037] After the construction is completed, the graph structure is input into the evolved graph convolutional network constructed based on EvolveGCN, at each time step, the state propagation path is dynamically updated through the graph convolution parameter evolution unit, in the process, the feasibility enhanced convolution structure is called to perform real-time inhibition and activation processing on the reachability of the scheduling path, resource bottleneck constraint on the edge and execution controllability of the node, and only the state change is propagated on the adjustable path, and finally the node state prediction result is output, including the predicted water level, predicted flood peak arrival time and flow rate evolution trend.

[0038] In the priority score construction module, the system calculates three score dimensions of basic risk, dynamic threat and regional vulnerability according to the state prediction result, in order to prevent a single index from dominating the ranking, the system respectively applies a normalization function and an upper limit inhibition function to each index, and constructs a nonlinear weighted priority score model, and each scheduling object such as a pump station, a gate, a dam, an emergency device and the like enters a scheduling action sequence according to the score result.

[0039] The system continuously receives feedback data in the actual execution process, including water level change rate, scheduling response delay, execution result code and the like, and updates the score set and the scheduling sequence in real time, when the water level rise of a certain node exceeds the original estimation, the system automatically increases the risk score of the node; if a path fails continuously for two times, the system sets the corresponding edge as unreachable, the scheduling path is forced to be interrupted, and is eliminated in the next generation. When the execution response delay exceeds the set threshold, the system increases the execution constraint intensity of the target node, and reduces the probability of future repeated scheduling.

[0040] In the actual trial operation stage, 202 historical flood process data are selected for comparison test, and the comparison mode adopts the full-flow simulation of the system and the existing traditional scheduling method. The following is a comparison effect data table, which shows the advantages of the application from the angles of prediction accuracy, response delay, scheduling coverage rate, execution success rate and the like.

[0041] Table 1: Performance comparison table of flood priority scheduling intelligent evaluation system and traditional scheduling scheme

[0042] In terms of node state prediction, the water level prediction mean square error of the system is 0.172 meters, which is significantly better than the 0.381 meters of the traditional scheme, with an error reduction of more than half, indicating that the use of evolutionary graph convolutional network for hydrological evolution simulation is more accurate. In the prediction of flood peak arrival time, the system controls the error within 6.2 minutes, while the average error of the traditional scheme is 15.7 minutes, with a prediction accuracy improvement of nearly 60%, which helps to respond to pre-dispatching.

[0043] The system response speed shows obvious differences. The system realizes an average dynamic score update period of 0.44 seconds, while the traditional scheme does not support score updating and is difficult to adapt to changes in flood situation. In terms of average delay in dispatch response, the system is only 32.1 seconds, while the traditional scheme is 85.6 seconds, with a response timeliness improvement of more than 60%, ensuring that critical dispatch instructions can be executed in time within the window period.

[0044] In terms of dispatch reliability, the instruction execution success rate of the system reaches 96.7%, which is much higher than the 81.2% of the traditional scheme, and the feasibility structure field can be updated in real time according to the execution result. During the trial operation, the cumulative number of field updates reached 1829 times, showing strong adaptive scheduling capability. The high-risk node coverage rate and multi-round dispatch hit rate reached 93.8% and 87.1% respectively, compared with 76.5% and 62.9% of the traditional scheme, enhancing the protection capability of the dispatch to key areas and the accuracy of multi-round linkage.

[0045] For the response to sudden failure paths, the system can complete path elimination and structure adjustment within an average of 1.27 seconds, while the traditional scheme does not have this capability, showing the advantage of structure closed-loop control. Finally, in terms of system running efficiency, the average time consumption of the system for single-round full-graph calculation is 28.4 seconds, which is much lower than the 119.6 seconds of the traditional scheme, with a processing efficiency improvement of more than 76%.

[0046] Overall, the system is superior to the traditional dispatching scheme in terms of prediction accuracy, response efficiency, execution stability and structure adaptive capability, and can support intelligent dispatching tasks in complex basins and improve the dispatching closed-loop capability of multi-modal flood warning.

[0047] The above is only the preferred specific embodiment of the present application, but the protection scope of the present application is not limited thereto. Any skilled person in the art can make equivalent substitutions or changes to the technical solutions and inventive concepts of the present application within the technical scope disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. An intelligent evaluation method for flood priority scheduling for multimodal flood warning, characterized by: The steps include: Step 1: Collect rainfall, water level and flow rate data, and preprocess them to generate standardized multimodal input data; Step 2: Construct a river network graph structure, represent observation points as nodes, represent river connections as edges, extract node and edge features, and generate graph structure input data; Step 3: Input the graph structure input data into the evolutionary graph convolutional network built based on EvolveGCN. A feasibility-enhanced convolutional structure is used to propagate node state features. The scheduling path reachability, resource bottleneck, and execution constraint features are embedded in the convolution process, and a state propagation feature set is output. Step 4: Generate a node state prediction vector based on the state propagation feature set; Step 5: Calculate three indicators of basic risk, dynamic threat, and regional vulnerability based on the node state prediction vector. Apply a nonlinear constraint mechanism including an upper limit suppression function and a normalization function to construct a priority score feature set and generate a node priority score set. Step 6: Generate a scheduling action sequence based on the node priority score set, send the scheduling action sequence to the control object for execution, collect water level change data, flow rate change data and response status information of the corresponding node during the execution process, and build a scheduling execution feedback data set; Step 7: Trigger the dynamic update and reordering of the node priority score set and the scheduling action sequence based on the scheduling execution feedback data set, update the feasibility enhancement convolution structure based on the feedback data set, and generate the final scheduling action instruction set.

2. The intelligent evaluation method for flood priority scheduling for multimodal flood warning according to claim 1 is characterized in that: The preprocessing in step 1 includes: performing time alignment, missing completion, anomaly removal and normalization processing on rainfall, water level and flow rate data to generate standardized multimodal input data.

3. The intelligent evaluation method for flood priority scheduling for multimodal flood warning according to claim 2 is characterized in that: The evolutionary graph convolutional network constructed based on EvolveGCN in step 3 includes a graph convolution parameter initialization unit, a graph convolution parameter evolution unit, a gated recursive unit, a node state update unit, a historical state fusion unit, a state storage unit, and an output scheduling unit, specifically: In the graph convolution initialization phase, the graph convolution parameter initialization unit receives graph structure input data at the input end of the evolved graph convolution network, divides the graph structure input data into several time-step graph structure snapshots, and initializes the node state propagation weight matrix and bias term for each layer of the graph convolution structure based on the time-step graph structure snapshots; At the beginning of each time step, the graph convolution parameter evolution unit calls the convolution weight set of the previous time step in the parameter evolution path, combines the node state change information of the graph structure snapshot of the current time step, and performs parameter update operation through the gated recursive unit to generate the convolution weight set of the current time step; In the convolution propagation phase at each time step, the node state update unit calls the historical state propagation feature set for each graph structure node, performs the graph structure convolution operation based on the convolution weight set of the current time step and the adjacent edge weight information, and calculates the node state propagation feature. The operation is completed in parallel for all nodes in the graph structure; After state propagation, the historical state fusion unit calls the state propagation features of the node in the past time steps in the state propagation path of the current time step, and uses splicing, linear fusion or residual connection to complete the historical information fusion and generate the fused node state propagation feature representation; After each time step, the state storage unit records the state propagation features of all nodes in the current time step in the time series feature buffer, which serves as the input basis for the graph convolution parameter evolution and state fusion in the next time step. After all time steps are completed, the output scheduling unit sequentially integrates the node state propagation feature sets generated by each time step and outputs the final state propagation feature set.

4. The intelligent evaluation method for flood priority scheduling for multimodal flood warning according to claim 3 is characterized in that: The feasibility enhancement convolutional structure in step 3 includes a path reachability control unit, a resource bottleneck weight adjustment unit, and an execution constraint activation unit, specifically: Before the node state update unit performs the graph structure convolution operation, the path reachability control unit calls the scheduling path reachability features in the graph structure input data, performs path reachability discrimination processing on the adjacent nodes of each graph structure node, removes unreachable adjacent nodes from the feature aggregation path, and retains reachable adjacent nodes to form the state propagation input set; During the graph convolution operation, the resource bottleneck weight adjustment unit calls the field representing the resource capacity in the graph input data, performs the propagation channel weight adjustment operation, maps the resource bottleneck feature value corresponding to the edge into the propagation weight adjustment factor, and adjusts the weight value of the corresponding edge in the convolution weight set on an edge-by-edge basis. After the state propagation feature is generated, the execution constraint activation unit calls the execution constraint feature field in the graph structure input data, compares the execution constraint state of each node with the set threshold, marks the nodes that do not meet the conditions and terminates the state propagation feature output of the node in the current time step, and retains the state propagation features of the nodes that meet the conditions.

5. The intelligent evaluation method for flood priority scheduling for multimodal flood warning according to claim 4 is characterized in that: The nonlinear constraint mechanism in step 5 is executed by the scoring construction unit, specifically including: After receiving the node state prediction vector, the scoring construction unit extracts the basic risk index used to describe the flood intensity, the dynamic threat index used to describe the state change rate, and the regional vulnerability index used to describe the population exposure degree, forming a three-dimensional scoring index set. The scoring construction unit applies nonlinear numerical suppression to each indicator in the three-dimensional scoring indicator set. The basic risk indicator performs interval stretching through a normalization function, the dynamic threat indicator performs gradient suppression through an exponential transformation, and the regional vulnerability indicator performs saturation value restriction through an upper limit suppression function. After performing nonlinear processing, the scoring construction unit calls the set weight factor set to perform weighted combination calculation on the three processed indicators to construct a node priority scoring feature set.

6. The intelligent evaluation method for flood priority scheduling for multimodal flood warning according to claim 5 is characterized in that: The control objects in step six include gates, pumping stations, dams and emergency response equipment. The control objects perform opening and closing, pumping, water release and deployment operations according to the scheduling action sequence, and collect corresponding status information to generate a scheduling execution feedback data set.

7. The intelligent evaluation method for flood priority scheduling for multimodal flood warning according to claim 6 is characterized in that: The dynamic update and reordering operations in step seven include: After receiving the scheduling execution feedback data set, the water level change increment, flow velocity change amplitude and execution response status identifier corresponding to each node are extracted, the basic risk indicator of each node in the node priority score set is updated to the maximum value, the dynamic threat indicator is replaced by the difference, and the regional vulnerability indicator is scaled and corrected to generate an updated node priority score set; Match the score values ​​in the updated node priority score set with the target node numbers bound to the actions in the original scheduling action sequence, add a new score field to each action in the scheduling action sequence, sort them in descending order according to the score values, and regenerate the scheduling action sequence; For nodes marked as failed in the scheduling execution feedback data set, records with a response status of "failed" are extracted. The scheduling path reachability feature field in the edge feature vector on the corresponding path is set to 0, the resource bottleneck feature field of the corresponding edge is decremented, and the execution constraint feature field of the target node is incremented to form an updated feasibility-enhanced convolutional structure. The reordered scheduling action sequence and the updated feasibility-enhanced convolutional structure are jointly input to generate the final scheduling action instruction set.

8. A flood priority dispatch intelligent assessment system for multimodal flood early warning, applied to the flood priority dispatch intelligent assessment method for multimodal flood early warning according to any one of claims 1 to 7, characterized in that: Includes the following modules: Multimodal data preprocessing module, used to collect rainfall, water level and flow rate data, and complete time alignment, anomaly removal and normalization processing; The graph structure construction module is used to construct the preprocessed data into a river network graph structure, extract node features and edge features, and generate graph structure input data; The evolutionary graph convolution processing module is used to receive graph structure input data, perform graph convolution operations based on EvolveGCN, and output a set of state propagation features; The state prediction generation module is used to generate a node state prediction vector based on the state propagation feature set to represent the future hydrological state of the node; The priority scoring building module is used to calculate three types of indicators and impose nonlinear constraints to generate a set of node priority scores; The scheduling action generation module is used to generate a scheduling action sequence based on the score set and send it to each control object for execution; The feedback processing and reordering module is used to parse feedback data, update the scoring set and scheduling sequence, and output the final scheduling action instruction set.

Images