A city flood disaster monitoring and early warning system and method

Abstract

The application discloses a kind of urban flood disaster monitoring and early warning system and method, it is related to monitoring and early warning field, including: using three-layer directional hydraulic atlas to configure urban flood water power topology structure, configuration node is to bear urban flood key hydraulic state, configuration directed edge is to express water flow direction and flux capacity, set spatial proximity relationship mapping multi-source observation data to node, set the unified expression and value range of node state quantity.Through three-layer directional hydraulic atlas, multi-source observation is unified to node / edge representation, and with short sliding window recovery rate and variance rise and other indicators to build node critical signs, can identify nonlinear critical evolution before traditional absolute threshold triggers, finally with MIP to solve minimum intervention set and issue to edge execution unit and do post-validation, to realize lower false alarm rate, higher response specificity and quantifiable risk reduction effect under limited resources.

Description

Technical Field

[0001] This invention relates to the field of monitoring and early warning, and in particular to an urban flood disaster monitoring and early warning system and method. Background Technology

[0002] Urban flood monitoring and early warning require simultaneously characterizing surface runoff, underground drainage networks, and the backflow / overflow coupling relationship between the two to reflect the physical connectivity and potential energy driving characteristics of water flow. Existing projects typically use single-layer grids or isolated pipe network models, and multi-source observations are difficult to effectively integrate on a unified topology, resulting in insufficient joint judgment capabilities on propagation paths, node-driven potential energy, and observation uncertainties. Based on a three-layer directed hydraulic topology, multi-source data can be mapped to nodes and directed edges, allowing water depth, inflow, pipe network saturation, and observation confidence to be expressed within the same framework and used for subsequent risk assessment and scheduling.

[0003] However, there are still two problems with the existing technology. First, most current early warning systems rely on absolute thresholds or single observation indicators, which are difficult to provide reliable precursors before nonlinear transitions occur. When the system shows short-term dynamic signs such as slowed recovery rate or increased variance, they are often ignored. Second, when observation resources are limited, existing scheduling is mostly based on risk level or static coverage priority, without simultaneously coupling and prioritizing propagation capability and information uncertainty, resulting in wasted or mismatched observation and intervention resources. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by providing an urban flood disaster monitoring and early warning system and method. This system aims to estimate water depth recovery rate, autoregressive trends, and variance changes at MDHG nodes using a short sliding window, and to construct a critical symptom index for LCSD nodes. This allows for the identification of early warning signs of critical deterioration before exceeding thresholds. The system calculates propagation potential using node hydraulic potential energy and transmission impedance, and couples this with observation / model / public reporting uncertainties to form a PUC value. This PUC value is used to prioritize observation and intervention targets, thereby updating probabilities and causal attributions within online latency, ultimately driving integer programming and edge-loop execution of the minimum intervention path.

[0005] Therefore, this application provides an urban flood disaster monitoring and early warning system and method, including the following steps:

[0006] Step S100: Configure the urban flood hydraulic topology using a three-layer directional hydraulic map, configure nodes to carry the key hydraulic states of urban flooding, configure directed edges to express the direction of water flow and flux capacity, set spatial proximity relationships to map multi-source observation data to nodes, and set a unified expression and value range for node state variables.

[0007] Step S200: Obtain the node water depth time series, configure a short sliding window to perform recovery rate estimation, configure node fluctuation and vulnerability indicators, configure node critical symptom index, set measurement confidence constraints on the calculation conditions of the node critical symptom index, and perform node critical state verification. Based on the node critical state, perform linkage control and state marking.

[0008] Step S300: The calculation method of using nodal hydraulic potential energy as the propagation driving force is adopted. The propagation potential of the node is calculated by using the hydraulic potential energy difference and the transmission impedance. The node uncertainty metric is configured and calculated to reflect the confidence of observation and model. The coupling ranking value PUC is configured and the selection rules under resource constraints are configured.

[0009] Step S400: Set up the observation resource set and define the observation resource attributes, configure the total amount of information uncertainty to quantify the observation value base, use marginal information gain to measure the value of a single observation, use benefit-cost ratio to prioritize observation tasks, execute greedy assignment, and update the observation plan online.

[0010] Step S500: Configure the Counterfactual Graph Neural Predictor (CGNP), generate a counterfactual scene set, label node flooding and path attribution, set the prediction output, set the training loss, and supervise node prediction, path attribution and confidence calibration. Define the interpretable attribution calculation of edge and path contribution, and define the online conditional update and linkage rules with VoITask.

[0011] Step S600: Configure the intervention candidate set, use CGNP to estimate the marginal benefit of the intervention, set the minimum intervention path integer programming expression, set the interactive correction of the impact of the intervention combination, and set the edge closed-loop distribution and local rollback execution mechanism.

[0012] In some specific embodiments, step S100 specifically includes:

[0013] Step S100.1: Use a three-layer directional hydraulic map to configure the urban flood hydraulic topology structure, configure nodes to carry the key hydraulic states of urban flooding, and configure directed edges to express the direction of water flow and flux capacity.

[0014] Step S100.2: Set the spatial proximity relationship to map multi-source observation data to nodes, and set the unified expression and value range of node state variables.

[0015] The water depth of the nodes is expressed in meters, the inflow rate in cubic meters per second, the saturation of the drainage network is expressed as a continuous value between 0 and 1, and the measurement confidence level is expressed as a continuous value between 0 and 1, where the larger the value, the higher the reliability of the corresponding state quantity or observation result.

[0016] The measurement confidence level stored in the nodes is used to constrain subsequent flood analysis and early warning decisions. When the measurement confidence level of a node is less than or equal to 0.6, the weight of that node in the flood risk assessment is reduced.

[0017] When the measurement confidence level is less than or equal to 0.3 and the duration exceeds 10 minutes, the node is marked as a low-confidence node to avoid unreliable observations from misleading the overall flood warning results.

[0018] In some specific embodiments, step S200 specifically includes:

[0019] Step S200.1: Obtain the node water depth time series, configure a short sliding window to perform recovery rate estimation, configure node volatility and vulnerability indicators, and configure node critical symptom index.

[0020] Step S200.2: Set the calculation conditions for the critical symptom index of the measurement confidence constraint node, and perform node critical state verification. Based on the node critical state, perform linkage control and state marking.

[0021] In some specific embodiments, step S300 specifically includes:

[0022] Step S300.1: The calculation method of using the node hydraulic potential energy as the propagation driving force is adopted, and the node propagation potential energy is calculated by using the hydraulic potential energy difference and the transmission impedance.

[0023] Step S300.2: Configure and calculate node uncertainty metrics to reflect observation and model confidence, configure coupling ranking value PUC, and configure selection rules under resource constraints.

[0024] In some specific embodiments, step S400 specifically includes:

[0025] Step S400.1: Set up the observation resource set, define the observation resource attributes, configure the total amount of information uncertainty to quantify the observation value base, and use marginal information gain to measure the value of a single observation.

[0026] Step S400.2: Prioritize observation tasks by comparing benefits and costs, execute greedy assignment, and update the observation plan online.

[0027] In some specific embodiments, step S500 specifically includes:

[0028] Step S500.1: Configure the counterfactual graph neural predictor, generate a counterfactual scene set, label node flooding and path attribution, set the prediction output, set the training loss, and supervise node prediction, path attribution, and confidence calibration.

[0029] Step S500.2: Define the interpretable attribution calculation of edge and path contribution, and define the online conditional update and linkage rules with VoITask.

[0030] In some specific embodiments, step S600 specifically includes:

[0031] Step S600.1: Configure the intervention candidate set, use CGNP to estimate the marginal benefit of the intervention, and set the minimum intervention path integer programming expression.

[0032] Step S600.2: Set the interactive correction of the impact of the intervention combination, and set the edge closed-loop distribution and local rollback execution mechanism.

[0033] If multiple interventions have significant interaction effects, the combined marginal benefit is denoted as... Furthermore, the optimization objective is extended to a quadratic or mixed integer programming problem that includes interactive correction terms to avoid double-counting.

[0034] The optimized set of interventions is distributed through edge execution units: these units include pump station SCADA control modules, valve remote control units, emergency vehicle command terminals, and on-site emergency teams for each selected intervention. Before execution, a local security check must be performed and recorded. :

[0035] like =If rollback is possible, then the executed command includes rollback conditions and rollback instructions. If the error cannot be reversed, then additional manual confirmation or automatic threshold verification is required.

[0036] An urban flood disaster monitoring and early warning system includes the following modules:

[0037] The three-layer hydraulic module is used to configure the urban flood hydraulic topology using a three-layer directional hydraulic map, configure nodes to carry the key hydraulic states of urban flooding, configure directed edges to express the direction of water flow and flux capacity, set spatial proximity relationships to map multi-source observation data to nodes, and set a unified expression and value range for node state quantities.

[0038] The node symptom module is used to acquire the node water depth time series, configure a short sliding window to perform recovery rate estimation, configure node fluctuation and vulnerability indicators, configure the node critical symptom index, set the measurement confidence constraint for the calculation conditions of the node critical symptom index, perform node critical state verification, and perform linkage control and state marking based on the node critical state.

[0039] The propagation potential coupling module is used to calculate the propagation potential of nodes by using the hydraulic potential energy of nodes as the propagation driving force, and to calculate the propagation potential of nodes by using the hydraulic potential energy difference and transmission impedance. It also configures and calculates the node uncertainty metric to reflect the confidence of observations and models, configures the coupling ranking value PUC, and configures the selection rules under resource constraints.

[0040] The information value scheduling module is used to set up the observation resource set, define the attributes of the observation resources, configure the total amount of information uncertainty to quantify the base of observation value, use marginal information gain to measure the value of a single observation, use benefit-cost ratio to prioritize observation tasks, execute greedy assignment, and update the observation plan online.

[0041] The causal attribution module is used to configure the Counterfactual Graph Neural Predictor (CGNP), generate a counterfactual scenario set, label node flooding and path attribution, set the prediction output, set the training loss, and supervise node prediction, path attribution and confidence calibration. It also defines the interpretable attribution calculation of edge and path contributions, and defines the online conditional update and linkage rules with VoITask.

[0042] The minimum intervention module is used to configure the intervention candidate set, estimate the marginal benefit of the intervention using CGNP, set the integer programming expression of the minimum intervention path, set the interactive correction of the impact of the intervention combination, and set the edge closed-loop delivery and local rollback execution mechanism.

[0043] In summary, the urban flood disaster monitoring and early warning system and method provided in this application unifies multi-source observations to node / edge representations through a three-layer directional hydraulic spectrum, and constructs node critical signs using indicators such as recovery rate and variance increase of short sliding windows. This can identify nonlinear critical evolution before traditional absolute threshold triggering. Finally, the minimum intervention set is solved using MIP and distributed to edge execution units for posterior verification, thereby achieving a lower false alarm rate, higher response targeting, and quantifiable risk reduction effect under limited resources. Attached Figure Description

[0044] Figure 1 This is an overall flowchart of an urban flood disaster monitoring and early warning method provided in the embodiments of this application.

[0045] Figure 2 This is an overall framework diagram of an urban flood disaster monitoring and early warning system provided in an embodiment of this application. Detailed Implementation

[0046] Please refer to Figure 1 The document illustrates a flowchart of an embodiment of an urban flood disaster monitoring and early warning system and method according to the present disclosure.

[0047] like Figure 1As shown, a method for monitoring and early warning of urban flooding disasters includes the following steps:

[0048] Step S100: Configure the urban flood hydraulic topology using a three-layer directional hydraulic map, configure nodes to carry the key hydraulic states of urban flooding, configure directed edges to express the direction of water flow and flux capacity, set spatial proximity relationships to map multi-source observation data to nodes, and set a unified expression and value range for node state variables.

[0049] Step S200: Obtain the node water depth time series, configure a short sliding window to perform recovery rate estimation, configure node fluctuation and vulnerability indicators, configure node critical symptom index, set measurement confidence constraints on the calculation conditions of the node critical symptom index, and perform node critical state verification. Based on the node critical state, perform linkage control and state marking.

[0050] Step S300: The calculation method of using nodal hydraulic potential energy as the propagation driving force is adopted. The propagation potential of the node is calculated by using the hydraulic potential energy difference and the transmission impedance. The node uncertainty metric is configured and calculated to reflect the confidence of observation and model. The coupling ranking value PUC is configured and the selection rules under resource constraints are configured.

[0051] Step S400: Set up the observation resource set and define the observation resource attributes, configure the total amount of information uncertainty to quantify the observation value base, use marginal information gain to measure the value of a single observation, use benefit-cost ratio to prioritize observation tasks, execute greedy assignment, and update the observation plan online.

[0052] Step S500: Configure the Counterfactual Graph Neural Predictor (CGNP), generate a counterfactual scene set, label node flooding and path attribution, set the prediction output, set the training loss, and supervise node prediction, path attribution and confidence calibration. Define the interpretable attribution calculation of edge and path contribution, and define the online conditional update and linkage rules with VoITask.

[0053] Step S600: Configure the intervention candidate set, use CGNP to estimate the marginal benefit of the intervention, set the minimum intervention path integer programming expression, set the interactive correction of the impact of the intervention combination, and set the edge closed-loop distribution and local rollback execution mechanism.

[0054] In some specific embodiments, step S100 specifically includes:

[0055] Step S100.1: Use a three-layer directional hydraulic map to configure the urban flood hydraulic topology structure, configure nodes to carry the key hydraulic states of urban flooding, and configure directed edges to express the direction of water flow and flux capacity.

[0056] A three-layer directional hydraulic map is used to model the urban flood flow transmission relationship. The three-layer directional hydraulic map includes a surface layer, a drainage network layer, and a return / overflow coupling layer. The surface layer is used to express the spatial confluence relationship of surface runoff, the drainage network layer is used to express the discharge capacity and connectivity of the underground drainage network, and the return / overflow coupling layer is used to express the bidirectional hydraulic connection between the drainage network and the surface and the network, forming a directional topology that can simultaneously reflect the direction of surface and groundwater hydraulic transmission.

[0057] In the three-layer directional hydraulic map, a node structure is configured. The nodes include road nodes, manhole nodes, pump station nodes, shaft nodes, and low-lying plot nodes. Each node is assigned a unique spatial identifier. Four types of state variables are stored in the node: water depth, inflow rate, drainage network saturation, and measurement confidence level, which are used to describe the flood hydraulic state of the node at the current moment.

[0058] The three-layer directional hydraulic spectrum is configured with a directed edge structure to connect adjacent nodes. The directed edge is used to express the possible transmission direction of water flow between nodes. A transmission impedance parameter and a maximum throughput parameter are set in each directed edge. The transmission impedance is used to describe the ease with which water flows through the connection path, and the maximum throughput is used to limit the maximum water volume transmission capacity of the connection path per unit time.

[0059] Based on spatial proximity, rainfall observation data, rain-measuring radar inversion data, ground water level sensing data, drainage network telemetry data, and terrain elevation data are mapped to corresponding nodes in the three-layer directional hydraulic map. When the horizontal distance between the observation data acquisition location and a certain node is less than or equal to 30 meters, the observation data is mapped to the corresponding node.

[0060] When the same observation data satisfies the mapping conditions of multiple nodes, it is preferentially mapped to the node with the closest spatial distance.

[0061] Step S100.2: Set the spatial proximity relationship to map multi-source observation data to nodes, and set the unified expression and value range of node state variables.

[0062] The water depth of the nodes is expressed in meters, the inflow rate in cubic meters per second, the saturation of the drainage network is expressed as a continuous value between 0 and 1, and the measurement confidence level is expressed as a continuous value between 0 and 1, where the larger the value, the higher the reliability of the corresponding state quantity or observation result.

[0063] The measurement confidence level stored in the nodes is used to constrain subsequent flood analysis and early warning decisions. When the measurement confidence level of a node is less than or equal to 0.6, the weight of that node in the flood risk assessment is reduced.

[0064] When the measurement confidence level is less than or equal to 0.3 and the duration exceeds 10 minutes, the node is marked as a low-confidence node to avoid unreliable observations from misleading the overall flood warning results.

[0065] In some specific embodiments, step S200 specifically includes:

[0066] Step S200.1: Obtain the node water depth time series, configure a short sliding window to perform recovery rate estimation, configure node volatility and vulnerability indicators, and configure node critical symptom index.

[0067] Water depth sensors are deployed at the nodes of the drainage pipe network to acquire corresponding node water depth time series data for each drainage pipe network node. The node water depth time series is formed by the water depth sensors continuously collecting data according to a preset sampling period and uploading it to the data processing unit.

[0068] Based on the acquired node water depth time series data, taking a single drainage network node as the object, a short sliding time window is configured to segment the node water depth time series. The window length of the short sliding time window is set to 5 to 15 minutes, with a default setting of 10 minutes.

[0069] When the sampling period of the water depth sensor is less than or equal to 10 seconds, the node water depth time series is fully incorporated into the short sliding time window for processing according to the original sampling time sequence.

[0070] Based on the node water depth time series within the short sliding time window, a first-order autoregressive approximation model is used to estimate the dynamic recovery characteristics of the node water depth to obtain the corresponding node water depth recovery rate parameter. The first-order autoregressive approximation model is used to describe the autocorrelation evolution relationship between adjacent node water depths, and the node water depth recovery rate parameter is determined using an in-window least squares estimation method or an exponentially weighted least squares estimation method.

[0071] Based on the node water depth time series data within the same short sliding time window, the corresponding node water depth variance value is calculated, and the rate of change of node water depth variance between adjacent short sliding time windows is further calculated to characterize the changing trend of node water depth fluctuation intensity.

[0072] Meanwhile, flow monitoring devices or pipeline status monitoring devices deployed in the drainage network are used to obtain the network saturation data of the corresponding drainage network nodes. The drainage network saturation is used to characterize the proportion of the pipeline cross-section occupied by water. Based on the drainage network saturation data within adjacent short sliding time windows, the growth rate of drainage network saturation is calculated, and its change intensity is expressed in minutes.

[0073] The threshold for the rate of change of node water depth variance is set at 1.2, which means that a 20% increase in node water depth variance compared to the previous time window is considered a significant increase. The threshold for the rate of increase of drainage network saturation is set at 0.05 per minute, which is used to characterize the rapid aggravation of the saturation state of the drainage network.

[0074] Based on the change in the node water depth recovery rate parameter between adjacent short sliding time windows, the node water depth variance change rate, and the drainage network saturation growth rate, a node critical symptom index LCSD is constructed to quantify the degree to which the operating state of drainage network nodes approaches a critical state.

[0075] The Node Critical Symptom Index (LCSD) is constructed using a weighted linear combination method. The change in node water depth recovery rate is used to characterize the degree of weakening of the system's recovery capability, the change rate of node water depth variance is used to characterize the degree of enhancement of system volatility, and the growth rate of drainage network saturation is used to characterize the accumulation rate of structural bearing pressure. Each weight parameter is a non-negative weight and satisfies normalization constraints.

[0076] The threshold for determining the critical symptom index (LCSD) of a node is set to 0.04. When the LCSD reaches or exceeds this threshold, the corresponding drainage network node is determined to have entered a critical state.

[0077] Step S200.2: Set the calculation conditions for the critical symptom index of the measurement confidence constraint node, and perform node critical state verification. Based on the node critical state, perform linkage control and state marking.

[0078] Based on the operational status of the water depth sensor, the evaluation results of data integrity and anomaly rate, a corresponding measurement confidence parameter is configured for each drainage network node to characterize the reliability of the node's water depth time series data.

[0079] The node water depth recovery rate estimation, node water depth variance change rate calculation, and node critical symptom index (LCSD) determination are performed only when the measurement confidence parameter is greater than or equal to 0.6. When the measurement confidence parameter is between 0.3 and 0.6, the node critical symptom index (LCSD) result is reduced and used for internal evaluation. When the measurement confidence parameter is less than or equal to 0.3, the node critical symptom index of the drainage network node is not used as the basis for automatic early warning, and the node is marked as a low-confidence node to participate in observation optimization configuration.

[0080] Based on the node critical symptom index LCSD reaching the judgment threshold, a multivariate consistency check is further performed based on three types of observations: node water depth growth rate, drainage network saturation growth rate, and node inflow rate. When at least two types of observations exceed their respective set thresholds simultaneously within the same short sliding time window, the confidence level of the corresponding drainage network node's critical state judgment is improved.

[0081] Meanwhile, based on the drainage network topology, a multi-node directed hydraulic graph structure is constructed to identify the set of neighboring nodes that are directly connected to the drainage network nodes. When a preset proportion of nodes in the set of neighboring nodes show an upward trend in the critical state index of nodes consistent with that of the drainage network nodes within the same time window, it is confirmed that the critical state of the drainage network nodes has spatial consistency.

[0082] When a drainage network node is determined to be in a critical state, the operating status of the drainage network node is marked as critical, and the node is included in the priority processing node set for subsequent propagation risk assessment and active observation scheduling.

[0083] Within a preset time range after a node enters a critical state, the active observation and scheduling mechanism for that node is triggered first to increase the acquisition density of node water depth time series and drainage network operation status data.

[0084] When the critical symptom index of a node continues to rise within a subsequent time window and the measurement confidence level remains within the preset reliable range, the drainage network node is further marked as a rapidly deteriorating state, and the corresponding rapid response assessment process is executed.

[0085] In some specific embodiments, step S300 specifically includes:

[0086] Step S300.1: The calculation method of using the node hydraulic potential energy as the propagation driving force is adopted, and the node propagation potential energy is calculated by using the hydraulic potential energy difference and the transmission impedance.

[0087] The hydraulic potential energy of each node is calculated using observations and estimations from a constructed three-layer directional hydraulic map. This node hydraulic potential energy is obtained by combining three factors: node water depth, node topographic elevation, and node drainage network pressure. The node water depth is directly collected by water depth sensors deployed in manholes or on the surface. The node topographic elevation is provided by a lidar digital elevation model or urban GIS data. The node drainage network pressure is preferentially collected directly by pressure sensors deployed in pump stations or pipelines. When these pressure sensors are unavailable, the node drainage network pressure is obtained by back-calculating using engineering approximations or numerical models based on measured flow rates and pipe diameters of adjacent pipelines. The node hydraulic potential energy is used to characterize the driving state of a node relative to its neighboring nodes, providing physical quantity input for propagation potential energy calculation.

[0088] The propagation potential of nodes is calculated based on the nodal hydraulic potential energy and the transmission impedance of directed edges in the MDHG. A node representation is used. Its dissemination potential is specifically as follows:

[0089]

[0090] In the formula: For the first The hydraulic potential energy of a node is obtained by combining water depth, terrain elevation, and pipeline pressure. In order to be with the first Nodes connected by directed edges The directly connected first The hydraulic potential energy of the node, For directed edges The transmission impedance is initially calibrated according to road slope, pavement cover type, or pipe diameter and material, and supports online updates. For the first The connectivity importance coefficient of a node is determined by multiplying the node's out-degree by its importance level. The connectivity amplification factor is set to . .

[0091] The obtained Perform normalization to map to The interval, normalized expression is:

[0092]

[0093] In the formula: The set of all nodes in MDHG For the first The normalized propagation potential value of a node.

[0094] Step S300.2: Configure and calculate node uncertainty metrics to reflect observation and model confidence, configure coupling ranking value PUC, and configure selection rules under resource constraints.

[0095] Configuration number The uncertainty measure of a node is ,Will The uncertainty is constructed as a weighted sum of the following three items: sensor observation noise index, model ensemble prediction variance index, and uncertainty discount of public perception report. The sensor observation noise index is obtained from the self-diagnostic information of the nodal depth sensor and the statistical analysis of historical observation fluctuations. The model ensemble prediction variance index is calculated from the output variance of several prediction models running in parallel. The uncertainty discount of public perception report is calculated by combining the image quality score reported by the mobile terminal, the degree of blurring of the location information, and the confidence level of historical reports. The weighted linear combination is expressed as:

[0096]

[0097] In the formula: For the first Normalized index of sensor observation noise at nodes, value range , For the first The normalization index of the prediction variance of the model set of nodes, with a range of values. , For the first Normalized index of uncertainty in public perception at nodes, value range , The weights are non-negative and satisfy the following conditions: , set as ,Will Mapped to The interval, the larger the value, the higher the number. The higher the uncertainty of the node.

[0098] Based on the Normalized value of node propagation potential With uncertainty The propagation uncertainty coupling ranking value PUC is constructed using a product form and is used to quantify the impact of the propagation uncertainty on the current time step. The comprehensive necessity for nodes to implement priority observation or priority intervention is defined by PUC as follows:

[0099]

[0100] In the formula: For the first The propagation uncertainty of the nodes is coupled and ordered, and the range is mapped to after normalization. .

[0101] based on Size setting priority threshold: when At that time, the first The node is listed as a high-priority candidate node, when At that time, the first The node is listed as a medium-priority candidate node, when At that time, the first The node column is a low-priority candidate node, and the threshold can be adjusted within the range of 0.30–0.50 as a subordinate limit.

[0102] Based on all nodes The values ​​are arranged in descending order to form a priority queue. In situations with limited resources, nodes are selected sequentially by priority until the resource budget is met or the preset cumulative PUC coverage is reached. The default cumulative PUC coverage is set to 60% of the total PUC of the node set. For nodes with equal PUC... The nodes are ranked according to their LCSD index and the spatial distance to key protected targets, such as hospitals or shelters, as the ordering rule.

[0103] when The first When a node is marked as a high-priority candidate node, priority is given to triggering the action against the first node within a time window of 0 to 5 minutes. The value information of nodes drives the observation task; when observation resources are insufficient, observation is performed node-wise. Resource allocation is performed based on the ratio of the marginal utility of available observational resources, and subsequent observations show that the... The LCSD index of a node or When the measurement confidence continues to deteriorate, priority should be given to the first Nodes are incorporated into the minimum intervention path assessment to enable rapid response.

[0104] In some specific embodiments, step S400 specifically includes:

[0105] Step S400.1: Set up the observation resource set, define the observation resource attributes, configure the total amount of information uncertainty to quantify the observation value base, and use marginal information gain to measure the value of a single observation.

[0106] Set up observation resource set This refers to a complete list of observation resources that the implementing entity can allocate; the set of observation resources... It includes a drone platform, a mobile measurement vehicle, a portable flow probe, a portable depth gauge, and a public sensing interface.

[0107] For each observation resource in the set Configure observation attributes: observation cost Measured by relative cost or time hours, and observation delay. The effective coverage radius is measured and observed in seconds or minutes. Measurements in meters and observation validity coefficient Used to represent resources At the node The ability to reduce uncertainty, with a range of values. .

[0108] The public perception interface performs irreversible feature extraction and location obfuscation on the device side before reporting to protect privacy. In this method, public reports are regarded as low-cost but high-noise observation resources.

[0109] The total information uncertainty of the configuration system before active observation is implemented. To quantify the current scale of reducible uncertainty across the entire network, the total amount of information uncertainty is specifically given as follows:

[0110]

[0111] In the formula: The set of all nodes in MDHG To calculate the first Node propagation uncertainty coupling ordering value For the first Node uncertainty measure This represents the total weighted uncertainty of the system before active observations are implemented.

[0112] Using the first Nodes are composed of resources The reduction in system uncertainty after the observation is performed is taken as the marginal information gain of the observation task. The specific definition is as follows:

[0113]

[0114] In the formula: The total weighted uncertainty of the system before observation is performed. In order to complete the first Nodes use resources The observations are used to update the total weighted uncertainty of the system. For the first Nodes are composed of resources Marginal information gain after observation.

[0115] To ensure engineering feasibility, the following observation-guided uncertainty update model is used to estimate the uncertainty at each node after observation, thereby constructing... Specifically:

[0116]

[0117] In the formula: In order to observe the first Nodes and using resources After the update Uncertainty at the node, For the first observation Uncertainty at the node, For resources In the The observation validity coefficient at the node, and its range of values. , In MDHG and the first The set of neighboring nodes that are directly connected to a node through directed edges. For the first Node pair The propagation attenuation coefficient of the influence of neighboring nodes, and its range of values. The path weights are obtained by standardization in MDHG, and QT represents other nodes.

[0118] based on structure:

[0119]

[0120] The symbols in the formula have the same meaning as those in the formulas above, so they will not be explained further here.

[0121] Step S400.2: Prioritize observation tasks by comparing benefits and costs, execute greedy assignment, and update the observation plan online.

[0122] Using the benefit-cost ratio of observation tasks As a sorting criterion, it is defined as follows:

[0123]

[0124] In the formula: For the first Nodes are composed of resources Marginal information gain of the observation. For resources The observation cost, measured in relative cost or man-hours, For the first Nodes are composed of resources The benefit-cost ratio of performing observations is a key indicator; a higher value indicates a greater information gain per unit cost.

[0125] based on For all candidate observation tasks Sort in descending order and in resource budget Within the time window constraint, a greedy assignment method is used to select the task sequence.

[0126] The iterative assignment process is executed, which consists of four steps.

[0127] The first rule is: forming a set of high-priority candidate nodes. ,right Each node With each available resource calculate .

[0128] The second rule is: Select the one with the largest... The observation mission is Add the task to the observation execution queue and update the resource budget. .

[0129] The third point is: Execute observation tasks and preprocess and quality verify the returned observation data, and update the uncertainty of relevant nodes online based on qualified observation results. The Node Critical Symptom Index (LCSD) and the estimated parameters of related edges in MDHG.

[0130] The fourth point is: Recalculation , and Repeat steps 1–3 until the resource budget is exhausted or the current time window ends.

[0131] Set default response and resource parameters for dependency constraints: Observe the response window Minutes, default resource cost reference scale: drones Mobile measuring vehicle Portable probe Public perception interface Observation effectiveness coefficient The default value range is The default cumulative PUC coverage rate is set to The above values ​​may be given as optional limitations in the dependent claims to cover implementation variations.

[0132] In some specific embodiments, step S500 specifically includes:

[0133] Step S500.1: Configure the counterfactual graph neural predictor, generate a counterfactual scene set, label node flooding and path attribution, set the prediction output, set the training loss, and supervise node prediction, path attribution, and confidence calibration.

[0134] A graph neural network predictor named Counterfactual Graph Neural Predictor is configured. The Counterfactual Graph Neural Predictor is CGNP. CGNP is defined as MDHGCGNP, MDHG, and the set of engineering calibration parameters for each directed edge. Counterfactual training of CGNP refers to the introduction of a large number of counterfactual scenarios generated by a physical hydrodynamic numerical simulator and an engineering condition injector as supervision data during the training phase.

[0135] A set of counterfactual scenarios is generated using a physical hydrodynamic numerical simulator and an engineering condition injector. Each scene The parameters include: spatiotemporal distribution of rainfall, pump station operating conditions, location and extent of valve blockage, and initial saturation of the pipeline network. A numerical simulator is used to analyze each scenario. Obtain the actual submerged label set of nodes and path-level true propagation contribution .

[0136] The decoder head of CGNP outputs the set of node flooding probabilities respectively. With edge contribution set ,in Indicates in the scene Next The predicted flooding probability of a node. Indicates in the scene Next The predicted contribution score of a directed edge.

[0137] To enable CGNP to have interpretable propagation prediction capabilities in rare / extreme cases, a training loss consisting of three components is employed: node prediction loss. Path attribution monitoring loss With confidence calibration loss The overall training objective is to minimize:

[0138]

[0139]

[0140] In the formula: A collection of counterfactual scenarios, The set of all nodes in MDHG Let MDHG be the set of all directed edges. This is the binary cross-entropy loss function, which can be replaced with mean squared error when the true label is continuous water depth. For the model in the scene Next to the Node flooding probability prediction For the scene Next The actual submerged labels of the nodes are provided by the numerical simulator. The model predicts the first Contribution score of each edge For the scene The actual propagation contribution below was obtained and normalized by post-processing the simulator path tracing. For the scene Next The confidence target label of the node. and It is a non-negative hyperparameter used to balance the three losses.

[0141] Step S500.2: Define the interpretable attribution calculation of edge and path contribution, and define the online conditional update and linkage rules with VoITask.

[0142] To map internal attention or transmission intensity to interpretable edge contributions, a path-based attention accumulation attribution method is employed, which calculates the cumulative attribution of attention in a single forward propagation. The contribution of an edge is defined as the normalized cumulative weight of the path attention, specifically:

[0143]

[0144] In the formula: This represents a directed path from the source node to a target node. Representing a path Includes edges , For the edge The instantaneous attention or transmission strength in the current forward propagation is output by the edge attention head or learnable edge weight function of CGNP. The denominator is the cumulative normalized amount of attention over all candidate paths, ensuring that the edge contribution can be used for comparison and ranking after normalization. The path set is obtained by MDHG topological finite search, and the upper limit of the path length can be limited to 1 to 3 transitions to control the computational complexity.

[0145] New observations returned by VoITask during online operation Encode and conditionalize the first The internal representation of a node enables fast updates; the simplified conditional expression is as follows:

[0146]

[0147] In the formula: For the first observation The internal representation vector of a node. To observe the encoder, the observation will be... Mapped to a vector This is the conditional gain coefficient, with a default value of 0. .

[0148] Will Inject into the message passing layer and complete the re-forward inference within the target's online latency. The target's online latency should not exceed 60 seconds, and for high-priority targets, it should not exceed 1 minute. The linkage rule is set as follows: if the cumulative contribution percentage of several edges in the set... And the predicted probability of the corresponding target node If so, the upstream nodes involved in the above edges will be added to the high-priority observation object set first, so that VoITask can prioritize the scheduling of observation resources.

[0149] The default value is set to: the size of the counterfactual scenario set. to The maximum length of the path enumeration is 3 for each transition, and the causal attribution weight is... Calibration weights Online conditional gain Online target update delay Seconds, high priority Second.

[0150] In some specific embodiments, step S600 specifically includes:

[0151] Step S600.1: Configure the intervention candidate set, use CGNP to estimate the marginal benefit of the intervention, and set the minimum intervention path integer programming expression.

[0152] Configure intervention candidate set For each candidate intervention Configuration: Execution Cost Execution delay The set of nodes that can be affected Local rollback capability identifier .

[0153] Using CGNP's counterfactual conditionation capacity estimation in the implementation of interventions The reduction in the weighted risk of the system is then defined by a single intervention. The marginal benefit is:

[0154]

[0155] In the formula: The set of all nodes in MDHG For the first The node weights are preferably selected. This is used to couple the necessity of propagation with the risk value of nodes. The output of CGNP at the current moment, without intervention. Node flooding probability In order to intervene As the conditional input, the CGNP prediction is the first Node flooding probability A higher value indicates intervention. The greater the contribution to reducing the weighted risk of the system, It can be directly output by the conditionalization module of CGNP, or obtained by performing Monte Carlo inference on several counterfactual scenarios and calculating the average value.

[0156] Given a resource budget With the maximum number of interventions Construct binary choice variables under constraints. When choosing intervention hour The goal is to achieve the maximum reduction of the system's weighted risk within the budget using the minimum set of interventions, which can be expressed using 0 / 1 integer programming:

[0157]

[0158] In the formula: For intervention Expected marginal benefit For intervention The execution cost is measured in units compatible with Cost(r). For available intervention budget, To set the maximum number of interventions that can be deployed simultaneously, .

[0159] To achieve a greedy approximation that prioritizes both efficiency and cost, a heuristic can be used: Prioritizing efficiency / cost ratios.

[0160]

[0161] according to Descending order of intervention until budget or Exhausted, in the formula: The risk of reducing utility due to intervention in unit cost.

[0162] Step S600.2: Set the interactive correction of the impact of the intervention combination, and set the edge closed-loop distribution and local rollback execution mechanism.

[0163] If multiple interventions have significant interaction effects, the combined marginal benefit is denoted as... Furthermore, the optimization objective is extended to a quadratic or mixed-integer programming problem that includes interactive correction terms to avoid double-counting, specifically:

[0164]

[0165] In the formula: Indicates simultaneous implementation and Additional joint gain corrections can be negative to indicate redundancy.

[0166] The optimized set of interventions is distributed through edge execution units: these units include pump station SCADA control modules, valve remote control units, emergency vehicle command terminals, and on-site emergency teams for each selected intervention. Before execution, a local security check must be performed and recorded. :

[0167] like =If rollback is possible, then the executed command includes rollback conditions and rollback instructions. If the error cannot be reversed, then additional manual confirmation or automatic threshold verification is required.

[0168] Execution delay target: High-priority interventions are issued and initiated within 0–5 minutes, and interventions with lack of on-site executability or authorization trigger alternative solutions.

[0169] Immediately after the intervention is implemented, the observation and operation logs should be incorporated into the post-validation process: within the time window after implementation. Collect new observational data And use CGNP to perform counterfactual posterior prediction to estimate the actual risk reduction. Define the performance ratio:

[0170]

[0171] In the formula: For observation-based The actual system weighted risk reduction obtained from online CGNP inference is divided by a denominator of the expected total risk reduction. Then a threshold is recommended. Or when the posterior flooding probability of a critical target node is still higher than the allowable threshold, for example... and If the value is high, a replanning process is triggered: the current remaining budget, the latest observations, and edge execution feedback are used as new inputs to repeat the MIP solution and distribution process. Observations of the mass channel are used in the posterior calculation according to their measurement confidence level. Reduction, for example, by factor ,default .

[0172] All intervention decisions must be recorded: selected intervention set, optimization target value, and estimated value. Timestamps issued, edge execution unit responses, and posterior observations. and Indicator value.

[0173] Please refer to Figure 2 The document illustrates the flow of one embodiment of an urban flood disaster monitoring and early warning system according to this disclosure.

[0174] An urban flood disaster monitoring and early warning system includes the following modules:

[0175] The three-layer hydraulic module is used to configure the urban flood hydraulic topology using a three-layer directional hydraulic map, configure nodes to carry the key hydraulic states of urban flooding, configure directed edges to express the direction of water flow and flux capacity, set spatial proximity relationships to map multi-source observation data to nodes, and set a unified expression and value range for node state quantities.

[0176] The node symptom module is used to acquire the node water depth time series, configure a short sliding window to perform recovery rate estimation, configure node fluctuation and vulnerability indicators, configure the node critical symptom index, set the measurement confidence constraint for the calculation conditions of the node critical symptom index, perform node critical state verification, and perform linkage control and state marking based on the node critical state.

[0177] The propagation potential coupling module is used to calculate the propagation potential of nodes by using the hydraulic potential energy of nodes as the propagation driving force, and to calculate the propagation potential of nodes by using the hydraulic potential energy difference and transmission impedance. It also configures and calculates the node uncertainty metric to reflect the confidence of observations and models, configures the coupling ranking value PUC, and configures the selection rules under resource constraints.

[0178] The information value scheduling module is used to set up the observation resource set, define the attributes of the observation resources, configure the total amount of information uncertainty to quantify the base of observation value, use marginal information gain to measure the value of a single observation, use benefit-cost ratio to prioritize observation tasks, execute greedy assignment, and update the observation plan online.

[0179] The causal attribution module is used to configure the Counterfactual Graph Neural Predictor (CGNP), generate a counterfactual scenario set, label node flooding and path attribution, set the prediction output, set the training loss, and supervise node prediction, path attribution and confidence calibration. It also defines the interpretable attribution calculation of edge and path contributions, and defines the online conditional update and linkage rules with VoITask.

[0180] The minimum intervention module is used to configure the intervention candidate set, estimate the marginal benefit of the intervention using CGNP, set the integer programming expression of the minimum intervention path, set the interactive correction of the impact of the intervention combination, and set the edge closed-loop delivery and local rollback execution mechanism.

[0181] In practical applications, the urban flood monitoring system first constructs a three-layer directional hydraulic map, representing the surface layer, drainage network layer, and return coupling layer. Roads, manholes, pumping stations, and low-lying areas are configured as nodes. Each node stores water depth, inflow rate, network saturation, and measurement confidence level, with the confidence level ranging from 0 to 1. Directed edges describe the direction of water flow and are configured with maximum flux and impedance parameters. Multi-source rainfall, water level, and telemetry data are mapped to corresponding nodes within a 30-meter spatial radius and updated in a unified time sequence.

[0182] Next, the critical signs of nodes are calculated based on short-term water depth changes, with a time window of 10 minutes. The trend of water depth recovery rate and the rate of change of variance are calculated. When the variance growth rate is not less than 1.2 and the pipeline saturation growth rate reaches 0.05 per minute, the node stability decreases significantly. The comprehensive index forms the critical sign value of the node, with a threshold of 0.04. When the threshold is reached, it is marked as a critical node, which is only effective for nodes with a confidence level of not less than 0.6.

[0183] Subsequently, the node propagation potential is calculated and normalized based on the hydraulic potential energy difference, topological connectivity and edge flux capacity. The node uncertainty consists of observation noise and model variance, which is also normalized. The propagation potential and uncertainty are multiplied to form a coupled ranking value PUC. Nodes with a PUC of not less than 0.50 are classified as high priority and participate in subsequent scheduling as key observation and intervention targets.

[0184] Subsequently, active observation scheduling is executed based on information value criteria. Observation cost and effectiveness coefficients are configured for UAVs, mobile measurement equipment, and public perception interfaces. The reduction in system uncertainty is used as an information gain index. The benefit-cost ratio of observation tasks is calculated and sorted by value. Tasks with the highest benefit-cost ratio are executed first. The observation results are updated in real time with node status and PUC value to complete the online closed loop.

[0185] Then, the propagation prediction adopts a counterfactual prediction model based on graph structure. During the training phase, counterfactual samples with different rainfall intensities, pipeline failures and pumping station operating conditions are introduced to enable the model to learn the causal contribution relationship of the propagation path. The model outputs the node flooding probability and path contribution. When the target node flooding probability is not less than 0.50 and the upstream contribution exceeds 0.60, the intervention plan is triggered.

[0186] Finally, the intervention plan generates an execution plan based on the principle of minimum intervention, configuring execution costs and response delays for pump station activation, valve adjustment, and local control. Under budget constraints, the intervention combination with the greatest risk reduction benefits is selected. After the intervention is issued to the edge execution unit, the effect is verified. When the actual risk reduction is less than 80% of the expected value, the system re-plans and iteratively executes.

Claims

1. A method for monitoring and early warning of urban flooding disasters, characterized in that, Includes the following steps: S100: A three-layer directional hydraulic map is used to configure the urban flood hydraulic topology. Nodes are configured to carry the key hydraulic states of urban flooding. Directed edges are configured to express the direction of water flow and flux capacity. Spatial proximity relationships are set to map multi-source observation data to nodes. The unified expression and value range of node state quantities are set. S200: Obtain the node water depth time series, configure a short sliding window to perform recovery rate estimation, configure node fluctuation and vulnerability indicators, configure node critical symptom index, set measurement confidence constraints on the calculation conditions of the node critical symptom index, and perform node critical state verification. Based on the node critical state, perform linkage control and state marking. S300, adopts the calculation method of nodal hydraulic potential energy as the propagation driving force, uses hydraulic potential energy difference and transmission impedance to calculate nodal propagation potential, configures and calculates nodal uncertainty measure to reflect the confidence of observation and model, configures coupling ranking value PUC, and configures selection rules under resource constraints. S400: Set up the set of observation resources and define the attributes of the observation resources. Configure the total amount of information uncertainty to quantify the base value of the observation. Use marginal information gain to measure the value of a single observation. Use the benefit-cost ratio to prioritize the observation tasks. Execute greedy assignment and update the observation plan online. S500, configure the Counterfactual Graph Neural Predictor (CGNP), generate a counterfactual scene set, and label node flooding and path attribution, set the prediction output, set the training loss, and supervise node prediction, path attribution and confidence calibration, define the interpretable attribution calculation of edge and path contribution, and define the online conditional update and linkage rules with VoITask. S600: Configure the intervention candidate set, use CGNP to estimate the marginal benefit of the intervention, set the minimum intervention path integer programming expression, set the interactive correction of the impact of the intervention combination, and set the edge closed-loop distribution and local rollback execution mechanism.

2. The method for monitoring and early warning of urban flooding disasters according to claim 1, characterized in that, S100 specifically includes: S100.

1. A three-layer directional hydraulic map is used to configure the urban flood hydraulic topology structure, with nodes configured to carry the key hydraulic states of urban flooding, and directed edges configured to express the direction of water flow and flux capacity. A three-layer directional hydraulic map is used to model the urban flood flow transmission relationship. The three-layer directional hydraulic map includes a surface layer, a drainage network layer, and a return / overflow coupling layer. The surface layer is used to express the spatial confluence relationship of surface runoff, the drainage network layer is used to express the discharge capacity and connectivity of the underground drainage network, and the return / overflow coupling layer is used to express the bidirectional hydraulic connection between the drainage network and the surface and the network, forming a directional topology that can simultaneously reflect the direction of surface and groundwater hydraulic transmission. In the three-layer directional hydraulic map, a node structure is configured. The nodes include road nodes, manhole nodes, pump station nodes, shaft nodes, and low-lying plot nodes. A unique spatial identifier is set for each node. Four types of state variables are stored in the node: water depth, inflow rate, drainage network saturation, and measurement confidence level, which are used to describe the flood hydraulic state of the node at the current moment. The three-layer directional hydraulic spectrum is configured with a directed edge structure to connect adjacent nodes. The directed edge is used to express the possible transmission direction of water flow between nodes. In each directed edge, a transmission impedance parameter and a maximum throughput parameter are set. The transmission impedance is used to describe the ease with which water flows through the connection path, and the maximum throughput is used to limit the maximum water flow transmission capacity of the connection path per unit time. Based on spatial proximity, rainfall observation data, rain-measuring radar inversion data, ground water level sensing data, drainage network telemetry data, and terrain elevation data are mapped to corresponding nodes in the three-layer directional hydraulic map. When the horizontal distance between the observation data acquisition location and a certain node is less than or equal to 30 meters, the observation data is mapped to the corresponding node. When the same observation data satisfies multiple node mapping conditions, it is preferentially mapped to the node with the closest spatial distance. S100.2, Set up spatial proximity relationships to map multi-source observation data to nodes, and set a unified expression and value range for node state variables; The water depth of the nodes is expressed in meters, the inflow rate in cubic meters per second, the saturation of the drainage network is expressed as a continuous value between 0 and 1, and the measurement confidence level is expressed as a continuous value between 0 and 1, where the larger the value, the higher the confidence level of the corresponding state quantity or observation result. The measurement confidence level stored in the nodes is used to constrain subsequent flood analysis and early warning decisions. When the measurement confidence level of a node is less than or equal to 0.6, the weight of that node in the flood risk assessment is reduced. When the measurement confidence level is less than or equal to 0.3 and the duration exceeds 10 minutes, the node is marked as a low-confidence node to avoid unreliable observations from misleading the overall flood warning results.

3. The method for monitoring and early warning of urban flooding disasters according to claim 1, characterized in that, S200 specifically includes: S200.

1. Obtain the node water depth time series, configure a short sliding window to perform recovery rate estimation, configure node volatility and vulnerability indicators, and configure node critical symptom index. Water depth sensors are deployed at the nodes of the drainage network to acquire corresponding node water depth time series data for each drainage network node. The node water depth time series is formed by the water depth sensors continuously collecting and uploading the data to the data processing unit according to a preset sampling period. Based on the node water depth time series within the short sliding time window, a first-order autoregressive approximation model is used to estimate the dynamic recovery characteristics of the node water depth to obtain the corresponding node water depth recovery rate parameter. The first-order autoregressive approximation model is used to describe the autocorrelation evolution relationship between adjacent node water depths, and the node water depth recovery rate parameter is determined by the in-window least squares estimation method or the exponentially weighted least squares estimation method. Based on the node water depth time series data within the same short sliding time window, the corresponding node water depth variance value is calculated, and the node water depth variance change rate between adjacent short sliding time windows is further calculated to characterize the changing trend of node water depth fluctuation intensity. Meanwhile, flow monitoring devices or pipeline status monitoring devices deployed in the drainage network are used to obtain the network saturation data of the corresponding drainage network nodes. The drainage network saturation is used to characterize the proportion of the pipeline cross-section occupied by water. Based on the drainage network saturation data in adjacent short sliding time windows, the growth rate of drainage network saturation is calculated, and its change intensity is expressed in minutes. Based on the change of the node water depth recovery rate parameter between adjacent short sliding time windows, the node water depth variance change rate, and the drainage network saturation growth rate, a node critical symptom index LCSD is constructed to quantify the degree to which the operating state of drainage network nodes approaches a critical state. The node critical symptom index LCSD is constructed using a weighted linear combination method. The change in node water depth recovery rate is used to characterize the degree of weakening of the system's recovery capacity, the change rate of node water depth variance is used to characterize the degree of enhancement of system volatility, and the growth rate of drainage network saturation is used to characterize the accumulation rate of structural bearing pressure. Each weight parameter is a non-negative weight and satisfies normalization constraints. The threshold for determining the critical symptom index (LCSD) of a node is set to 0.

04. When the LCSD reaches or exceeds this threshold, the corresponding drainage network node is determined to have entered a critical state. S200.2 Set the calculation conditions for the critical symptom index of the measurement confidence constraint node, and perform node critical state verification. Based on the node critical state, perform linkage control and state marking. Based on the evaluation results of the water depth sensor's operating status, data integrity, and anomaly rate, a corresponding measurement confidence parameter is configured for each drainage network node to characterize the reliability of the node's water depth time series data. Only when the measurement confidence parameter is greater than or equal to 0.6, the node water depth recovery rate estimation, node water depth variance change rate calculation, and node critical symptom index LCSD determination are performed. When the measurement confidence parameter is between 0.3 and 0.6, the node critical symptom index LCSD result is reduced and used for internal evaluation. When the measurement confidence parameter is less than or equal to 0.3, the node critical symptom index of the drainage network node is not used as the basis for automatic early warning, and the node is marked as a low-confidence node to participate in observation optimization configuration. Based on the node critical symptom index LCSD reaching the judgment threshold, a multivariate consistency check is further performed based on three types of observations: node water depth growth rate, drainage network saturation growth rate, and node inflow rate. When at least two types of observations exceed their respective set thresholds simultaneously within the same short sliding time window, the confidence level of the corresponding drainage network node's critical state judgment is improved. Meanwhile, based on the multi-node directed hydraulic graph structure constructed from the drainage network topology, the set of neighboring nodes directly connected to the drainage network node is identified. When a preset proportion of nodes in the set of neighboring nodes show an upward trend in the critical state index of the same node as the drainage network node within the same time window, it is confirmed that the critical state of the drainage network node has spatial consistency. When a drainage network node is determined to be in a critical state, the operating status of the drainage network node is marked as critical, and the node is included in the priority processing node set for subsequent propagation risk assessment and active observation scheduling.

4. The method for monitoring and early warning of urban flooding disasters according to claim 1, characterized in that, The S300 specifically includes: S300.1 Calculation method using node hydraulic potential energy as the propagation driving force, and calculation of node propagation potential energy using hydraulic potential energy difference and transmission impedance; The hydraulic potential energy of each node is calculated using the observations and estimations of each node in the constructed three-layer directional hydraulic map. The node hydraulic potential energy is obtained by combining three factors: node water depth, node topographic elevation, and node drainage network pressure. The node water depth is directly collected by water depth sensors deployed in manholes or on the ground surface. The node topographic elevation is provided by lidar digital elevation model or urban GIS data. The node drainage network pressure is preferentially collected by pressure sensors deployed in pump stations or pipelines. When the pressure sensors are unavailable, the node drainage network pressure is obtained by back-calculation using engineering approximation or numerical model information based on the measured flow rate and pipe diameter information of adjacent pipelines. The node hydraulic potential energy is used to characterize the driving state of the node relative to adjacent nodes and provides physical quantity input for propagation potential calculation. The propagation potential of nodes is calculated based on the nodal hydraulic potential energy and the transmission impedance of directed edges in the MDHG. A node representation is used. The potential for dissemination; The obtained Perform normalization to map to interval; S300.2 Configure and calculate node uncertainty metrics to reflect observation and model confidence, configure coupling ranking value PUC, and configure selection rules under resource constraints; Configuration number The uncertainty measure of a node is ,Will The model is constructed as a weighted sum of the following three items: sensor observation noise index, model ensemble prediction variance index, and public perception report uncertainty discount. The sensor observation noise index is obtained from the self-diagnostic information of the node water depth sensor and the statistical analysis of historical observation fluctuations. The model ensemble prediction variance index is obtained from the output variance of several prediction models running in parallel. The public perception report uncertainty discount is obtained by combining the image quality score reported by the mobile terminal, the degree of blurring of the location information, and the confidence level of historical reports. Based on the Normalized value of node propagation potential With uncertainty The propagation uncertainty coupling ranking value PUC is constructed using a product form and is used to quantify the impact of the propagation uncertainty on the current time step. The comprehensive necessity of prioritizing observation or intervention at nodes; Based on all nodes The values ​​are arranged in descending order to form a priority queue. In situations with limited resources, nodes are selected sequentially by priority until the resource budget is met or the preset cumulative PUC coverage is reached. The default cumulative PUC coverage is set to 60% of the total PUC of the node set. For nodes with equal PUC... The nodes are ranked according to their LCSD index and the spatial distance to key protected targets, such as hospitals or shelters, as the ordering rule.

5. The method for monitoring and early warning of urban flooding disasters according to claim 1, characterized in that, The S400 specifically includes: S400.1 Set up a set of observation resources, define the attributes of the observation resources, configure the total amount of information uncertainty to quantify the base of observation value, and use marginal information gain to measure the value of a single observation; Set up observation resource set This refers to a complete list of observation resources that the implementing entity can allocate; the set of observation resources... It includes a drone platform, a mobile measurement vehicle, a portable flow probe, a portable depth gauge, and a public sensing interface; For each observation resource in the set Configure observation attributes: observation cost Measured by relative cost or time hours, and observation delay. The effective coverage radius is measured and observed in seconds or minutes. Measurements in meters and observation validity coefficient Used to represent resources At the node The ability to reduce uncertainty, with a range of values. ; The public perception interface performs irreversible feature extraction and location obfuscation on the device side before reporting to protect privacy. In this method, public reports are regarded as low-cost but high-noise observation resources. The total information uncertainty of the configuration system before active observation is implemented. To quantify the current scale of reducible uncertainty across the entire network; Using the first Nodes are composed of resources The reduction in system uncertainty after the observation is performed is taken as the marginal information gain of the observation task. ; To ensure engineering feasibility, the following observation-guided uncertainty update model is used to estimate the uncertainty at each node after observation, thereby constructing... ; S400.2 Prioritize observation tasks by comparing benefits and costs, execute greedy assignment, and update the observation plan online; Using the benefit-cost ratio of observation tasks As a sorting criterion; based on For all candidate observation tasks Sort in descending order and in resource budget Within the time window constraint, a greedy assignment method is used to select the task sequence; Execute an iterative assignment process, which includes four steps; The first rule is: forming a set of high-priority candidate nodes. ,right Each node With each available resource calculate ; The second rule is: Select the one with the largest... The observation mission is Add the task to the observation execution queue and update the resource budget. ; The third point is: Execute observation tasks and preprocess and quality verify the returned observation data, and update the uncertainty of relevant nodes online based on qualified observation results. , Node Critical Symptom Index (LCSD) and Estimated Parameters of Related Edges in MDHG; The fourth point is: Recalculation , and Repeat steps 1–3 until the resource budget is exhausted or the current time window ends.

6. The method for monitoring and early warning of urban flooding disasters according to claim 1, characterized in that, The S500 specifically includes: S500.1 Configure the counterfactual graph neural predictor, generate a counterfactual scene set, label node flooding and path attribution, set the prediction output, set the training loss, and supervise node prediction, path attribution and confidence calibration. A graph neural network predictor named Counterfactual Graph Neural Predictor is configured. The Counterfactual Graph Neural Predictor is CGNP. CGNP is MDHGC. CGNP is MDHG and the set of engineering calibration parameters for each directed edge. Counterfactual training of CGNP refers to the introduction of a large number of counterfactual scenarios generated by the physical hydrodynamic numerical simulator and the engineering condition injector as supervision data during the training phase. A set of counterfactual scenarios is generated using a physical hydrodynamic numerical simulator and an engineering condition injector. Each scene The parameters include: spatiotemporal distribution of rainfall, pump station operating conditions, location and extent of valve blockage, and initial saturation of the pipeline network. A numerical simulator is used to analyze each scenario. Obtain the actual submerged label set of nodes and path-level true propagation contribution ; The decoder head of CGNP outputs the set of node flooding probabilities respectively. With edge contribution set ,in Indicates in the scene Next The predicted flooding probability of a node. Indicates in the scene Next The predicted contribution score of a directed edge; To enable CGNP to have interpretable propagation prediction capabilities in rare / extreme cases, a training loss consisting of three components is employed: node prediction loss. Path attribution monitoring loss With confidence calibration loss ; S500.2 Define the interpretable attribution calculation of edge and path contribution, and define the online conditional update and linkage rules with VoITask; To map internal attention or transmission intensity to interpretable edge contributions, a path-based attention accumulation attribution method is employed, which calculates the cumulative attribution of attention in a single forward propagation. The contribution of an edge is defined as the normalized cumulative weight of path attention; New observations returned by VoITask during online operation Encode and conditionalize the first The internal representation of nodes enables fast updates; Will Inject into the message passing layer and complete the re-forward inference within the target's online latency. The target's online latency should not exceed 60 seconds, and for high-priority targets, it should not exceed 1 minute. The linkage rule is set as follows: if the cumulative contribution percentage of several edges in the set... And the predicted probability of the corresponding target node If so, the upstream nodes involved in the above edges will be added to the high-priority observation object set first, so that VoITask can prioritize the scheduling of observation resources.

7. The method for monitoring and early warning of urban flooding disasters according to claim 1, characterized in that, The S600 specifically includes: S600.1 Configure the intervention candidate set, use CGNP to estimate the marginal benefit of the intervention, and set the minimum intervention path integer programming expression; Configure intervention candidate set For each candidate intervention Configuration: Execution Cost Execution delay The set of nodes that can be affected Local rollback capability identifier ; Using CGNP's counterfactual conditionation capacity estimation in the implementation of interventions The amount by which the weighted risk of the system is reduced; Given a resource budget With the maximum number of interventions Construct binary choice variables under constraints. When choosing intervention hour The goal is to achieve the maximum reduction of system-weighted risk within the budget with the minimum set of interventions. S600.2, Set the interactive correction of the impact of intervention combination, and set the edge closed-loop distribution and local rollback execution mechanism; If multiple interventions have significant interaction effects, the combined marginal benefit is denoted as... Furthermore, the optimization objective is extended to a quadratic or mixed integer programming problem that includes interactive correction terms to avoid double-counting. The optimized set of interventions is distributed through edge execution units: these units include pump station SCADA control modules, valve remote control units, emergency vehicle command terminals, and on-site emergency teams for each selected intervention. Before execution, a local security check must be performed and recorded. : like =If rollback is possible, then the executed command includes rollback conditions and rollback instructions. =If it cannot be rolled back, then additional manual confirmation or automatic threshold verification is required; Execution delay target: High-priority interventions are issued and initiated within 0–5 minutes, and interventions with limited feasibility or lack of authorization on-site trigger alternative solutions; Immediately after the intervention is implemented, the observation and operation logs should be incorporated into the post-validation process: within the time window after implementation. Collect new observational data And use CGNP to perform counterfactual posterior prediction to estimate the actual risk reduction. ; All intervention decisions must be recorded: selected intervention set, optimization target value, and estimated value. Timestamps issued, edge execution unit responses, and posterior observations. and Indicator value.

8. An urban flood disaster monitoring and early warning system, implementing the urban flood disaster monitoring and early warning method as described in any one of claims 17, characterized in that, Includes the following modules: The three-layer hydraulic module is used to configure the urban flood hydraulic topology using a three-layer directional hydraulic map, configure nodes to carry the key hydraulic states of urban flooding, configure directed edges to express the direction of water flow and flux capacity, set spatial proximity relationships to map multi-source observation data to nodes, and set a unified expression and value range of node state quantities. The node symptom module is used to acquire the node water depth time series, configure a short sliding window to perform recovery rate estimation, configure node fluctuation and vulnerability indicators, configure the node critical symptom index, set the measurement confidence constraint for the calculation conditions of the node critical symptom index, and perform node critical state verification. Based on the node critical state, it performs linkage control and state marking. The propagation potential coupling module is used to calculate the propagation potential of nodes by using nodal hydraulic potential energy as the propagation driving force, and to calculate the node propagation potential by using hydraulic potential energy difference and transmission impedance. It also configures and calculates node uncertainty metrics to reflect the confidence of observations and models, configures the coupling ranking value PUC, and configures selection rules under resource constraints. The information value scheduling module is used to set up the observation resource set, define the attributes of the observation resources, configure the total amount of information uncertainty to quantify the observation value base, use marginal information gain to measure the value of a single observation, use benefit-cost ratio to prioritize observation tasks, execute greedy assignment, and update the observation plan online. The causal attribution module is used to configure the Counterfactual Graph Neural Predictor (CGNP), generate a counterfactual scenario set, label node flooding and path attribution, set the prediction output, set the training loss, and supervise node prediction, path attribution and confidence calibration. It also defines the interpretable attribution calculation of edge and path contribution, and defines the online conditional update and linkage rules with VoITask. The minimum intervention module is used to configure the intervention candidate set, estimate the marginal benefit of the intervention using CGNP, set the integer programming expression of the minimum intervention path, set the interactive correction of the impact of the intervention combination, and set the edge closed-loop delivery and local rollback execution mechanism.

Images