Adaptive Neural Network Flood Evolution Simulation and Risk Assessment System and Method

Abstract

This invention discloses an adaptive neural network method for flood evolution simulation and risk assessment, comprising the following steps: Step 1, generating multi-source hydrological and meteorological preprocessing data; Step 2, constructing a river network topology map and generating node feature vectors and edge feature vectors to form a spatiotemporal graph input data; Step 3, inputting a liquid time constant neural network to perform continuous time state updates, generating a set of node hydrological state prediction vectors; Step 4, injecting hydrodynamic physical consistency residuals to generate a set of physically constrained hydrological state prediction vectors; Step 5, performing adaptive lag compensation to generate a set of lag-compensated hydrological state prediction vectors; Step 6, mapping and generating predicted data for flow, water level, and peak arrival time; Step 7, outputting a risk classification map and a list of early warning thresholds. This invention achieves accurate flood process prediction and dynamic generation of risk maps, improving early warning accuracy and response efficiency.

Description

Technical Field

[0001] This invention relates to the field of flood risk assessment technology, and in particular to an adaptive neural network flood evolution simulation and risk assessment system and method. Background Technology

[0002] Existing flood evolution prediction and risk assessment methods mostly rely on traditional hydrological and hydrodynamic models, often driven by a single data source, such as using only ground rain gauge and water level data for simulation. In watersheds with complex river network structures and large differences in the spatiotemporal distribution of rainfall, these methods are difficult to accurately reflect the impact of upstream water flow on downstream flood peaks. The parameters of traditional models are mostly fixed values, which are not adaptable to changes in watershed conditions. The time lag effect of rainfall-runoff relationship and the response capability of rapid changes in flood peak are weak, which can easily lead to large deviations in flood peak flow prediction and delays in peak arrival time.

[0003] In terms of spatial representation, existing methods have limited ability to structurally model watershed observation sections and river segments. The construction of node and edge features is incomplete, making it difficult to effectively map multi-source hydrological and meteorological data to spatiotemporal feature expressions that reflect hydrological processes. Rule-based risk assessment methods rely on fixed thresholds and lack dynamic coupling with real-time prediction results, resulting in a lack of specificity and timeliness in flood risk classification. The application of existing deep learning models in flood evolution prediction is mostly focused on sequence prediction, lacking integration with hydrodynamic constraints. The prediction results have deviations in terms of physical consistency, and the models are insufficient in adapting to sudden rainfall changes and handling the stability of receding water sections.

[0004] In dealing with lag effects, conventional methods mostly rely on static corrections and cannot be adaptively adjusted in combination with the real-time status of the watershed. This makes it difficult to control prediction errors during the formation and propagation stages of flood peaks. At the risk output level, most methods fail to spatially overlay flood evolution prediction results with the exposure of watershed infrastructure, lack a refined assessment of the impact range of key facilities, and lack the ability to dynamically generate warning threshold lists. These problems limit the accuracy of flood prediction, response speed, and the scientific nature of risk assessment when dealing with complex hydrological scenarios.

[0005] Therefore, how to provide an adaptive neural network flood evolution simulation and risk assessment system and method is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose an adaptive neural network flood evolution simulation and risk assessment system and method. This invention integrates multi-source hydrological and meteorological data, river network topology modeling and hydrodynamic physical consistency constraints, and combines adaptive lag compensation and dynamic risk classification to achieve accurate flood process prediction and risk map generation. It has the advantages of high prediction accuracy, fast response speed and strong interpretability of risk assessment results.

[0007] The adaptive neural network flood evolution simulation and risk assessment method according to embodiments of the present invention includes the following steps:

[0008] Step 1: Collect watershed rainfall observation data and river water level observation data, and preprocess them to obtain multi-source hydrological and meteorological preprocessed data;

[0009] Step 2: Construct a river network topology map based on multi-source hydrological and meteorological preprocessing data, construct the observation cross-section as node feature vectors, construct the river segment as edge feature vectors, and form a spatiotemporal map input data;

[0010] Step 3: Input the spatiotemporal graph input data into the liquid time constant neural network with embedded graph attention mechanism, perform continuous time state updates, and generate a set of node hydrological state prediction vectors;

[0011] Step 4: Inject hydrodynamic physical consistency residuals into the nodal hydrological state prediction vector set to generate a physically constrained hydrological state prediction vector set.

[0012] Step 5: Perform adaptive lag compensation on the set of physically constrained hydrological state prediction vectors to generate a set of lag-compensated hydrological state prediction vectors.

[0013] Step 6: Map the set of delayed compensation hydrological state prediction vectors to flow prediction data, water level prediction data and flood peak arrival time prediction data to form flood evolution prediction results data;

[0014] Step 7: Based on the flood evolution prediction results, calculate the inundation depth, flow velocity, and arrival time data under different annual exceedance probabilities. Combine this with the spatial data of watershed infrastructure to generate flood risk level data and key facility exposure data, and output flood risk classification map data and flood warning threshold list data.

[0015] Optionally, step one includes: performing time series missing value filling, outlier removal, and time granularity unification processing on the watershed rainfall observation data to generate rainfall preprocessed data; performing time series missing value filling, outlier removal, and station benchmark unification processing on the river water level observation data to generate water level preprocessed data; and synchronizing and registering the rainfall preprocessed data and water level preprocessed data according to a unified time benchmark to construct multi-source hydrological and meteorological preprocessed data.

[0016] Optionally, step two includes: extracting water level preprocessing data and time indexes corresponding to each observation section from multi-source hydrological and meteorological preprocessing data; sequentially combining water level preprocessing data of the same observation section under continuous time indexes to construct node feature vectors; extracting rainfall preprocessing data, roughness data, and gradient data corresponding to each river segment from multi-source hydrological and meteorological preprocessing data; sequentially combining data of the same river segment under continuous time indexes to construct edge feature vectors; combining node feature vectors and edge feature vectors according to the upstream and downstream relationships connecting river segments to generate a river network topology map; arranging all node feature vectors and edge feature vectors in the river network topology map according to time indexes to form spatiotemporal map input data.

[0017] Optionally, the processing of the liquid time constant neural network with the embedded graph attention mechanism in step three includes:

[0018] The spatiotemporal graph input data is sorted in ascending order by time index. The node feature vector and edge feature vector under each time index are extracted. The learnable time constant parameter is initialized for each node. Normalization is performed according to the range of node feature vector to obtain the initial value of the time constant. The graph attention parameter is initialized for each edge and bound to the edge feature vector corresponding to the edge. The node feature vector under the first time index is used as the initial node hydrological state prediction vector.

[0019] Under the current time index, the current node feature vector is connected with the hydrological state prediction vector of the node in the previous time index to form a query vector. For each incoming edge connected to a node, the hydrological state prediction vector of the previous time index node of the adjacent node is connected with the edge feature vector of the incoming edge to form a key vector and the corresponding value vector.

[0020] Perform a dot product operation on the query vector and key vector according to their corresponding dimensions to obtain a relevance score. Multiply the relevance score element-wise with the graph attention parameters of the incoming edges, and add the result of the linear transformation of the edge feature vector to obtain an unnormalized attention score. Apply the graph attention parameters of the corresponding edges to the attention score. Normalize the scores of all incoming edges of the same node to obtain a set of attention weights. Perform a weighted summation on the corresponding value vectors using the attention weights to generate a graph attention aggregation vector.

[0021] The graph attention aggregation vector is concatenated with the current node feature vector and input into the liquid time constant unit. The update coefficient is calculated based on the time constant parameter and the time index interval. The hydrological state prediction vector of the previous time index node and the state vector of the candidate node are weighted and updated to obtain the hydrological state prediction vector of the current node. The calculation is repeated on all time indices, and the node hydrological state prediction vectors of each time index are collected to form a set of node hydrological state prediction vectors.

[0022] Optionally, step four includes:

[0023] Under each time index, the hydrological state prediction vector of the current node and the hydrological state prediction vector of the adjacent node are read from the set of node hydrological state prediction vectors. Combined with the corresponding edge feature vectors, the inflow and outflow of the current node are summed respectively within the same time index. The difference between the sum of inflow and outflow is calculated to obtain the residual of the continuity equation.

[0024] Under each time index, the time index interval is allocated based on the difference between the current node's hydrological state prediction vector and the previous time index node's hydrological state prediction vector. The residual of the momentum equation is calculated by combining the water level gradient along the river channel direction and the friction slope determined by the roughness and flow velocity.

[0025] Under each time index, the residuals of the continuity equation and the residuals of the momentum equation are weighted and synthesized into a hydrodynamic physical consistency residual vector according to the adaptive weights. The adaptive weights are determined based on the monotonic mapping of the residual magnitude, the time index interval, and the time constant parameter.

[0026] Under each time index, the hydrological state prediction vector of the current node is linearly corrected based on the hydrodynamic physical consistency residual vector and the residual injection gain coefficient to generate a physically constrained hydrological state prediction vector. The residual injection gain coefficient is determined by the time constant parameter and the time index interval and is limited to a preset closed interval.

[0027] All time indices are processed repeatedly, and the physical constraint hydrological state prediction vectors are collected according to the time index to form a set of physical constraint hydrological state prediction vectors.

[0028] Optionally, step five includes:

[0029] Under each time index, the physical constraint hydrological state prediction vector of the current node and the physical constraint hydrological state prediction vector of several historical time indices are extracted from the set of physical constraint hydrological state prediction vectors. The cumulative curve of the corresponding rainfall process and the cumulative curve of the flow process are calculated. The time difference between the rainfall peak and the flow peak is determined based on the difference in the characteristic inflection point position of the two cumulative curves, and the time delay estimate of the current node is generated.

[0030] Under each time index, the time delay estimate and corresponding time constant parameter of the current node are extracted, and the time constant parameter of the previous time index is extracted. Normalization is performed on the time delay estimate and time constant parameter to obtain the standardized time delay and standardized time constant. The rate of change of the time constant parameter between adjacent time indices is calculated to obtain the change in time constant. Weight pairs are determined based on the standardized time delay and the change in time constant, so that the sum of the weight pairs is one. Linear fusion is performed on the standardized time delay and standardized time constant according to the weight pairs to generate unrestricted correction coefficients. Monotonic normalization and upper and lower bound pruning are performed on the unrestricted correction coefficients to restrict the results to a preset closed interval to obtain adaptive time delay correction coefficients.

[0031] Under each time index, the alignment position of the current node's physical constraint hydrological state prediction vector on the time axis is adjusted according to the adaptive time delay correction coefficient. Interpolation update is performed on the hydrological state prediction vector of the node that needs to be moved forward, and delayed smoothing update is performed on the hydrological state prediction vector of the node that needs to be moved backward.

[0032] At each time index, the node hydrological state prediction vector after adaptive lag compensation is output. The output vectors of all nodes are collected according to the time index to form a set of lag-compensated hydrological state prediction vectors.

[0033] Optionally, the flood risk classification map data and the flood warning threshold list data specifically include:

[0034] The flood evolution prediction results obtained by mapping the delayed compensation hydrological state prediction vector set are used to extract flow prediction data, water level prediction data and flood peak arrival time prediction data. Combined with the corresponding rainfall and watershed infrastructure spatial data, the inundation depth data, flow velocity data and arrival time data of each grid cell are calculated.

[0035] The inundation depth, flow velocity, and time of arrival data of each grid cell are matched with a preset set of risk classification conditions. The risk classification condition set consists of inundation depth thresholds, flow velocity thresholds, and time of arrival thresholds, and is divided into four levels:

[0036] When the flooding depth is greater than or equal to three meters, the flow velocity is greater than or equal to two meters per second, and the arrival time is less than or equal to one hour, the grid cell is marked as Level 1 risk;

[0037] When the flooding depth is between two and three meters, the flow velocity is between one and two meters per second, or the arrival time is between one and three hours, and the conditions for Level 1 risk are not met, the grid cell is marked as Level 2 risk.

[0038] When the flooding depth is between 0.5 and 2 meters, the flow velocity is between 0.5 and 1 meter per second, or the arrival time is between 3 and 6 hours, and the conditions for level I and II risks are not met, the grid cell is marked as level III risk.

[0039] When the inundation depth is less than 0.5 meters, the flow velocity is less than 0.5 meters per second, and the arrival time is greater than 6 hours, the grid cell is marked as low risk;

[0040] Under each time index, the risk level labels of all raster cells are combined into flood risk level data. The flood risk level data is spatially overlaid with the critical facility exposure data. The types and quantities of critical facilities covered under each risk level are extracted, and a flood warning threshold list is generated and output.

[0041] The adaptive neural network flood evolution simulation and risk assessment system according to an embodiment of the present invention includes the following modules:

[0042] The data acquisition and preprocessing module is used to collect watershed rainfall observation data and river water level observation data and generate multi-source hydrological and meteorological preprocessed data;

[0043] The river network topology construction module is used to transform multi-source hydrological and meteorological preprocessing data into node feature vectors and edge feature vectors and form spatiotemporal graph input data.

[0044] The liquid time constant neural network processing module is used to input spatiotemporal map data and perform continuous time state updates to generate a set of node hydrological state prediction vectors;

[0045] The physical consistency residual injection module is used to inject hydrodynamic physical consistency residuals into the set of node hydrological state prediction vectors to generate a set of physically constrained hydrological state prediction vectors.

[0046] The adaptive lag compensation module is used to perform adaptive lag compensation on the set of hydrological state prediction vectors under physical constraints to generate a set of hydrological state prediction vectors with lag compensation.

[0047] The flood evolution result generation module is used to map the set of delayed compensation hydrological state prediction vectors into flow prediction data, water level prediction data and flood peak arrival time prediction data.

[0048] The risk assessment and early warning generation module is used to generate flood risk classification map data and flood early warning threshold list data.

[0049] The beneficial effects of this invention are:

[0050] (1) By constructing multi-source hydrological and meteorological preprocessing data and river network topology map, the observation section and river segment are transformed into node feature vector and edge feature vector. A liquid time constant neural network with embedded graph attention mechanism is introduced to realize continuous time update prediction of watershed hydrological status, improve the prediction accuracy of flood peak flow and peak occurrence time, and reduce calculation deviation under complex river network conditions.

[0051] (2) Inject hydrodynamic physical consistency residuals into the set of nodal hydrological state prediction vectors, and combine them with an adaptive lag compensation mechanism to dynamically correct the rainfall-runoff time lag, improve the response speed and stability of flood evolution prediction in the stages of sudden rainfall changes and rapid formation of flood peaks, and enhance the smoothness and physical rationality of the recession section prediction.

[0052] (3) The prediction results after lag compensation are converted into flow, water level and flood peak arrival time data. Combined with the spatial data of watershed infrastructure, the inundation depth, flow velocity and arrival time are calculated based on the overshoot probability of different years. Flood risk classification is realized according to multiple condition thresholds, and a list of flood warning thresholds is dynamically generated to improve the pertinence of risk assessment and the enforceability of warning. Attached Figure Description

[0053] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0054] Figure 1 This is an overall flowchart of the adaptive neural network flood evolution simulation and risk assessment method proposed in this invention;

[0055] Figure 2 This is a schematic diagram of the hydrodynamic physical consistency residual injection and adaptive hysteresis compensation process proposed in this invention. Detailed Implementation

[0056] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0057] refer to Figure 1 and Figure 2 An adaptive neural network system and method for flood evolution simulation and risk assessment includes the following steps:

[0058] Step 1: Collect watershed rainfall observation data and river water level observation data, and preprocess them to obtain multi-source hydrological and meteorological preprocessed data;

[0059] Step 2: Construct a river network topology map based on multi-source hydrological and meteorological preprocessing data, construct the observation cross-section as node feature vectors, construct the river segment as edge feature vectors, and form a spatiotemporal map input data;

[0060] Step 3: Input the spatiotemporal graph input data into the liquid time constant neural network with embedded graph attention mechanism, perform continuous time state updates, and generate a set of node hydrological state prediction vectors;

[0061] Step 4: Inject hydrodynamic physical consistency residuals into the nodal hydrological state prediction vector set to generate a physically constrained hydrological state prediction vector set.

[0062] Step 5: Perform adaptive lag compensation on the set of physically constrained hydrological state prediction vectors to generate a set of lag-compensated hydrological state prediction vectors.

[0063] Step 6: Map the set of delayed compensation hydrological state prediction vectors to flow prediction data, water level prediction data and flood peak arrival time prediction data to form flood evolution prediction results data;

[0064] Step 7: Based on the flood evolution prediction results, calculate the inundation depth, flow velocity, and arrival time data under different annual exceedance probabilities. Combine this with the spatial data of watershed infrastructure to generate flood risk level data and key facility exposure data, and output flood risk classification map data and flood warning threshold list data.

[0065] In this embodiment, step one includes: performing time series missing value filling, outlier removal, and time granularity unification processing on the watershed rainfall observation data to generate rainfall preprocessed data; performing time series missing value filling, outlier removal, and station benchmark unification processing on the river water level observation data to generate water level preprocessed data; and synchronizing and registering the rainfall preprocessed data and water level preprocessed data according to a unified time benchmark to construct multi-source hydrological and meteorological preprocessed data.

[0066] In this embodiment, step two includes: extracting water level preprocessing data and time indexes corresponding to each observation section from multi-source hydrological and meteorological preprocessing data; sequentially combining water level preprocessing data of the same observation section under continuous time indexes to construct node feature vectors; extracting rainfall preprocessing data, roughness data, and gradient data corresponding to each river segment from multi-source hydrological and meteorological preprocessing data; sequentially combining data of the same river segment under continuous time indexes to construct edge feature vectors; combining node feature vectors and edge feature vectors according to the upstream and downstream relationships connecting river segments to generate a river network topology map; arranging all node feature vectors and edge feature vectors in the river network topology map according to time indexes to form spatiotemporal map input data.

[0067] In this embodiment, the processing of the liquid time constant neural network with the embedded graph attention mechanism in step three includes:

[0068] The spatiotemporal graph input data is sorted in ascending order by time index. The node feature vector and edge feature vector under each time index are extracted. The learnable time constant parameter is initialized for each node. Normalization is performed according to the range of node feature vector to obtain the initial value of the time constant. The graph attention parameter is initialized for each edge and bound to the edge feature vector corresponding to the edge. The node feature vector under the first time index is used as the initial node hydrological state prediction vector.

[0069] Under the current time index, the current node feature vector is connected with the hydrological state prediction vector of the node in the previous time index to form a query vector. For each incoming edge connected to a node, the hydrological state prediction vector of the previous time index node of the adjacent node is connected with the edge feature vector of the incoming edge to form a key vector and the corresponding value vector.

[0070] Perform a dot product operation on the query vector and key vector according to their corresponding dimensions to obtain a relevance score. Multiply the relevance score element-wise with the graph attention parameters of the incoming edges, and add the result of the linear transformation of the edge feature vector to obtain an unnormalized attention score. Apply the graph attention parameters of the corresponding edges to the attention score. Normalize the scores of all incoming edges of the same node to obtain a set of attention weights. Perform a weighted summation on the corresponding value vectors using the attention weights to generate a graph attention aggregation vector.

[0071] The graph attention aggregation vector is concatenated with the current node's feature vector and input into the liquid time constant unit. Update coefficients are calculated based on the time constant parameter and the time index interval. A weighted update is performed on the hydrological state prediction vector of the previous time index node and the candidate node's state vector to obtain the current node's hydrological state prediction vector. This process is repeated across all time indices, and the node hydrological state prediction vectors for each time index are collected to form a node hydrological state prediction vector set. Specifically, this includes:

[0072] The graph attention aggregation vector is concatenated with the current node's feature vector and input into the liquid time constant unit. For each node, at the time index... The following is the basis:

[0073] ;

[0074] in, This represents the node index, corresponding to the observation section in the river network topology diagram. Indicates a time index. This represents the feature vector of the node at the current time index. This represents the graph attention aggregation vector. This represents a vector concatenation operation. Represents a learnable weight vector. This represents a learnable bias scalar. Represents the smoothing and nonnegativity function. A small positive constant used for numerical stability;

[0075] Calculate the time constant parameter using attention-weighted incoming edge feature vectors:

[0076] ;

[0077] in, Represents a node The set of incoming adjacent nodes, Indicates the index of the upstream adjacent node. Indicates from node Pointing to node The attention weights are non-negative and sum to 1; From node Pointing to node The edge feature vectors;

[0078] The modulation factor is obtained:

[0079] ;

[0080] in, Represents the linear rectification activation function. Represents a learnable weight vector. This represents a learnable bias scalar. This represents the modulation factor calculated from the edge feature aggregation result;

[0081] Forming an effective time constant:

[0082] ;

[0083] in, This represents the effective time constant after modulation;

[0084] Based on adjacent time index interval Calculate the continuous-time update coefficients:

[0085] ;

[0086] in, This indicates the interval between the current time and the previous time index. Represents the original continuous-time update coefficients;

[0087] Through bounded mapping:

[0088] ;

[0089] in, This indicates the lower and upper bounds of the update coefficients. Indicates the reference center value. Represents the temperature coefficient, a positive real number. This represents the Sigmoid function. Represents numerically stable update coefficients;

[0090] Hydrological state prediction vector for the previous time index node With candidate node state vector renew:

[0091]

[0092] in, This represents the hydrological state prediction vector of the node indexed in the previous time period. This represents the candidate node state vector generated by the liquid time constant unit. This represents the hydrological state prediction vector for the node at the current time index.

[0093] Perform a weighted update to obtain the hydrological state prediction vector for the current node. Repeat the calculation on all time indices and collect the node hydrological state prediction vectors for each time index to form a set of node hydrological state prediction vectors.

[0094] In this embodiment, step four includes:

[0095] Under each time index, the hydrological state prediction vector of the current node and the hydrological state prediction vector of the adjacent node are read from the set of node hydrological state prediction vectors. Combined with the corresponding edge feature vectors, the inflow and outflow of the current node are summed respectively within the same time index. The difference between the sum of inflow and outflow is calculated to obtain the residual of the continuity equation.

[0096] Under each time index, the time index interval is allocated based on the difference between the current node's hydrological state prediction vector and the previous time index node's hydrological state prediction vector. The residual of the momentum equation is calculated by combining the water level gradient along the river channel direction and the friction slope determined by the roughness and flow velocity.

[0097] Under each time index, the residuals of the continuity equation and the residuals of the momentum equation are weighted and synthesized into a hydrodynamic physical consistency residual vector according to the adaptive weights. The adaptive weights are determined based on the monotonic mapping of the residual magnitude, the time index interval, and the time constant parameter.

[0098] Under each time index, the hydrological state prediction vector of the current node is linearly corrected based on the hydrodynamic physical consistency residual vector and the residual injection gain coefficient to generate a physically constrained hydrological state prediction vector. The residual injection gain coefficient is determined by the time constant parameter and the time index interval and is limited to a preset closed interval.

[0099] All time indices are processed repeatedly, and the physical constraint hydrological state prediction vectors are collected according to the time index to form a set of physical constraint hydrological state prediction vectors.

[0100] Specifically, it includes:

[0101] At each time index, the hydrological state prediction vector of the current node and the hydrological state prediction vectors of its adjacent nodes are obtained from the set of node hydrological state prediction vectors. The continuity equation residual is then calculated by combining this residual with the corresponding edge feature vectors.

[0102] ;

[0103] in, Represents a node The set of incoming adjacent nodes, Represents a node The set of out-adjacent nodes, Indicates time index The following nodes Flow to Node The flow rate is calculated based on the hydrological state prediction vectors of adjacent nodes and the corresponding edge feature vectors. Indicates time index The following nodes Flow to Node Traffic;

[0104] Calculate the residual of the momentum equation, which is:

[0105] ;

[0106] in, Indicates time index Next node Total flow Time index interval Gravitational acceleration constant Time Index Next node water level height The water level gradient along the river channel is calculated based on the water level difference between adjacent nodes and the length of the river segment. Time Index The friction slope is calculated based on roughness and flow velocity.

[0107] The physical consistency residual vector is obtained by combining the residuals of the continuity equation and the momentum equation with weighted coefficients.

[0108] ;

[0109] in, These are the residual weighting coefficients for the continuity equation. These are the residual weighting coefficients for the momentum equation;

[0110] The physical consistency residual vector is used as a correction factor and applied to the hydrological state prediction vector of the current node to generate a physically constrained hydrological state prediction vector.

[0111] ;

[0112] in, This is the residual correction step size coefficient. This is the corrected nodal hydrological state prediction vector;

[0113] The calculation is repeated across all time indices, and the physical constraint hydrological state prediction vectors for each time index are collected to form a set of physical constraint hydrological state prediction vectors.

[0114] In this embodiment, step five includes:

[0115] Under each time index, the physical constraint hydrological state prediction vector of the current node and the physical constraint hydrological state prediction vector of several historical time indices are extracted from the set of physical constraint hydrological state prediction vectors. The cumulative curve of the corresponding rainfall process and the cumulative curve of the flow process are calculated. The time difference between the rainfall peak and the flow peak is determined based on the difference in the characteristic inflection point position of the two cumulative curves, and the time delay estimate of the current node is generated.

[0116] Under each time index, the estimated time delay and corresponding time constant parameter of the current node are extracted, and the time constant parameter of the previous time index is also extracted. Normalization is performed on the estimated time delay and time constant parameter to obtain the standardized time delay and standardized time constant. The rate of change of the time constant parameter between adjacent time indices is calculated to obtain the change in time constant. Weight pairs are determined based on the standardized time delay and the change in time constant, ensuring the sum of the weight pairs is one. Linear fusion is performed on the standardized time delay and standardized time constant according to the weight pairs to generate unrestricted correction coefficients. Monotonic normalization and upper / lower bound pruning are performed on the unrestricted correction coefficients to restrict the results to a preset closed interval, obtaining an adaptive time constant. The time delay correction coefficient is calculated as follows: Normalization and upper / lower bound pruning of the unrestricted correction coefficient include: for each time index, reading the unrestricted correction coefficient of the current node and a preset normalization target range; subtracting the minimum value of the unrestricted correction coefficients of all nodes within that time index from the unrestricted correction coefficient; dividing the result by the difference between the maximum and minimum values ​​of the unrestricted correction coefficients of all nodes within that time index to obtain the normalization result; comparing the normalization result with a preset lower bound value, replacing it with the lower bound value when it is less than the lower bound value; comparing the normalization result with a preset upper bound value, replacing it with the upper bound value when it is greater than the upper bound value; and outputting the adaptive time delay correction coefficient limited to a preset closed interval.

[0117] Under each time index, the alignment position of the current node's physical constraint hydrological state prediction vector on the time axis is adjusted according to the adaptive time delay correction coefficient. Interpolation update is performed on the hydrological state prediction vector of the node that needs to be moved forward, and delayed smoothing update is performed on the hydrological state prediction vector of the node that needs to be moved backward.

[0118] The specific processing for "performing interpolation updates and performing delayed smooth updates" includes:

[0119] For the hydrological state prediction vectors of nodes that need to be moved forward, interpolation updates are performed: Under each time index, the forward movement step size is calculated based on the adaptive time lag correction coefficient, and the current time index is denoted as t; when the current forward movement step size is less than one time step, the prediction vectors at time indices t and t+1 are selected, and an interpolation vector is generated according to the linear weights corresponding to the forward movement step size, and the interpolation vector is written to t; when the current forward movement step size is not less than one time step, linear interpolation and writing are repeated at unit time steps until all forward movement steps are completed; when there is no t+1 available at the end, the end vector copy is used as the interpolation input;

[0120] For the hydrological state prediction vectors of nodes that need to be shifted, a delayed smoothing update is performed: Under each time index, the shift step size is calculated based on the adaptive time lag correction coefficient, and the current time index is denoted as t; when the shift step size is less than one time step, the prediction vector at t and the prediction vector at t+1 are distributed and written at t and t+1 according to the linear weights corresponding to the shift step size; when the shift step size is not less than one time step, the prediction vector at t is first written to the position t+shift step position, and the remaining part less than one step is distributed proportionally between two adjacent target indices; a fixed window weighted average smoothing is performed on the written position and adjacent positions, and the weights decrease monotonically with the distance from the target index; when the boundary is crossed at the beginning or end, the beginning or end vector is used for boundary filling;

[0121] At each time index, the node hydrological state prediction vector after adaptive lag compensation is output. The output vectors of all nodes are collected according to the time index to form a set of lag-compensated hydrological state prediction vectors.

[0122] In this embodiment, the flood risk classification map data and the flood warning threshold list data specifically include:

[0123] The flood evolution prediction results obtained by mapping the delayed compensation hydrological state prediction vector set are used to extract flow prediction data, water level prediction data and flood peak arrival time prediction data. Combined with the corresponding rainfall and watershed infrastructure spatial data, the inundation depth data, flow velocity data and arrival time data of each grid cell are calculated.

[0124] The inundation depth, flow velocity, and time of arrival data of each grid cell are matched with a preset set of risk classification conditions. The risk classification condition set consists of inundation depth thresholds, flow velocity thresholds, and time of arrival thresholds, and is divided into four levels:

[0125] When the flooding depth is greater than or equal to three meters, the flow velocity is greater than or equal to two meters per second, and the arrival time is less than or equal to one hour, the grid cell is marked as Level 1 risk;

[0126] When the flooding depth is between two and three meters, the flow velocity is between one and two meters per second, or the arrival time is between one and three hours, and the conditions for Level 1 risk are not met, the grid cell is marked as Level 2 risk.

[0127] When the flooding depth is between 0.5 and 2 meters, the flow velocity is between 0.5 and 1 meter per second, or the arrival time is between 3 and 6 hours, and the conditions for level I and II risks are not met, the grid cell is marked as level III risk.

[0128] When the inundation depth is less than 0.5 meters, the flow velocity is less than 0.5 meters per second, and the arrival time is greater than 6 hours, the grid cell is marked as low risk;

[0129] Under each time index, the risk level labels of all raster cells are combined into flood risk level data. The flood risk level data is spatially overlaid with the critical facility exposure data. The types and quantities of critical facilities covered under each risk level are extracted, and a flood warning threshold list is generated and output.

[0130] The adaptive neural network flood evolution simulation and risk assessment system according to an embodiment of the present invention includes the following modules:

[0131] The data acquisition and preprocessing module is used to collect watershed rainfall observation data and river water level observation data and generate multi-source hydrological and meteorological preprocessed data;

[0132] The river network topology construction module is used to transform multi-source hydrological and meteorological preprocessing data into node feature vectors and edge feature vectors and form spatiotemporal graph input data.

[0133] The liquid time constant neural network processing module is used to input spatiotemporal map data and perform continuous time state updates to generate a set of node hydrological state prediction vectors;

[0134] The physical consistency residual injection module is used to inject hydrodynamic physical consistency residuals into the set of node hydrological state prediction vectors to generate a set of physically constrained hydrological state prediction vectors.

[0135] The adaptive lag compensation module is used to perform adaptive lag compensation on the set of physically constrained hydrological state prediction vectors to generate a set of lag-compensated hydrological state prediction vectors.

[0136] The flood evolution result generation module is used to map the set of delayed compensation hydrological state prediction vectors into flow prediction data, water level prediction data and flood peak arrival time prediction data.

[0137] The risk assessment and early warning generation module is used to generate flood risk classification map data and flood early warning threshold list data.

[0138] Example 1:

[0139] To verify the feasibility of this invention in practice, it was applied to flood evolution prediction and risk assessment in a medium-sized river basin. This basin has a complex river network structure, many tributaries, and large spatial distribution differences in rainfall processes. Traditional methods relying on a single hydrological model have problems such as prediction lag, large peak error, and inaccurate risk classification in this area. In similar scenarios in the past, the arrival time of the flood peak was often delayed, and the risk of flooding of some key infrastructure was not identified in advance, resulting in a passive nature of flood control scheduling.

[0140] In practical applications, rainfall observation stations and river water level observation stations are first deployed within the watershed. Combined with satellite and radar rainfall inversion data, multi-source hydrological and meteorological preprocessed data of rainfall and water level are formed. After time alignment and missing data filling processing, a river network topology map containing the connection relationship between observation sections and river segments is constructed. The observation section information is converted into node feature vectors, and the river segment information is converted into edge feature vectors, forming complete spatiotemporal map input data.

[0141] The aforementioned spatiotemporal graph is input into a liquid time constant neural network with an embedded graph attention mechanism. The set of node hydrological state prediction vectors is obtained through continuous time state updates. During the prediction process, the network uses graph attention to calculate the weighted information of adjacent nodes and adjusts the state update speed in combination with the time constant parameter. This can automatically shorten the response time when heavy rainfall occurs and maintain stable prediction during the steady phase. Subsequently, the physical consistency residual is injected into the node hydrological state prediction vector to ensure that the changes in flow and water level satisfy the constraints of the continuity equation and momentum equation, thereby improving the physical rationality of the prediction results.

[0142] When dealing with the rainfall-runoff time lag effect, an adaptive lag compensation mechanism is used to adjust the time axis position and smooth the state values ​​of the physically constrained hydrological state prediction vector, ensuring alignment between flood peak prediction and actual measurements. In this process, the time lag estimate is combined with the time constant parameter to generate an adaptive time lag correction coefficient. Different nodes can obtain targeted compensation amplitudes during the flood peak formation and recession stages, reducing the deviation in peak time prediction.

[0143] The compensated prediction results are mapped to future flow forecast data, water level forecast data, and flood peak arrival time forecast data. Combined with the spatial distribution information of watershed infrastructure, the gridded inundation depth, flow velocity, and arrival time are calculated based on the exceedance probability of different years. Flood risk level data is generated according to the graded thresholds, and the exposure data of key facilities is superimposed to form a flood risk grading map and a flood warning threshold list. In actual operation, the risk grading map provides an intuitive spatial reference for flood control command, while the warning threshold list clarifies the triggering conditions under each risk level.

[0144] To verify the effectiveness, the prediction results of the method of this invention were compared and analyzed with those of a traditional single hydrological model during the same flood process. The invention controlled the maximum relative error of peak flow prediction within 2%, the peak time prediction error was 0 hours, and the root mean square error of flow change in the receding section was reduced by about 45% compared with the traditional method. In terms of risk classification accuracy, the consistency rate between the flood risk level generated by this invention and the actual inundation situation reached over 92%, and the early identification rate of risk areas of level II and above was improved by nearly 30%. For the impact assessment of critical infrastructure, this invention can provide early warning 6 hours before the arrival of the flood peak, making the deployment of defense measures more proactive.

[0145] The introduction of the method of this invention improves the accuracy and timeliness of flood evolution prediction and enhances the scientific nature of risk assessment. Especially in scenarios involving complex watersheds and multi-source data fusion, it demonstrates superior stability and reliability compared to traditional technologies. Table 1 shows the comparison data of the method of this invention and traditional methods on several key indicators in practical applications.

[0146] Table 1: Comparison of the Effectiveness of Flood Evolution Prediction and Risk Assessment

[0147]

[0148] As can be seen from the comparison results in Table 1, the method of this invention outperforms traditional methods in several core indicators of flood evolution prediction and risk assessment. Regarding hydrological prediction accuracy, the maximum relative error of peak flow prediction is reduced from 6.8% to 1.9% using traditional methods, a reduction of 72%. The peak occurrence time prediction error is reduced from 2 hours to 0.5 hours, improving the alignment accuracy of peak time. The root mean square error of the receding flow is also reduced from 184 m³ / s to 101 m³ / s. The flow rate was reduced by 45% (m³ / s), effectively improving the stability of the receding flood process. Regarding the accuracy of risk classification, the consistency rate between risk levels and actual measurements increased from 78% to 92%, an increase of 14 percentage points. The early identification rate of level II and above risk areas increased from 61% to 79%, an increase of 30%, making the deployment of defense measures more targeted and timely. In terms of early warning response capabilities, the method of this invention can provide effective early warnings 6 hours before the flood peak arrives, increasing the lead time by 4 hours compared to traditional methods. Simultaneously, the accuracy of the critical facility coverage list increased from 80% to 95%, an increase of 15 percentage points, ensuring that key protection targets can be accurately identified. In terms of computational efficiency, the calculation time for a single prediction was shortened from 95 seconds to 54 seconds, an efficiency improvement of 43%. The utilization rate of multi-source data fusion also increased from 65% to 96%, an increase of 31%, indicating that this invention can more fully utilize multi-source observation information, thereby achieving comprehensive optimization in terms of accuracy, timeliness, and risk identification.

[0149] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An adaptive neural network-based method for flood evolution simulation and risk assessment, characterized in that, Includes the following steps: Step 1: Collect watershed rainfall observation data and river water level observation data, and preprocess them to obtain multi-source hydrological and meteorological preprocessed data; Step 2: Construct a river network topology map based on multi-source hydrological and meteorological preprocessing data, construct the observation cross-section as node feature vectors, construct the river segment as edge feature vectors, and form a spatiotemporal map input data; Step 3: Input the spatiotemporal graph input data into the liquid time constant neural network with embedded graph attention mechanism, perform continuous time state updates, and generate a set of node hydrological state prediction vectors; Step 4: Inject hydrodynamic physical consistency residuals into the nodal hydrological state prediction vector set to generate a physically constrained hydrological state prediction vector set. Specifically, under each time index, the hydrological state prediction vector of the current node and the hydrological state prediction vector of the adjacent nodes are read from the set of hydrological state prediction vectors of the nodes. Combined with the corresponding edge feature vectors, the inflow and outflow of the current node are summed respectively within the same time index. The difference between the sum of inflow and outflow is calculated to obtain the residual of the continuity equation. Under each time index, the time index interval is allocated based on the difference between the current node's hydrological state prediction vector and the previous time index node's hydrological state prediction vector. The residual of the momentum equation is calculated by combining the water level gradient along the river channel direction and the friction slope determined by the roughness and flow velocity. Under each time index, the residuals of the continuity equation and the residuals of the momentum equation are weighted and synthesized into a hydrodynamic physical consistency residual vector according to the adaptive weights. The adaptive weights are determined based on the monotonic mapping of the residual magnitude, the time index interval, and the time constant parameter. Under each time index, the hydrological state prediction vector of the current node is linearly corrected based on the hydrodynamic physical consistency residual vector and the residual injection gain coefficient to generate a physically constrained hydrological state prediction vector. The residual injection gain coefficient is determined by the time constant parameter and the time index interval and is limited to a preset closed interval. All time indices are processed repeatedly, and the physical constraint hydrological state prediction vectors are collected according to the time index to form a set of physical constraint hydrological state prediction vectors. Step 5: Perform adaptive lag compensation on the set of physically constrained hydrological state prediction vectors to generate a set of lag-compensated hydrological state prediction vectors. Step 6: Map the set of delayed compensation hydrological state prediction vectors to flow prediction data, water level prediction data and flood peak arrival time prediction data to form flood evolution prediction results data; Step 7: Based on the flood evolution prediction results, calculate the inundation depth, flow velocity, and arrival time data under different annual exceedance probabilities. Combine this with the spatial data of watershed infrastructure to generate flood risk level data and key facility exposure data, and output flood risk classification map data and flood warning threshold list data.

2. The adaptive neural network flood evolution simulation and risk assessment method according to claim 1, characterized in that, Step one includes: performing time series missing value filling, outlier removal, and unified time granularity processing on the watershed rainfall observation data to generate rainfall preprocessed data; performing time series missing value filling, outlier removal, and unified station benchmark processing on the river water level observation data to generate water level preprocessed data; and synchronizing and registering the rainfall preprocessed data and water level preprocessed data according to a unified time benchmark to construct multi-source hydrological and meteorological preprocessed data.

3. The adaptive neural network flood evolution simulation and risk assessment method according to claim 2, characterized in that, Step two includes: extracting water level preprocessing data and time indexes corresponding to each observation section from multi-source hydrological and meteorological preprocessing data; combining water level preprocessing data of the same observation section under continuous time index in sequence to construct node feature vectors; extracting rainfall preprocessing data, roughness data, and gradient data corresponding to each river segment from multi-source hydrological and meteorological preprocessing data; combining data of the same river segment under continuous time index in sequence to construct edge feature vectors; combining node feature vectors and edge feature vectors according to the upstream and downstream relationships connecting river segments to generate a river network topology map; arranging all node feature vectors and edge feature vectors in the river network topology map according to time index to form spatiotemporal map input data.

4. The adaptive neural network flood evolution simulation and risk assessment method according to claim 3, characterized in that, The processing of the liquid time constant neural network for the embedded graph attention mechanism in step three includes: The spatiotemporal graph input data is sorted in ascending order by time index. The node feature vector and edge feature vector under each time index are extracted. The learnable time constant parameter is initialized for each node. Normalization is performed according to the range of node feature vector to obtain the initial value of the time constant. The graph attention parameter is initialized for each edge and bound to the edge feature vector corresponding to the edge. The node feature vector under the first time index is used as the initial node hydrological state prediction vector. Under the current time index, the current node feature vector is connected with the hydrological state prediction vector of the node in the previous time index to form a query vector. For each incoming edge connected to a node, the hydrological state prediction vector of the previous time index node of the adjacent node is connected with the edge feature vector of the incoming edge to form a key vector and the corresponding value vector. Perform a dot product operation on the query vector and key vector according to their corresponding dimensions to obtain a relevance score. Multiply the relevance score element-wise with the graph attention parameters of the incoming edges, and add the result of the linear transformation of the edge feature vector to obtain an unnormalized attention score. Apply the graph attention parameters of the corresponding edges to the attention score. Normalize the scores of all incoming edges of the same node to obtain a set of attention weights. Perform a weighted summation on the corresponding value vectors using the attention weights to generate a graph attention aggregation vector. The graph attention aggregation vector is concatenated with the current node feature vector and input into the liquid time constant unit. The update coefficient is calculated based on the time constant parameter and the time index interval. The hydrological state prediction vector of the previous time index node and the state vector of the candidate node are weighted and updated to obtain the hydrological state prediction vector of the current node. The calculation is repeated on all time indices, and the node hydrological state prediction vectors of each time index are collected to form a set of node hydrological state prediction vectors.

5. The adaptive neural network flood evolution simulation and risk assessment method according to claim 1, characterized in that, Step five includes: Under each time index, the physical constraint hydrological state prediction vector of the current node and the physical constraint hydrological state prediction vector of several historical time indices are extracted from the set of physical constraint hydrological state prediction vectors. The cumulative curve of the corresponding rainfall process and the cumulative curve of the flow process are calculated. The time difference between the rainfall peak and the flow peak is determined based on the difference in the characteristic inflection point position of the two cumulative curves, and the time delay estimate of the current node is generated. Under each time index, the time delay estimate and corresponding time constant parameter of the current node are extracted, and the time constant parameter of the previous time index is extracted. Normalization is performed on the time delay estimate and time constant parameter to obtain the standardized time delay and standardized time constant. The rate of change of the time constant parameter between adjacent time indices is calculated to obtain the change in time constant. Weight pairs are determined based on the standardized time delay and the change in time constant, so that the sum of the weight pairs is one. Linear fusion is performed on the standardized time delay and standardized time constant according to the weight pairs to generate unrestricted correction coefficients. Monotonic normalization and upper and lower bound pruning are performed on the unrestricted correction coefficients to restrict the results to a preset closed interval to obtain adaptive time delay correction coefficients. Under each time index, the alignment position of the current node's physical constraint hydrological state prediction vector on the time axis is adjusted according to the adaptive time delay correction coefficient. Interpolation update is performed on the hydrological state prediction vector of the node that needs to be moved forward, and delayed smoothing update is performed on the hydrological state prediction vector of the node that needs to be moved backward. At each time index, the node hydrological state prediction vector after adaptive lag compensation is output. The output vectors of all nodes are collected according to the time index to form a set of lag-compensated hydrological state prediction vectors.

6. The adaptive neural network flood evolution simulation and risk assessment method according to claim 5, characterized in that, The flood risk classification map data and flood warning threshold list data specifically include: The flood evolution prediction results obtained by mapping the delayed compensation hydrological state prediction vector set are used to extract flow prediction data, water level prediction data and flood peak arrival time prediction data. Combined with the corresponding rainfall and watershed infrastructure spatial data, the inundation depth data, flow velocity data and arrival time data of each grid cell are calculated. The inundation depth, flow velocity, and time of arrival data of each grid cell are matched with a preset set of risk classification conditions. The risk classification condition set consists of inundation depth thresholds, flow velocity thresholds, and time of arrival thresholds, and is divided into four levels: When the flooding depth is greater than or equal to three meters, the flow velocity is greater than or equal to two meters per second, and the arrival time is less than or equal to one hour, the grid cell is marked as Level 1 risk; When the flooding depth is between two and three meters, the flow velocity is between one and two meters per second, or the arrival time is between one and three hours, and the conditions for Level 1 risk are not met, the grid cell is marked as Level 2 risk. When the flooding depth is between 0.5 and 2 meters, the flow velocity is between 0.5 and 1 meter per second, or the arrival time is between 3 and 6 hours, and the conditions for level I and II risks are not met, the grid cell is marked as level III risk. When the inundation depth is less than 0.5 meters, the flow velocity is less than 0.5 meters per second, and the arrival time is greater than 6 hours, the grid cell is marked as low risk; Under each time index, the risk level labels of all grid cells are combined into flood risk level data. The flood risk level data is spatially overlaid with the critical facility exposure data. The types and quantities of critical facilities covered under each risk level are extracted, and a flood warning threshold list is generated and output.

7. An adaptive neural network flood evolution simulation and risk assessment system, applied to the adaptive neural network flood evolution simulation and risk assessment method according to any one of claims 1 to 6, characterized in that, Includes the following modules: The data acquisition and preprocessing module is used to collect watershed rainfall observation data and river water level observation data and generate multi-source hydrological and meteorological preprocessed data; The river network topology construction module is used to transform multi-source hydrological and meteorological preprocessing data into node feature vectors and edge feature vectors and form spatiotemporal graph input data. The liquid time constant neural network processing module is used to input spatiotemporal map data and perform continuous time state updates to generate a set of node hydrological state prediction vectors; The physical consistency residual injection module is used to inject hydrodynamic physical consistency residuals into the set of node hydrological state prediction vectors to generate a set of physically constrained hydrological state prediction vectors. The adaptive lag compensation module is used to perform adaptive lag compensation on the set of physically constrained hydrological state prediction vectors to generate a set of lag-compensated hydrological state prediction vectors. The flood evolution result generation module is used to map the set of delayed compensation hydrological state prediction vectors into flow prediction data, water level prediction data and flood peak arrival time prediction data. The risk assessment and early warning generation module is used to generate flood risk classification map data and flood early warning threshold list data.

Images