Flood path prediction method based on particle swarm adaptive clustering

Abstract

The invention discloses a flood path prediction method based on particle swarm adaptive clustering, and the method comprises the steps: S1, collecting multi-source heterogeneous data, and carrying out the preprocessing of the multi-source heterogeneous data; s2, constructing a graph neural network model based on the multi-source heterogeneous data; s3, performing key propagation path prediction on the target area by using the graph neural network model; s4, performing sub-region division on the target region according to the initial reference information; s5, for each sub-region, initializing a particle swarm and setting flood propagation path parameters of particles; s6, carrying out iterative optimization on the particle swarm based on a self-adaptive clustering strategy; s7, inputting the new multi-source heterogeneous data into the graph neural network model to update the prediction result of the key propagation path; and S8, forming a flood path prediction result of the whole target area. According to the method, dynamic particle swarm optimization refined to the sub-region level also provides higher reliability for accurately identifying waterlogging-prone points and endangered regions, and the accuracy of flood prevention and drainage scheme planning and early warning information issuing is improved.

Description

Technical Field

[0001] The present invention relates to the field of flood technology, and in particular to a flood path prediction method based on particle swarm adaptive grouping. Background Art

[0002] With the frequent occurrence of extreme climate events and the continuous acceleration of urbanization, flood disasters have become an important factor threatening the safety of people's lives and property in many areas. In order to effectively carry out flood prevention warning and emergency response, flood path prediction technology has received widespread attention in recent years. On the one hand, accurately predicting the propagation trajectory of floods in river basins or urban areas is crucial to formulating timely and effective defense measures and personnel evacuation strategies; on the other hand, with the rapid growth of multi-source data from remote sensing monitoring, hydrological observation and meteorological forecasting, how to make full use of and integrate these data has become a hot topic in current research.

[0003] In the existing technology, common flood path prediction methods are mainly divided into two categories: one is a numerical simulation method based on physical mechanisms, which uses fluid dynamics models to calculate the flow velocity, flow direction and inundation range of floods in a region; the other is to rely on machine learning or data-driven models to predict the propagation process of floods by training historical hydrological and meteorological data. The existing methods have improved the accuracy and efficiency of flood path prediction to a certain extent, but there are still some problems that are difficult to ignore. First, the physical simulation method has very strict requirements on boundary conditions and initial conditions and often has high computational overhead. In scenarios with complex terrain and multi-source heterogeneous data coupling, it is difficult to achieve real-time or near real-time prediction. Secondly, data-driven models often rely on a large amount of high-quality training data, have insufficient adaptability to sudden abnormal conditions (floods caused by extreme rainfall or geological disasters), and are easily affected by regional differences and data missing in terms of prediction accuracy.

[0004] In addition, existing technologies generally lack the ability to dynamically couple multi-source data and cannot fully explore the combined impact of terrain undulation, vegetation distribution and meteorological factors on flood propagation paths. Most studies only update data in offline mode, and the prediction results often lag behind the actual flood evolution speed, making it difficult to meet the real-time needs of emergency response. On the other hand, although some flood path prediction methods based on traditional optimization algorithms have certain optimization capabilities, they are easy to fall into local optimal solutions in small areas or complex terrains, and lack the ability to adapt to multi-target or multi-region differentiated predictions.

[0005] In summary, existing technologies have shortcomings in real-time prediction with high temporal and spatial resolution, dynamic fusion of multi-source heterogeneous data, and adaptive optimization for complex terrain. It is difficult to quickly and accurately identify the key paths of flood propagation and conduct refined early warnings, which brings challenges to current flood prevention and emergency management. For this reason, how to improve the prediction accuracy in a small area and adapt to changing hydrological conditions while utilizing multi-source data for real-time updates has become a technical problem that needs to be urgently solved in the field of flood path prediction. Summary of the invention

[0006] One object of the present invention is to propose a flood path prediction method based on particle swarm adaptive clustering. The dynamic particle swarm optimization refined to the sub-region level of the present invention also provides higher reliability for accurately identifying flood-prone points and endangered areas, and improves the accuracy of flood control and drainage plan planning and early warning information release.

[0007] A flood path prediction method based on particle swarm adaptive clustering according to an embodiment of the present invention includes the following steps:

[0008] S1. Collect multi-source heterogeneous data and pre-process the multi-source heterogeneous data;

[0009] S2. constructing a graph neural network model based on the multi-source heterogeneous data, taking the regional features in the historical flood data as nodes, and taking the interaction relationship formed by the terrain features and the meteorological data as edges;

[0010] S3. Use the graph neural network model to predict the key propagation path of the target area and output the initial reference information of the key propagation path;

[0011] S4. Divide the target area into sub-areas according to the initial reference information, and determine the boundary range of each sub-area;

[0012] S5. For each sub-region, initialize the particle group and set the flood propagation path parameters of the particles, so that each particle represents the flow velocity, flow direction or residence time parameter related to the flood propagation in the sub-region;

[0013] S6. Iteratively optimize the particle swarm based on the adaptive swarming strategy, calculate the fitness value and dynamically adjust the particle swarm size and algorithm parameters during the iteration process, and use the local data in the sub-region to correct the search direction and speed of the particles in real time;

[0014] S7. After receiving new multi-source heterogeneous data, the new multi-source heterogeneous data is input into the graph neural network model to update the prediction results of the key propagation paths, and the sub-region division and particle swarm parameters are adjusted accordingly;

[0015] S8. Integrate the particle swarm optimization results of each sub-region to form the overall flood path prediction result of the target area.

[0016] Optionally, the S1 includes the following steps:

[0017] S11. Collect historical flood data D h , including the flood propagation path, peak flow and duration of the i-th historical flood event;

[0018] S12. Obtaining terrain feature data D t , including the slope, aspect and surface roughness of the i-th geographic unit;

[0019] S13. Collect meteorological data D m , including rainfall intensity, wind speed and temperature at the i-th moment in the target area;

[0020] S14. Obtain water level data from hydrological station D w , including the water depth, velocity and flow rate of the i-th hydrological station;

[0021] S15. Obtain satellite remote sensing image data D r , including the water distribution, vegetation coverage and flooded range within the target area of ​​the i-th satellite image;

[0022] S16. Standardize the multi-source heterogeneous data obtained in S11-S15 to construct a unified multi-source heterogeneous data set D:

[0023]

[0024] in, represents the set of the i-th flood events in the historical flood data, d hi Contains information about the ith flood event, a total of n flood events, Represents a collection of terrain feature data, g j is the topographical feature of the jth geographic unit, with a total of m geographic units, represents a collection of meteorological data, r k is the meteorological data at the kth time point, with a total of p time points, Represents the collection of water level data of hydrological stations, s l is the real-time water level data of the lth hydrological station, with a total of q hydrological stations. represents the collection of satellite remote sensing image data, v o is the o-th satellite remote sensing image data, with a total of k images, ∪ represents the set union operator, which is used to aggregate the single elements in each type of data;

[0025] S17. Map all data in the multi-source heterogeneous dataset D to the interval [0,1] through linear normalization, and use the time interpolation method to complete the missing time point data, and the spatial interpolation method to supplement the missing spatial data in the region, and obtain the preprocessed multi-source heterogeneous dataset D c .

[0026] Optionally, S2 includes the following steps:

[0027] S21. From the preprocessed multi-source heterogeneous dataset D c In the process, the historical flood data D in the target area and the water level data of the hydrological station are selected as the node feature sources to construct the node set N = {n1, n2, ..., n u}, where n i ,i∈{1,2,...,u} represents the historical flood event characteristics contained in the i-th node and the corresponding hydrological station water level characteristics;

[0028] S22. Combining the multi-source heterogeneous dataset D c The terrain feature data, meteorological data and satellite remote sensing image data in the database are used to construct a multi-relation adjacency matrix. Where R represents the number of different relationship types, The node connectivity under the rth relationship type is shown, and the edge set E = {e ij};

[0029] S23. Define node feature matrix Where u represents the number of nodes, f represents the joint feature dimension contained in each node, and the characteristics of each node are composed of the characteristics of historical flood events, the characteristics of water level data of hydrological stations, and the local geographical and meteorological attributes corresponding to the node;

[0030] S24. Build a graph neural network model GNN. Based on the multi-relation adjacency matrix and the node feature matrix X, introduce a multi-head attention mechanism to perform weighted aggregation on different relationship types. Define the node feature update as:

[0031]

[0032] Among them, H (l) Represents the node feature matrix of the lth layer of the graph neural network. Initially, H (0) =X,W (l) is the trainable weight matrix of the lth layer, α (m,r) is the weight coefficient of the mth attention head under the rth type of relationship, M represents the number of attention heads, σ represents the activation function, and ∥ represents the concatenation operation.

[0033] Optionally, the edge set is constructed based on the following multi-source relationship:

[0034] The spatial proximity between nodes is identified using the slope and aspect in terrain feature data;

[0035] Using rainfall intensity or wind speed in meteorological data to identify temporal correlations between nodes;

[0036] The regional flooding range in satellite remote sensing image data is used to identify the flood impact correlation between nodes;

[0037] The water depth and flow information in the water level data of the integrated hydrological station are used to calculate the hydrodynamic coupling strength between nodes.

[0038] Optionally, S3 includes the following steps:

[0039] S31. Calculate the node feature matrix H based on the graph neural network model (L) , where L represents the total number of layers of the graph neural network and the node feature matrix H (L) It represents the importance and dynamic characteristics of each node in the flood propagation network under the comprehensive conditions of multi-source heterogeneous data;

[0040] S32. Based on the node feature matrix H (L) For each node n in the target area i Calculate its node importance score I(n i ), which is used to evaluate the criticality of a node in the flood propagation process:

[0041]

[0042] in, Indicates that under the rth type of relationship, node n i The node n to which it is connected j The flood impact weight, Represents node n in layer L i The flood propagation characteristics of , R represents the number of multi-source relationship types, and u is the total number of nodes;

[0043] S33. According to the node importance score I(n i ) in descending order, and select the K nodes with the highest scores to form the key node set N k ;

[0044] S34. For the key node set N k For each node in, extract its node features The connection relationship in the multi-relation adjacency matrix is ​​used to identify the flood propagation path between key nodes by traversing the adjacent nodes, and generate a preliminary propagation path set:

[0045] P={p1,p2,…,p M};

[0046] Among them, p M represents the Mth propagation path, including the node sequence and the flow or water level changes between nodes in the path;

[0047] S35. For each path p in the preliminary propagation path set P j Calculate the path weight W(p j ), the path weight is used to quantify the priority of the path in flood propagation:

[0048]

[0049] Among them, I(n ki ) is the path p j The importance score of the node reflects the influence of the node in flood propagation, w ij is the edge e in the path ij The weight of represents the intensity of flood propagation between nodes;

[0050] S36. According to the path weight W(p j ) Sort the preliminary propagation path set P, select the first M' paths with the highest weight to form the key propagation path set, the key path set P k Represents the most likely flood propagation path within the target area:

[0051] P k ={p k1 ,p k2 ,…,p kM′}.

[0052] Optionally, S4 includes the following steps:

[0053] S41. According to the key propagation path set P k Extract the node set N on each path k and the connection relationship set E k , where the node set N k Represents the flood propagation area nodes on the key propagation path, and the connection relationship set E k Represents the flood propagation association between nodes;

[0054] S42. Based on the key propagation path set P k Calculate the node cluster center C in the target area = {c1, c2, ..., c q}, where q represents the number of sub-regions initially divided, and each cluster center c q Determined by:

[0055]

[0056] Among them, w jFor node n j The importance weight reflects the criticality of the node in flood propagation, p j Represents node n j The spatial position vector of

[0057] S43. Node set N in the target area k Divide the space and divide each node n j Assign to the nearest cluster center c i , the sub-region division is completed, and the sub-region set is represented as R:

[0058]

[0059] Among them, R i represents the i-th sub-region, including the cluster center c i All nodes closest to each other, ∥p j -c i ∥ represents node n j With cluster center c i The Euclidean distance of

[0060] S44. Based on the sub-region set R, the connection relationship set E between the nodes is combined k For each sub-region R i The boundary range is further optimized and the boundary function B of each sub-region is calculated. i :

[0061]

[0062] Among them, B i represents the boundary of the ith subregion, including all edges e connected to other subregions jk .

[0063] Optionally, S5 includes the following steps:

[0064] S51. For each sub-region R i , set the particle swarm size P for each sub-region i , in sub-region R i Construct particle swarm collections separately:

[0065]

[0066] S52. In the sub-area R i In the search space determined by the flow velocity, flow direction and residence time parameters related to flood propagation, the parameter vector of the particle is defined Represents the initial position of the jth particle:

[0067]

[0068] in, Indicates that the jth particle is in the sub-region R i The initial flow velocity reflects the magnitude of the flood flow speed. Indicates the initial flow angle, representing the direction of flood propagation. It represents the initial detention time, reflecting the detention or stay characteristics of the flood in the sub-area;

[0069] S53. Parameter vector for each particle Perform random initialization to make the particle swarm evenly distributed within the feasible range. The random initialization is:

[0070]

[0071] Among them, d∈{1,2,3} corresponds to the parameters of the three dimensions of flow velocity, flow direction and residence time, respectively. and They represent the minimum and maximum feasible values ​​of the dth parameter respectively, and rand(0,1) represents a random number in the interval [0,1];

[0072] S54. For each particle Set the initial velocity vector Let the initial velocity components of each dimension obey random or zero distribution, and initialize the particle swarm set G i Stored in sub-region R i in the particle swarm database;

[0073] S55. Output each sub-region R i The initial position and initial velocity of the inner particle group are used to represent the flood propagation path parameters of the sub-region and complete the initialization of the particle group in each sub-region.

[0074] Optionally, the S6 includes the following steps:

[0075] S61. In the initialized particle swarm set G i Based on the i Define the upper limit of the number of iterations and execute the particle swarm optimization process based on the adaptive swarming strategy;

[0076] S62. At each iteration, calculate the particles The fitness value of The fitness value reflects the performance of the particle in predicting the accuracy and stability of the flood propagation path. The fitness function is defined as:

[0077]

[0078] Among them, D iRepresents sub-region R i The local data within the sub-region, including topographic features, meteorological data, and water level information of hydrological stations, are used to measure the degree of fit of the current particle in the flood propagation process in the sub-region;

[0079] S63. According to the fitness value Particle The individual optimal position and the global optimal position gbest in the sub-region i Update and save the position with the highest fitness value as the individual optimal position of the particle and the global optimal position of the sub-region;

[0080] S64. Particles Speed With the position updated, the velocity is updated as:

[0081]

[0082] Among them, ω is the inertia weight, balancing the exploration and development capabilities, c1 and c2 are learning factors, rand1(0,1) and rand2(0,1) represent random numbers in the interval [0,1];

[0083] S65. During the iteration process, the local data changes in the sub-region are detected in real time. If the key monitoring indicators change significantly, the particle swarm size and algorithm parameters ω, c1, c2 are dynamically adjusted according to the adaptive clustering strategy, and the particles are The search direction and speed are modified, and some particle groups are reinitialized or merged to meet the dynamic propagation conditions of floods;

[0084] S66. Repeat the iterative steps S61-S65 until the upper limit of the number of iterations is reached or the convergence condition is met, and output the sub-region R i The optimal position set of the inner particle swarm {gbest i}, which represents the set of particle solutions with the highest fitness in the flood propagation path prediction of this sub-region.

[0085] The beneficial effects of the present invention are:

[0086] (1) The present invention integrates multi-source heterogeneous information such as terrain feature data, water level data from hydrological stations, meteorological data, and satellite remote sensing image data during particle swarm initialization and iterative optimization, and divides the target area into differentiated sub-regions based on the extraction of key propagation paths. It can independently iterate and dynamically adjust algorithm parameters for the local characteristics of each sub-region. Under different hydrological conditions or sudden increases in upstream water volume, the adaptive clustering mechanism can quickly locate susceptible areas and optimize particle swarm search, so that real-time flood path prediction can take into account both the overall situation and achieve higher accuracy in local areas.

[0087] (2) Based on the traditional physical model and empirical model, the present invention introduces an improved graph neural network to deeply explore the correlation between historical flood data, topography and meteorological conditions. By extracting key nodes and constructing a multi-relationship adjacency matrix, the main paths of flood propagation can be accurately captured. At the same time, the adaptive clustering strategy pays attention to the changes in key monitoring indicators of the sub-region in real time during the particle swarm iteration process, and can dynamically adjust the particle swarm scale and parameters, avoiding the shortcomings of a single particle swarm falling into local optimality or lack of adaptability when facing different hydrological scenarios, and realizing more flexible and efficient prediction of flood propagation processes with high temporal and spatial dynamics.

[0088] (3) After refining and screening the key propagation paths, the present invention performs independent particle swarm iterative search on each local area based on sub-region division, and merges the optimal solutions of each sub-region at the global level, thereby achieving comprehensive control of the overall flood propagation situation, and can continuously correct the flood velocity, flow direction and residence time prediction indicators under the real-time update of multi-source data, so that the flood control command department can be more proactive and targeted when responding to sudden disasters. At the same time, the dynamic particle swarm optimization refined to the sub-region level also provides higher reliability for accurately identifying flood-prone points and endangered areas, and improves the accuracy of flood control and drainage plan planning and early warning information release. BRIEF DESCRIPTION OF THE DRAWINGS

[0089] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0090] Figure 1 The present invention provides a flow chart of a flood path prediction method based on particle swarm adaptive clustering. DETAILED DESCRIPTION

[0091] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0092] refer to Figure 1 , a flood path prediction method based on particle swarm adaptive clustering, comprising the following steps:

[0093] S1. Collect multi-source heterogeneous data and pre-process the multi-source heterogeneous data;

[0094] S2. Build a graph neural network model based on multi-source heterogeneous data, taking the regional features in the historical flood data as nodes and the interaction between the terrain features and the meteorological data as edges;

[0095] S3. Use the graph neural network model to predict the key propagation path of the target area and output the initial reference information of the key propagation path;

[0096] S4. Divide the target area into sub-areas according to the initial reference information, and determine the boundary range of each sub-area;

[0097] S5. For each sub-region, initialize the particle group and set the flood propagation path parameters of the particles, so that each particle represents the flow velocity, flow direction or residence time parameter related to the flood propagation in the sub-region;

[0098] S6. Iteratively optimize the particle swarm based on the adaptive swarming strategy, calculate the fitness value and dynamically adjust the particle swarm size and algorithm parameters during the iteration process, and use the local data in the sub-region to correct the search direction and speed of the particles in real time;

[0099] S7. When receiving new multi-source heterogeneous data, the new multi-source heterogeneous data is input into the graph neural network model to update the prediction results of the key propagation path, and the sub-region division and particle swarm parameters are adjusted accordingly;

[0100] S8. Integrate the particle swarm optimization results of each sub-region to form the overall flood path prediction result of the target area.

[0101] In this implementation, S1 includes the following steps:

[0102] S11. Collect historical flood data D h , including the flood propagation path, peak flow and duration of the i-th historical flood event;

[0103] S12. Obtaining terrain feature data D t , including the slope, aspect and surface roughness of the i-th geographic unit;

[0104] S13. Collect meteorological data D m , including rainfall intensity, wind speed and temperature at the i-th moment in the target area;

[0105] S14. Obtain water level data from hydrological station D w , including the water depth, velocity and flow rate of the i-th hydrological station;

[0106] S15. Obtain satellite remote sensing image data D r , including the water distribution, vegetation coverage and flooded range within the target area of ​​the i-th satellite image;

[0107] S16. Standardize the multi-source heterogeneous data obtained in S11-S15 to construct a unified multi-source heterogeneous data set D:

[0108]

[0109] in, represents the set of the i-th flood events in the historical flood data, d hi Contains information about the ith flood event, a total of n flood events, Represents a collection of terrain feature data, g j is the topographical feature of the jth geographic unit, with a total of m geographic units, represents a collection of meteorological data, r k is the meteorological data at the kth time point, with a total of p time points, Represents the collection of water level data of hydrological stations, s l is the real-time water level data of the lth hydrological station, with a total of q hydrological stations. represents the collection of satellite remote sensing image data, v o is the o-th satellite remote sensing image data, with a total of k images, ∪ represents the set union operator, which is used to aggregate the single elements in each type of data;

[0110] S17. Map all data in the multi-source heterogeneous dataset D to the interval [0,1] through linear normalization, and use the time interpolation method to complete the missing time point data, and the spatial interpolation method to supplement the missing spatial data in the region, and obtain the preprocessed multi-source heterogeneous dataset D c .

[0111] In this implementation, S2 includes the following steps:

[0112] S21. From the preprocessed multi-source heterogeneous dataset D c In the process, the historical flood data D in the target area and the water level data of the hydrological station are selected as the node feature sources to construct the node set N = {n1, n2, ..., n u}, where n i ,i∈{1,2,...,u} represents the historical flood event characteristics contained in the i-th node and the corresponding hydrological station water level characteristics;

[0113] S22. Combining multi-source heterogeneous datasets D c The terrain feature data, meteorological data and satellite remote sensing image data in the database are used to construct a multi-relation adjacency matrix. Where R represents the number of different relationship types, The node connectivity under the rth relationship type is shown, and the edge set E = {e ij};

[0114] S23. Define node feature matrix Where u represents the number of nodes, f represents the joint feature dimension contained in each node, and the characteristics of each node are composed of the characteristics of historical flood events, the characteristics of water level data of hydrological stations, and the local geographical and meteorological attributes corresponding to the node;

[0115] S24. Build a graph neural network model GNN. Based on the multi-relation adjacency matrix and the node feature matrix X, introduce a multi-head attention mechanism to perform weighted aggregation on different relationship types. Define the node feature update as:

[0116]

[0117] Among them, H (l) Represents the node feature matrix of the lth layer of the graph neural network. Initially, H (0) =X,W (l) is the trainable weight matrix of the lth layer, α (m,r) is the weight coefficient of the mth attention head under the rth type of relationship, M represents the number of attention heads, σ represents the activation function, and ∥ represents the concatenation operation.

[0118] In this implementation, the edge set is constructed based on the following multi-source relationships:

[0119] The spatial proximity between nodes is identified using the slope and aspect in terrain feature data;

[0120] Using rainfall intensity or wind speed in meteorological data to identify temporal correlations between nodes;

[0121] The regional flooding range in satellite remote sensing image data is used to identify the flood impact correlation between nodes;

[0122] The water depth and flow information in the water level data of the integrated hydrological station are used to calculate the hydrodynamic coupling strength between nodes.

[0123] In this implementation, S3 includes the following steps:

[0124] S31. Calculate the node feature matrix H based on the graph neural network model (L) , where L represents the total number of layers of the graph neural network and the node feature matrix H (L) It represents the importance and dynamic characteristics of each node in the flood propagation network under the comprehensive conditions of multi-source heterogeneous data;

[0125] S32. Based on the node feature matrix H (L) For each node n in the target area i Calculate its node importance score I(n i ), which is used to evaluate the criticality of a node in the flood propagation process:

[0126]

[0127] in, Indicates that under the rth type of relationship, node n i The node n to which it is connected j The flood impact weight, Represents node n in layer L i The flood propagation characteristics of , R represents the number of multi-source relationship types, and u is the total number of nodes;

[0128] S33. According to the node importance score I(n i ) in descending order, and select the K nodes with the highest scores to form the key node set N k ;

[0129] S34. For the key node set N k For each node in, extract its node features The connection relationship in the multi-relation adjacency matrix is ​​used to identify the flood propagation path between key nodes by traversing the adjacent nodes, and generate a preliminary propagation path set:

[0130] P={p1,p2,…,p M};

[0131] Among them, p M represents the Mth propagation path, including the node sequence and the flow or water level changes between nodes in the path;

[0132] S35. For each path p in the preliminary propagation path set P j Calculate the path weight W(p j ), the path weight is used to quantify the priority of the path in flood propagation:

[0133]

[0134] Among them, I(n ki ) is the path p j The importance score of the node reflects the influence of the node in flood propagation, w ij is the edge e in the path ij The weight of represents the intensity of flood propagation between nodes;

[0135] S36. According to the path weight W(p j ) Sort the preliminary propagation path set P, select the first M' paths with the highest weight to form the key propagation path set, the key path set P k Represents the most likely flood propagation path within the target area:

[0136] P k ={p k1 ,p k2 ,…,pkM′}.

[0137] In this implementation, S4 includes the following steps:

[0138] S41. According to the key propagation path set P k Extract the node set N on each path k and the connection relationship set E k , where the node set N k Represents the flood propagation area nodes on the key propagation path, and the connection relationship set E k Represents the flood propagation association between nodes;

[0139] S42. Based on the key propagation path set P k Calculate the node cluster center C in the target area = {c1, c2, ..., c q}, where q represents the number of sub-regions initially divided, and each cluster center c q Determined by:

[0140]

[0141] Among them, w j For node n j The importance weight reflects the criticality of the node in flood propagation, p j Represents node n j The spatial position vector of

[0142] S43. Node set N in the target area k Divide the space and divide each node n j Assign to the nearest cluster center c i , the sub-region division is completed, and the sub-region set is represented as R:

[0143]

[0144] Among them, R i represents the i-th sub-region, including the cluster center c i All nodes closest to each other, ∥p j -c i ∥ represents node n j With cluster center c i The Euclidean distance of

[0145] S44. Based on the sub-region set R, the connection relationship set E between the nodes is combined k For each sub-region R i The boundary range is further optimized and the boundary function B of each sub-region is calculated. i :

[0146]

[0147] Among them, B i represents the boundary of the ith subregion, including all edges e connected to other subregions jk .

[0148] In this implementation, S5 includes the following steps:

[0149] S51. For each sub-region R i , set the particle swarm size P for each sub-region i , in sub-region R i Construct particle swarm collections separately:

[0150]

[0151] S52. In sub-area R i In the search space determined by the flow velocity, flow direction and residence time parameters related to flood propagation, the parameter vector of the particle is defined Represents the initial position of the jth particle:

[0152]

[0153] in, Indicates that the jth particle is in the sub-region R i The initial flow velocity reflects the magnitude of the flood flow speed. Indicates the initial flow angle, representing the direction of flood propagation. It represents the initial detention time, reflecting the detention or stay characteristics of the flood in the sub-area;

[0154] S53. Parameter vector for each particle Perform random initialization to make the particle swarm evenly distributed within the feasible range. The random initialization is:

[0155]

[0156] Among them, d∈{1,2,3} corresponds to the parameters of the three dimensions of flow velocity, flow direction and residence time, respectively. and They represent the minimum and maximum feasible values ​​of the dth parameter respectively, and rand(0,1) represents a random number in the interval [0,1];

[0157] S54. For each particle Set the initial velocity vector Let the initial velocity components of each dimension obey random or zero distribution, and initialize the particle swarm set G i Stored in sub-region R iin the particle swarm database;

[0158] S55. Output each sub-region R i The initial position and initial velocity of the inner particle group are used to represent the flood propagation path parameters of the sub-region and complete the initialization of the particle group in each sub-region.

[0159] In this implementation, S6 includes the following steps:

[0160] S61. In the initialized particle swarm set G i Based on the i Define the upper limit of the number of iterations and execute the particle swarm optimization process based on the adaptive swarming strategy;

[0161] S62. At each iteration, calculate the particles The fitness value of The fitness value reflects the performance of the particle in predicting the accuracy and stability of the flood propagation path. The fitness function is defined as:

[0162]

[0163] Among them, D i Represents sub-region R i The local data within the sub-region, including topographic features, meteorological data, and water level information of hydrological stations, are used to measure the degree of fit of the current particle in the flood propagation process in the sub-region;

[0164] S63. According to the fitness value Particle The individual optimal position pbest pij and the global optimal position gbest in the sub-region i Update and save the position with the highest fitness value as the individual optimal position of the particle and the global optimal position of the sub-region;

[0165] S64. Particles Speed With the position updated, the velocity is updated as:

[0166]

[0167] Among them, ω is the inertia weight, balancing the exploration and development capabilities, c1 and c2 are learning factors, rand1(0,1) and rand2(0,1) represent random numbers in the interval [0,1];

[0168] S65. During the iteration process, the local data changes in the sub-region are detected in real time. If the key monitoring indicators change significantly, the particle swarm size and algorithm parameters ω, c1, c2 are dynamically adjusted according to the adaptive clustering strategy, and the particles are The search direction and speed are modified, and some particle groups are reinitialized or merged to meet the dynamic propagation conditions of floods;

[0169] S66. Repeat the iterative steps S61-S65 until the upper limit of the number of iterations is reached or the convergence condition is met, and output the sub-region R i The optimal position set of the inner particle swarm {gbest i}, which represents the set of particle solutions with the highest fitness in the flood propagation path prediction of this sub-region.

[0170] Embodiment 1:

[0171] Example: Application of the flood path prediction method based on deep learning-assisted particle swarm adaptive clustering in a mountainous area flood disaster on June 15, 2024

[0172] At 15:30 on June 15, 2024, a hydrological monitoring station in a certain mountainous area issued a rainstorm warning signal. The precipitation in the area suddenly increased that day, reaching 120 mm in a short period of time, and was expected to continue until 20:00 that evening. At this time, the flood control departments of the low-lying villages along the river received the warning information and quickly activated the flood path prediction system of the present invention to monitor the real-time dynamics of flood propagation and guide residents to evacuate.

[0173] The target area includes five small villages, two cross-river bridges and some farmland irrigation facilities. The terrain is complex and the slope is steep. The heavy rain triggered the rapid convergence of floods at the two main water confluence points upstream and spread downstream along the low-lying areas. To cope with this complex scenario, the flood control department immediately collected multi-source data including real-time rainfall, river water level, flow rate and satellite remote sensing image data after receiving the heavy rain warning. The data covered an area of ​​approximately 150 square kilometers in the target area.

[0174] At 15:35, the first batch of real-time data collected by the deployed multi-source monitoring equipment included:

[0175] The water level monitoring point of the upstream hydrological station (No.: WS-01) showed that the water level rose sharply to 4.2 meters, and the flow rate reached 2.8 meters per second;

[0176] The rainfall monitoring point (No.: RS-02) measured the rainfall intensity during this period to be 40 mm per hour;

[0177] Satellite remote sensing images show that 1.8 square kilometers of low-lying areas upstream have been flooded, and the water flow path has spread in the southeast direction.

[0178] The system first standardizes the collected data and inputs the terrain, rainfall intensity and water level information into the constructed graph neural network model. The graph neural network model analyzes the current conditions based on historical data and generates preliminary reference information for key propagation paths:

[0179] Path 1: Propagating from WS-01 along the downstream direction of the river, it is expected to reach Village A in 30 minutes;

[0180] Path 2: Water flows in low-lying areas, which may threaten Bridge B in 45 minutes;

[0181] Based on the above critical path information, the system automatically divides the target area into 6 sub-areas (R1 to R6), each of which corresponds to different hydrological characteristics and terrain conditions. In sub-area R2 (near village A), the system simulates flood propagation through particle swarm algorithm. The parameters for particle initialization include:

[0182] The initial flow velocity is 3.2 m / s;

[0183] The initial flow direction is 15° south-east;

[0184] The estimated detention time is 12 minutes.

[0185] At 15:40, the particle swarm algorithm completed the 10th iterative optimization and obtained the global optimal solution of the particle swarm. The prediction results showed that the flood would submerge the low-lying areas of Village A at 16:10, with an estimated flooded area of ​​2.5 square kilometers and a water depth of more than 1.2 meters.

[0186] At 16:00, rainfall monitoring point RS-03 sent out an abnormal signal, and the rainfall intensity suddenly increased to 55 mm per hour. The system detected abnormal water level fluctuations in sub-area R4 (downstream of the dam) through an adaptive clustering strategy, automatically adjusted the size and algorithm parameters of the particle swarm in the R4 area, and re-simulated the propagation path of the flood in the low-lying area. The results showed that the flood flow rate accelerated from 3.5 m / s to 4.1 m / s, and it was expected to threaten the northern area of ​​​​Village B before 16:20.

[0187] Based on the predicted path and disaster area map generated by the system, the flood control department immediately launched the emergency evacuation plan for Village A and Village B, and mobilized materials to reinforce Bridge B.

[0188] During the entire flood propagation process, the present invention is compared and analyzed with the traditional HEC-RAS numerical model. The following Table 1 shows the key indicators:

[0189] Table 1 Comparison of key indicators between the method of the present invention and the traditional method in flood path prediction

[0190]

[0191]

[0192] By comparison, it can be seen that the present invention not only significantly shortens the processing time and simulation time, but also excels in propagation path accuracy and flooded area prediction accuracy. The prediction map finally generated by the system at 16:30 is basically consistent with the distribution of the actual flooded area, which wins precious response time for the flood control department, successfully avoids thousands of people from being threatened by floods, and reduces the loss of facilities in the area by about 30%.

[0193] This example verifies the superior performance of the present invention under complex dynamic flood propagation conditions, and shows significant advantages in real-time, refinement and decision support. The comparative data further illustrates the applicability and effectiveness of the method of the present invention in actual flood prevention work.

[0194] The present invention integrates multi-source heterogeneous information such as terrain feature data, water level data from hydrological stations, meteorological data and satellite remote sensing image data during particle swarm initialization and iterative optimization, and divides the target area into differentiated sub-regions based on the extraction of key propagation paths. It can independently iterate and dynamically adjust algorithm parameters for the local characteristics of each sub-region. Under different hydrological conditions or sudden increases in upstream water volume, the adaptive clustering mechanism can quickly locate susceptible areas and optimize particle swarm search, so that real-time flood path prediction can take into account both the overall situation and achieve higher accuracy in local areas.

[0195] Based on traditional physical models and empirical models, the present invention introduces an improved graph neural network to deeply mine the correlation between historical flood data, topography and meteorological conditions. By extracting key nodes and constructing a multi-relationship adjacency matrix, the main paths of flood propagation can be accurately captured. At the same time, the adaptive clustering strategy pays attention to the changes in key monitoring indicators of the sub-region in real time during the particle swarm iteration process, and can dynamically adjust the size and parameters of the particle swarm, avoiding the shortcomings of a single particle swarm falling into local optimality or lack of adaptability when facing different hydrological scenarios, and realizing more flexible and efficient prediction of flood propagation processes with high temporal and spatial dynamics.

[0196] After refining and screening the key propagation paths, the present invention performs independent particle swarm iterative search on each local area based on sub-area division, and merges the optimal solutions of each sub-area at the global level, thereby achieving comprehensive control of the overall flood propagation situation, and can continuously correct flood velocity, flow direction and residence time prediction indicators under the real-time update of multi-source data, so that flood control command departments can be more proactive and targeted when responding to sudden disasters. At the same time, dynamic particle swarm optimization refined to the sub-area level also provides higher reliability for accurately identifying flood-prone points and endangered areas, and improves the accuracy of flood control and drainage plan planning and early warning information release.

[0197] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A flood path prediction method based on particle swarm adaptive clustering, characterized in that: The steps include: S1. Collect multi-source heterogeneous data and pre-process the multi-source heterogeneous data; S2. constructing a graph neural network model based on the multi-source heterogeneous data, taking the regional features in the historical flood data as nodes, and taking the interaction relationship formed by the terrain features and the meteorological data as edges; S3. Use the graph neural network model to predict the key propagation path of the target area and output the initial reference information of the key propagation path; S4. Divide the target area into sub-areas according to the initial reference information, and determine the boundary range of each sub-area; S5. For each sub-region, initialize the particle group and set the flood propagation path parameters of the particles, so that each particle represents the flow velocity, flow direction or residence time parameter related to the flood propagation in the sub-region; S6. Iteratively optimize the particle swarm based on the adaptive swarming strategy, calculate the fitness value and dynamically adjust the particle swarm size and algorithm parameters during the iteration process, and use the local data in the sub-region to correct the search direction and speed of the particles in real time; S7. After receiving new multi-source heterogeneous data, the new multi-source heterogeneous data is input into the graph neural network model to update the prediction results of the key propagation paths, and the sub-region division and particle swarm parameters are adjusted accordingly; S8. Integrate the particle swarm optimization results of each sub-region to form the overall flood path prediction result of the target area.

2. The flood path prediction method based on particle swarm adaptive clustering according to claim 1 is characterized in that: The S1 comprises the following steps: S11. Collect historical flood data D h , including the flood propagation path, peak flow and duration of the i-th historical flood event; S12. Obtaining terrain feature data D t , including the slope, aspect and surface roughness of the i-th geographic unit; S13. Collect meteorological data D m , including rainfall intensity, wind speed and temperature at the i-th moment in the target area; S14. Obtain water level data from hydrological station D w , including the water depth, velocity and flow rate of the i-th hydrological station; S15. Obtain satellite remote sensing image data D r , including the water distribution, vegetation coverage and flooded range within the target area of ​​the i-th satellite image; S16. Standardize the multi-source heterogeneous data obtained in S11-S15 to construct a unified multi-source heterogeneous data set D: in, represents the set of the i-th flood events in the historical flood data, d hi Contains information about the ith flood event, a total of n flood events, Represents a collection of terrain feature data, g j is the topographical feature of the jth geographic unit, with a total of m geographic units, represents a collection of meteorological data, r k is the meteorological data at the kth time point, with a total of p time points, Represents the collection of water level data of hydrological stations, s l is the real-time water level data of the lth hydrological station, with a total of q hydrological stations. represents the collection of satellite remote sensing image data, v o is the o-th satellite remote sensing image data, with a total of k images, ∪ represents the set union operator, which is used to aggregate the single elements in each type of data; S17. Map all data in the multi-source heterogeneous dataset D to the interval [0,1] through linear normalization, and use the time interpolation method to complete the missing time point data, and the spatial interpolation method to supplement the missing spatial data in the region, and obtain the preprocessed multi-source heterogeneous dataset D c .

3. The flood path prediction method based on particle swarm adaptive clustering according to claim 1 is characterized in that: The S2 comprises the following steps: S21. From the preprocessed multi-source heterogeneous dataset D c In the process, the historical flood data D in the target area and the water level data of the hydrological station are selected as the node feature sources to construct the node set N = {n1, n2, ..., n u }, where n i ,i∈{1,2,...,u} represents the historical flood event characteristics contained in the i-th node and the corresponding hydrological station water level characteristics; S22. Combining the multi-source heterogeneous dataset D c The terrain feature data, meteorological data and satellite remote sensing image data in the database are used to construct a multi-relation adjacency matrix. Where R represents the number of different relationship types, The node connectivity under the rth relationship type is shown, and the edge set E = {e ij }; S23. Define node feature matrix Where u represents the number of nodes, f represents the joint feature dimension contained in each node, and the characteristics of each node are composed of the characteristics of historical flood events, the characteristics of water level data of hydrological stations, and the local geographical and meteorological attributes corresponding to the node; S24. Build a graph neural network model GNN. Based on the multi-relation adjacency matrix and the node feature matrix X, introduce a multi-head attention mechanism to perform weighted aggregation on different relationship types. Define the node feature update as: Among them, H (l) Represents the node feature matrix of the lth layer of the graph neural network. Initially, H (0) =X,W (l) is the trainable weight matrix of the lth layer, α (m,r) is the weight coefficient of the mth attention head under the rth type of relationship, M represents the number of attention heads, σ represents the activation function, and ∥ represents the concatenation operation.

4. The flood path prediction method based on particle swarm adaptive clustering according to claim 3 is characterized in that: The edge set is constructed based on the following multi-source relationship: The spatial proximity between nodes is identified using the slope and aspect in terrain feature data; Using rainfall intensity or wind speed in meteorological data to identify temporal correlations between nodes; The regional flooding range in satellite remote sensing image data is used to identify the flood impact correlation between nodes; The water depth and flow information in the water level data of the integrated hydrological station are used to calculate the hydrodynamic coupling strength between nodes.

5. The flood path prediction method based on particle swarm adaptive clustering according to claim 1 is characterized in that: The S3 comprises the following steps: S31. Calculate the node feature matrix H based on the graph neural network model (L) , where L represents the total number of layers of the graph neural network and the node feature matrix H (L) It represents the importance and dynamic characteristics of each node in the flood propagation network under the comprehensive conditions of multi-source heterogeneous data; S32. Based on the node feature matrix H (L) For each node n in the target area i Calculate its node importance score I(n i ), which is used to evaluate the criticality of a node in the flood propagation process: in, Indicates that under the rth type of relationship, node n i The node n to which it is connected j The flood impact weight, Represents node n in layer L i The flood propagation characteristics of , R represents the number of multi-source relationship types, and u is the total number of nodes; S33. According to the node importance score I(n i ) in descending order, and select the K nodes with the highest scores to form the key node set N k ; S34. For the key node set N k For each node in, extract its node features The connection relationship in the multi-relation adjacency matrix is ​​used to identify the flood propagation path between key nodes by traversing the adjacent nodes, and generate a preliminary propagation path set: P={p1,p2,…,p M }; Among them, p M represents the Mth propagation path, including the node sequence and the flow or water level changes between nodes in the path; S35. For each path p in the preliminary propagation path set P j Calculate the path weight W(p j ), the path weight is used to quantify the priority of the path in flood propagation: Among them, I(n ki ) is the path p j The importance score of the node reflects the influence of the node in flood propagation, w ij is the edge e in the path ij The weight of represents the intensity of flood propagation between nodes; S36. According to the path weight W(p j ) Sort the preliminary propagation path set P, select the first M' paths with the highest weight to form the key propagation path set, the key path set P k Represents the most likely flood propagation path within the target area: P k ={p k1 ,p k2 ,…,p kM′ }。 6. The flood path prediction method based on particle swarm adaptive clustering according to claim 1 is characterized in that: The S4 comprises the following steps: S41. According to the key propagation path set P k Extract the node set N on each path k and the connection relationship set E k , where the node set N k Represents the flood propagation area nodes on the key propagation path, and the connection relationship set E k Represents the flood propagation association between nodes; S42. Based on the key propagation path set P k Calculate the node cluster center C in the target area = {c1, c2, ..., c q }, where q represents the number of sub-regions initially divided, and each cluster center c q Determined by: Among them, w j For node n j The importance weight reflects the criticality of the node in flood propagation, p j Represents node n j The spatial position vector of S43. Node set N in the target area k Divide the space and divide each node n j Assign to the nearest cluster center c i , the sub-region division is completed, and the sub-region set is represented as R: Among them, R i represents the i-th sub-region, including the cluster center c i All nodes closest to each other, ∥p j -c i ∥ represents node n j With cluster center c i The Euclidean distance of S44. Based on the sub-region set R, the connection relationship set E between the nodes is combined k For each sub-region R i The boundary range is further optimized and the boundary function B of each sub-region is calculated. i : Among them, B i represents the boundary of the ith subregion, including all edges e connected to other subregions jk .

7. The flood path prediction method based on particle swarm adaptive clustering according to claim 1 is characterized in that: The S5 comprises the following steps: S51. For each sub-region R i , set the particle swarm size P for each sub-region i , in sub-region R i Construct particle swarm collections separately: S52. In the sub-area R i In the search space determined by the flow velocity, flow direction and residence time parameters related to flood propagation, the parameter vector of the particle is defined Represents the initial position of the jth particle: in, Indicates that the jth particle is in the sub-region R i The initial flow velocity reflects the magnitude of the flood flow speed. Indicates the initial flow angle, representing the direction of flood propagation. It represents the initial detention time, reflecting the detention or stay characteristics of the flood in the sub-area; S53. Parameter vector for each particle Perform random initialization to make the particle swarm evenly distributed within the feasible range. The random initialization is: Among them, d∈{1,2,3} corresponds to the parameters of the three dimensions of flow velocity, flow direction and residence time, respectively. and They represent the minimum and maximum feasible values ​​of the dth parameter respectively, and rand(0,1) represents a random number in the interval [0,1]; S54. For each particle Set the initial velocity vector Let the initial velocity components of each dimension obey random or zero distribution, and initialize the particle swarm set G i Stored in sub-region R i in the particle swarm database; S55. Output each sub-region R i The initial position and initial velocity of the inner particle group are used to represent the flood propagation path parameters of the sub-region and complete the initialization of the particle group in each sub-region.

8. The flood path prediction method based on particle swarm adaptive clustering according to claim 1 is characterized in that: The S6 comprises the following steps: S61. In the initialized particle swarm set G i Based on the i Define the upper limit of the number of iterations and execute the particle swarm optimization process based on the adaptive swarming strategy; S62. At each iteration, calculate the particles The fitness value of The fitness value reflects the performance of the particle in predicting the accuracy and stability of the flood propagation path. The fitness function is defined as: Among them, D i Represents sub-region R i The local data within the sub-region, including topographic features, meteorological data, and water level information of hydrological stations, are used to measure the degree of fit of the current particle in the flood propagation process in the sub-region; S63. According to the fitness value Particle The individual optimal position and the global optimal position gbest in the sub-region i Update and save the position with the highest fitness value as the individual optimal position of the particle and the global optimal position of the sub-region; S64. Particles Speed With the position updated, the velocity is updated as: Among them, ω is the inertia weight, balancing the exploration and development capabilities, c1 and c2 are learning factors, rand1(0,1) and rand2(0,1) represent random numbers in the interval [0,1]; S65. During the iteration process, the local data changes in the sub-region are detected in real time. If the key monitoring indicators change significantly, the particle swarm size and algorithm parameters ω, c1, c2 are dynamically adjusted according to the adaptive clustering strategy, and the particles are The search direction and speed are modified, and some particle groups are reinitialized or merged to meet the dynamic propagation conditions of floods; S66. Repeat the iterative steps S61-S65 until the upper limit of the number of iterations is reached or the convergence condition is met, and output the sub-region R i The optimal position set of the inner particle swarm {gbest i }, which represents the set of particle solutions with the highest fitness in the flood propagation path prediction of this sub-region.

Images