A flood area prediction method and system based on rainfall spatiotemporal position perception

Abstract

The application discloses a flood area prediction method and system based on rainfall space-time position perception, and comprises the following steps: acquiring meteorological forcing data sets, static characteristic data sets of the flood prediction area, and rainfall characteristic time series obtained by multidimensional clustering of each watershed and the outlet section flow time series of the watershed; constructing a flood prediction area model of the flood prediction area; and performing collaborative training on the flood prediction area model, wherein a regional model collaborative loss function is used as a loss function. Compared with the prior art, the flood area prediction method based on rainfall space-time position perception can significantly improve the perception ability of the regional prediction model to the space-time differences of watershed rainfall, effectively enhance the accuracy and reliability of flood prediction, and also provide a result reference for flood prediction in unmeasured areas.

Description

Technical Field

[0001] This invention belongs to the field of hydrological forecasting technology, specifically relating to a flood area forecasting method and system based on the spatial and temporal location perception of rainfall. Background Technology

[0002] In recent years, influenced by both global climate change and frequent extreme weather events, the phenomenon of short-duration heavy rainfall triggering torrential floods has become increasingly frequent, seriously threatening people's lives and property and causing enormous impacts on urban infrastructure, agricultural production, and energy supply systems. Accurate flood forecasting for flood-prone areas is a crucial support for flood control and disaster reduction, reservoir scheduling, and water resource management, and is of great significance for improving the ability of flood-prone areas to cope with extreme hydrological events. In flood forecasting model construction, rainfall, as the most core driving factor, directly determines the dynamic evolution of runoff generation and confluence processes in flood-prone areas due to its spatiotemporal distribution characteristics. Traditional flood forecasting methods typically treat rainfall as a time series input or perform spatial averaging, neglecting the spatial differences in rainfall, its temporal similarities, and its spatial matching relationship with the topography and river network structure of flood-prone areas. This results in insufficient characterization of key information such as flood peak values, rise and fall processes, and response lags in areas sensitive to the response of torrential rainfall centers, seriously affecting the timeliness and accuracy of regional-scale flood forecasts. Therefore, effectively sensing the fine spatiotemporal location information of rainfall events has become a key breakthrough in improving the accuracy of flood forecasting for flood-prone areas.

[0003] With the development of artificial intelligence technology, deep learning has shown great potential in the field of hydrological forecasting, especially Long Short-Term Memory (LSTM) networks, which have been widely used in flood forecasting tasks due to their excellent modeling ability of long-term dependencies in time series. Research shows that compared with traditional physical mechanism hydrological models and other machine learning methods, LSTM excels in capturing the nonlinear dynamic characteristics of rainfall-flood flow processes and can effectively learn complex patterns in historical hydrological and meteorological sequences. However, lumped LSTM models are usually trained on a single watershed, making it difficult to fully consider the spatiotemporal correlation information between flood processes and dynamic forcing data and static feature datasets across multiple watersheds, thus limiting their application at the regional scale. To overcome this limitation, researchers have proposed extending LSTM to multiple watersheds for regional modeling, achieving better performance than lumped LSTM. However, these methods still generally use average rainfall as the model input and lack an explicit perception mechanism for watershed rainfall events. Existing models fail to fully consider the spatiotemporal correlation information between rainfall-sensitive areas and flood processes, leading to lags or distortions in the prediction of regional flood responses triggered by localized heavy rainfall. Therefore, it is urgent to construct a flood regional forecasting framework that can fully perceive the dynamic information of the spatiotemporal location of rainfall. While retaining the powerful temporal modeling capabilities of LSTM, it is necessary to enhance the model's perception and response capabilities to the spatial distribution patterns of rainfall, so as to achieve more accurate and reliable regional flood forecasts. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a flood area forecasting method and system based on rainfall spatiotemporal location perception, which solves the problems of insufficient perception of rainfall spatiotemporal distribution characteristics and difficulty in effectively integrating multi-source static attributes and dynamic hydrological response information of watersheds in existing flood forecasting models.

[0005] The technical solution adopted in this invention is as follows:

[0006] This invention provides a flood area forecasting method based on rainfall spatiotemporal location perception, comprising:

[0007] Based on the DEM elevation data of the flood forecast area, the flood forecast area is divided into multiple watersheds; each watershed has a unique outlet section.

[0008] Obtain multiple typical historical rainfall scenarios; under each typical historical rainfall scenario, obtain the meteorological forcing dataset, static feature dataset, and rainfall feature time series obtained by multidimensional clustering of each watershed, as well as the outlet cross-sectional flow time series of the watershed;

[0009] A flood forecasting regional model is constructed for the flood forecasting area. Under each typical historical rainfall scenario, sample data from all watersheds are simultaneously input. For each watershed, the sample data includes meteorological forcing datasets, static feature datasets, and rainfall feature time series after multidimensional clustering as model inputs. The outlet section flow time series corresponding to the watershed is used as the output target. The loss function adopts the regional model collaborative loss function constructed by considering the simulated flow of the outlet section of all watersheds. The flood forecasting regional model is collaboratively trained to obtain a trained flood forecasting regional model.

[0010] A pre-trained flood forecasting regional model is used to forecast floods in the flood forecasting region.

[0011] Furthermore, the meteorological forcing dataset includes raster precipitation, air temperature, evapotranspiration, and solar radiation; the static feature dataset includes watershed area, average elevation, frequency of heavy precipitation, clay proportion of soil layer, sandy soil proportion of soil layer, loam proportion of soil layer, artificial surface cover proportion, agricultural land proportion, forest and semi-natural area proportion, wetland proportion, water body proportion, annual average precipitation, precipitation seasonality index, snowfall proportion, frequency of light precipitation, duration of heavy precipitation and duration of light precipitation, as well as the calculated watershed daily precipitation standard deviation, daily average global radiation, minimum air temperature, maximum air temperature, and potential evapotranspiration index.

[0012] Furthermore, the rainfall feature time series obtained from the multidimensional clustering of each watershed were acquired using the following method:

[0013] Step A1, for the watershed , B represents the total number of watersheds in the flood forecast area. Rasterization is performed to obtain N raster cells; the typical historical rainfall scenario has T time steps.

[0014] Step A2: Under the typical historical rainfall scenario, obtain each grid cell. At each time step Rainfall observations , , According to grid units The watershed The degree of coverage affects rainfall observations. The correction was performed to obtain the corrected rainfall value. Thus, grid cells are obtained. Rainfall value time series ;

[0015] Step A3: Considering the time-delay response characteristics of rainfall and watershed outlet cross-section flow, calculate the value of each grid cell. Rainfall value time series With the watershed Time series of outlet cross-section flow Similarity features ;

[0016] Step A4, according to each grid cell With the watershed The distance between the exit sections and the grid cells The DEM digital elevation is used to calculate the value of each raster cell. Cumulative confluence distance characteristics ;

[0017] Step A5, construct the watershed Three-dimensional temporal feature matrix :

[0018]

[0019] Three-dimensional temporal feature matrix Each element in For the aforementioned watershed Three-dimensional temporal feature matrix Normalization is performed to obtain the normalized three-dimensional temporal feature matrix. Each of its elements ;

[0020] Step A6, Set the weight matrix , These represent the weights for rainfall values, similarity features, and cumulative confluence distance features, respectively. ,and ;

[0021] Step A7, for the normalized three-dimensional temporal feature matrix With weight matrix Perform the Hadamard product to obtain the clustering feature matrix. Among them, the clustering feature matrix Each element in is ;

[0022] Step A8: Use the TimeSeriesKMeans clustering method to cluster the feature matrix. Clustering the elements in the data yields the following results: A cluster;

[0023] For each cluster, calculate the rainfall value of each element within the cluster at the same time step. The mean value is used to obtain the rainfall characteristic time series for each cluster;

[0024] Step A9: Set different weight matrices γ, and repeat steps A6 to A8 to obtain the watershed. Multiple time series of rainfall characteristics that characterize different spatial properties of rainfall.

[0025] Furthermore, step A2 specifically involves:

[0026] Step A21, obtain the watershed Raster rainfall matrix :

[0027]

[0028] Step A22, according to the watershed Boundary vector data and each raster cell Spatial relationships, to obtain the watershed raster coverage matrix ;in, Representative watershed grid cells The coverage rate, if the watershed The boundary vector data completely covers the raster cells. Then the coverage The value is 1; if there is partial coverage, then the coverage rate is 1. =Grid unit Covered area / grid unit The area;

[0029] Step A23, calculate the raster rainfall matrix. With raster coverage matrix The Hadamard product yields the corrected rainfall matrix. :

[0030]

[0031] Rainfall value matrix Each element in the grid is a grid cell. At time step Corrected rainfall values .

[0032] Furthermore, step A3 specifically involves:

[0033] Step A31, the watershed Time series of outlet cross-section flow Simplified representation: Its time step The corresponding element is ; the basin Each grid cell Rainfall value time series Simplified representation: Its time step The corresponding element is ;in, ; ;

[0034] Step A32, calculate the watershed First-order monotonic sequence of outlet cross-sectional flow and the basin Each grid cell Rainfall values ​​first-order monotonic sequence The formula is as follows:

[0035]

[0036]

[0037] in: , These are boundary conditions;

[0038] Step A33, calculate the watershed Second-order monotonic sequence of outlet cross-sectional flow and the basin Each grid cell Second-order monotonic sequence of rainfall values The formula is as follows:

[0039]

[0040]

[0041] in: , , , , which are boundary conditions;

[0042] Step A34, according to the watershed The time series of outlet cross-sectional flow, the first-order monotonic series of outlet cross-sectional flow, and the second-order monotonic series of outlet cross-sectional flow, at each time step A three-dimensional feature vector of the outlet cross-sectional flow is constructed. This leads to the three-dimensional characteristic matrix of the outlet cross-sectional flow. ;

[0043] According to the watershed Each grid cell The time series of rainfall values, the first-order monotonic series of rainfall values, and the second-order monotonic series of rainfall values, at each time step , constructing grid cells Rainfall three-dimensional feature vector ; and then the grid cells are constructed. Rainfall three-dimensional feature matrix ;

[0044] The following formula is used to calculate the three-dimensional characteristic matrix of the outlet cross-sectional flow. any time step Corresponding element With the three-dimensional feature matrix of rainfall any time step Corresponding element Euclidean distance :

[0045] ;

[0046] in, The smaller the value, the higher the degree of similarity; ; ;

[0047] Step A35, Three-dimensional feature matrix based on outlet cross-sectional flow and grid cells Rainfall three-dimensional feature matrix Based on the similarity of each element, construct the cumulative distance matrix DT:

[0048]

[0049] in: Three-dimensional characteristic matrix representing the outlet cross-sectional flow any time step Corresponding element With the three-dimensional feature matrix of rainfall any time step Corresponding element The cumulative distance, , The calculation method is as follows:

[0050] Based on the range of the time-delay response between basin rainfall and outlet cross-section flow, a time step is set. and time step Time difference constraints between: ;in, The time-delay response threshold;

[0051] Elements in the cumulative distance matrix DT If its time step and time step Time difference constraint not met Then let it Invalid value;

[0052] For satisfying the time difference constraint Each element The assignment method is as follows:

[0053] Set the initial boundary conditions as follows:

[0054]

[0055]

[0056] when at the same time ,at the same time When, the following formula is used to determine Value:

[0057]

[0058] Step A36: Calculate the optimal path based on the cumulative distance matrix DT;

[0059] Specifically, the plan uses the element in the lower left corner. Starting from the top right element Optimal path to destination ,in , representing the k-th element in the optimal path. This represents the number of elements in the optimal path, and This optimal path satisfies the following constraints:

[0060] Boundary conditions: , ;

[0061] Continuity and monotonicity: If the previous step of the path ,Next step Must meet , This ensures that W moves monotonically over time and can only align with its adjacent elements;

[0062] Step A37: Calculate the grid cells based on the planned optimal path. Rainfall value time series Time series of flow rates to the outlet section Minimum cumulative distance DTW ( , ):

[0063]

[0064] Step A38, calculate grid cells Rainfall value time series With the watershed Time series of outlet cross-section flow Similarity features :

[0065]

[0066] Similarity features The higher the score, the higher the value of the grid cell. Rainfall value time series and watershed The more similar the time series of the outlet cross-section flow, the better the performance of the grid cell. Hydrological response characteristics.

[0067] Furthermore, step A4 specifically involves:

[0068] Step A41, the watershed Each grid cell in A grid cell has 8 adjacent cells. The cells directly above, below, to the left, and to the right of the grid cell are called orthogonal neighbors; the cells at the top left, top right, bottom left, and bottom right of the grid cell are called intersecting neighbors.

[0069] Step A42, the watershed The location of the outlet section is used to set the outlet section station of the watershed; grid unit The recursive form of the cumulative confluence distance characteristic is:

[0070]

[0071]

[0072] in: For grid cells The infinitesimal distance of the flow direction, Raster resolution; Represents grid cells The cumulative confluence distance characteristic of its downstream adjacent grid cells in the direction of flow;

[0073] Grid cell The direct form of the cumulative confluence distance characteristic is:

[0074]

[0075] in: Expression grid unit Each grid cell passed through on the path to the outlet section station o of the watershed , Indicates the individual grid cells that have passed through. The infinitesimal distance of the flow direction.

[0076] Furthermore, the TimeSeriesKMeans clustering method is used to analyze the clustering feature matrix. When clustering the elements in the dataset, Euclidean distance is used to measure the differences between the elements:

[0077]

[0078] in: Represents the clustering feature matrix elements in With elements The Euclidean distance.

[0079] Furthermore, the collaborative loss function of the regional model is:

[0080]

[0081]

[0082] Where: B represents the total number of watersheds in the flood forecast area. The outlet cross-section discharge time series of watershed b is shown below. ; The simulated flow time series is for the outlet section of watershed b. ; Let be the normalized standard deviation of the outlet flow of watershed b; Let be the loss coefficient for watershed b; .

[0083] Furthermore, when training the flood forecasting regional model, the Nash efficiency coefficient (NSE) for each watershed is used as an evaluation index for forecast accuracy.

[0084]

[0085] in: Let be the Nash efficiency coefficient of watershed b; For watershed b at time step The outlet cross-sectional flow rate; For watershed b at time step Simulated flow rate at the outlet section; Let be the average discharge at the outlet section of watershed b. .

[0086] The present invention also provides a flood area forecasting system based on rainfall spatiotemporal location perception, comprising:

[0087] The flood forecasting area division module is used to divide the flood forecasting area into multiple watersheds based on the DEM elevation data of the flood forecasting area; each watershed has a unique outlet section;

[0088] The typical historical rainfall scenario acquisition module is used to acquire multiple typical historical rainfall scenarios; under each typical historical rainfall scenario, it acquires the meteorological forcing dataset, static feature dataset, and rainfall feature time series obtained by multidimensional clustering of each watershed, as well as the outlet section flow time series of the watershed.

[0089] The flood forecasting regional model construction and training module is used to construct the flood forecasting regional model for the flood forecasting region. Under each typical historical rainfall scenario, sample data from all watersheds are input simultaneously. For each watershed, the sample data includes meteorological forcing datasets, static feature datasets, and rainfall feature time series after multidimensional clustering as model inputs, and the outlet section flow time series corresponding to the watershed as the output target. The loss function adopts the regional model collaborative loss function constructed by considering the simulated flow of the outlet section of all watersheds. The flood forecasting regional model is collaboratively trained to obtain the trained flood forecasting regional model.

[0090] The flood forecasting module is used to forecast floods in the flood forecasting area using a trained flood forecasting area model.

[0091] The flood area forecasting method and system based on rainfall spatiotemporal location sensing provided by this invention has the following advantages:

[0092] Compared with existing technologies, the regional flood forecasting method based on precipitation spatiotemporal location perception constructed in this invention can significantly improve the regional forecasting model's ability to perceive the spatiotemporal differences in watershed rainfall, effectively enhance the accuracy and reliability of flood forecasting, and also provide result references for flood forecasting in unmeasured areas. Attached Figure Description

[0093] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0094] Figure 1 A flowchart of a flood area forecasting method based on rainfall spatiotemporal location perception provided by the present invention;

[0095] Figure 2 This is a similarity analysis result diagram of the dynamic time warping method based on monotonicity constraints in this invention;

[0096] Figure 3 This is a diagram showing the spatiotemporal clustering results of high-dimensional raster rainfall sequences based on the TimeSeriesKMeans method in this invention.

[0097] Figure 4 This is a bar chart of the NSE (Non-Sequential Array) histogram of the regional LSTM prediction results in this invention. Detailed Implementation

[0098] To make the technical problems solved, the technical solutions, and the beneficial effects of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and are not intended to limit the invention.

[0099] This invention relates to a flood regional forecasting method and system based on rainfall spatiotemporal location awareness. Targeting the dynamic response characteristics of rainfall-outlet cross-section discharge, a dynamic time warping method with monotonicity constraints (m-DTW) is proposed. This method effectively identifies the similarity features between rainfall sequences in different raster units and the outlet cross-section discharge response. It also combines cumulative confluence distance to characterize spatial confluence effects and employs the TimeSeriesKMeans method for spatiotemporal clustering analysis of high-dimensional raster rainfall sequences. Simultaneously, it fully leverages the advantages of LSTM networks in long-term time series modeling, integrating meteorological forcing data, static attributes, and the spatiotemporal pattern of rainfall to achieve accurate perception of watershed rainfall spatiotemporal information and detailed characterization of complex nonlinear hydrological processes. This method not only improves the model's ability to perceive dynamic spatiotemporal information of rainfall but also enhances the accuracy and robustness of regional-scale flood discharge forecasts. The overall design concept is clear, the implementation path is well-defined, and it possesses good scalability and practical application value.

[0100] This invention, while identifying the spatiotemporal characteristics of rainfall in a watershed, incorporates rainfall variation information and runoff characteristic information, which plays a significant role in promoting the development of spatiotemporal rainfall data mining and regional forecasting models. Figure 1 As shown, the specific steps include:

[0101] Step S1: Based on the DEM elevation data of the flood forecast area, the flood forecast area is divided into multiple watersheds; each watershed has a unique outlet section.

[0102] Step S2: Obtain multiple typical historical rainfall scenarios; under each typical historical rainfall scenario, obtain the meteorological forcing dataset, static feature dataset, and rainfall feature time series obtained by multidimensional clustering of each watershed, as well as the outlet section flow time series of the watershed.

[0103] The meteorological forcing dataset includes rasterized precipitation, temperature, evapotranspiration, and solar radiation; the static feature dataset includes watershed area, average elevation, frequency of heavy precipitation, clay proportion of soil layer, sandy soil proportion of soil layer, loam proportion of soil layer, artificial surface cover proportion, agricultural land proportion, forest and semi-natural area proportion, wetland proportion, water body proportion, annual average precipitation, precipitation seasonality index, snowfall proportion, frequency of light precipitation, duration of heavy precipitation and duration of light precipitation, as well as the watershed daily precipitation standard deviation, daily average global radiation, minimum temperature, maximum temperature, and potential evapotranspiration index calculated based on the rasterized precipitation, temperature, and evapotranspiration meteorological forcing dataset.

[0104] Step S3: Construct the flood forecasting area model for the flood forecasting area; under each typical historical rainfall scenario, simultaneously input sample data from all watersheds. For each watershed, the sample data includes meteorological forcing dataset, static feature dataset, and rainfall feature time series after multidimensional clustering as model input, and the outlet section flow time series corresponding to the watershed as output target. The loss function adopts the regional model collaborative loss function constructed by considering the simulated flow of the outlet section of all watersheds. The flood forecasting area model is collaboratively trained to obtain the trained flood forecasting area model.

[0105] Step S4: Use the trained flood forecasting area model to forecast floods in the flood forecasting area.

[0106] In step S2 of this invention, a unique algorithm model is used to obtain the rainfall characteristic time series obtained by multidimensional clustering of each watershed. The specific method is as follows:

[0107] Step A1, for the watershed , B represents the total number of watersheds in the flood forecast area. Rasterization is performed to obtain N raster cells; the typical historical rainfall scenario has T time steps.

[0108] Step A2: Under the typical historical rainfall scenario, obtain each grid cell. At each time step Rainfall observations , , According to grid units The watershed The degree of coverage affects rainfall observations. The correction was performed to obtain the corrected rainfall value. Thus, grid cells are obtained. Rainfall value time series ;

[0109] This step is specifically as follows:

[0110] Step A21, obtain the watershed Raster rainfall matrix :

[0111]

[0112] Step A22, according to the watershed Boundary vector data and each raster cell Spatial relationships, to obtain the watershed raster coverage matrix ;in, Representative watershed grid cells The coverage rate, if the watershed The boundary vector data completely covers the raster cells. Then the coverage The value is 1; if there is partial coverage, then the coverage rate is 1. =Grid unit Covered area / grid unit The area;

[0113] Step A23, calculate the raster rainfall matrix. With raster coverage matrix The Hadamard product yields the corrected rainfall matrix. :

[0114]

[0115] Rainfall value matrix Each element in the grid is a grid cell. At time step Corrected rainfall values .

[0116] Step A3: Considering the time-delay response characteristics between rainfall and watershed outlet flow, and the mismatch between the peak values ​​of rainfall and outlet flow, this invention proposes a dynamic time warping (m-DTW) calculation method that considers the monotonic constraint of rainfall-outlet flow, calculating the dynamic time warping (m-DTW) for each grid cell. Rainfall value time series With the watershed Time series of outlet cross-section flow Similarity features ;

[0117] This step is specifically as follows:

[0118] Step A31, the watershed Time series of outlet cross-section flow Simplified representation: Its time step The corresponding element is ; the basin Each grid cell Rainfall value time series Simplified representation: Its time step The corresponding element is ;in, ; ;

[0119] Step A32, calculate the watershed First-order monotonic sequence of outlet cross-sectional flow and the basin Each grid cell Rainfall values ​​first-order monotonic sequence The formula is as follows:

[0120]

[0121]

[0122] in: , These are boundary conditions;

[0123] Step A33, calculate the watershed Second-order monotonic sequence of outlet cross-sectional flow and the basin Each grid cell Second-order monotonic sequence of rainfall values The formula is as follows:

[0124]

[0125]

[0126] in: , , , , which are boundary conditions;

[0127] Step A34, according to the watershed The time series of outlet cross-sectional flow, the first-order monotonic series of outlet cross-sectional flow, and the second-order monotonic series of outlet cross-sectional flow, at each time step A three-dimensional feature vector of the outlet cross-sectional flow is constructed. This leads to the three-dimensional characteristic matrix of the outlet cross-sectional flow. ;

[0128] According to the watershed Each grid cell The time series of rainfall values, the first-order monotonic series of rainfall values, and the second-order monotonic series of rainfall values, at each time step , constructing grid cells Rainfall three-dimensional feature vector ; and then the grid cells are constructed. Rainfall three-dimensional feature matrix ;

[0129] The following formula is used to calculate the three-dimensional characteristic matrix of the outlet cross-sectional flow. any time step Corresponding element With the three-dimensional feature matrix of rainfall any time step Corresponding element Euclidean distance :

[0130] ;

[0131] in, The smaller the value, the higher the degree of similarity; ; ;

[0132] Step A35, Three-dimensional feature matrix based on outlet cross-sectional flow and grid cells Rainfall three-dimensional feature matrix Based on the similarity of each element, construct the cumulative distance matrix DT:

[0133]

[0134] in: Three-dimensional characteristic matrix representing the outlet cross-sectional flow any time step Corresponding element With the three-dimensional feature matrix of rainfall any time step Corresponding element The cumulative distance, , The calculation method is as follows:

[0135] Based on the range of the time-delay response between basin rainfall and outlet cross-section flow, a time step is set. and time step Time difference constraints between: ;in, For example, given the short time lag between rainfall and outlet flow in the watershed, a reasonable constraint on the time difference is set as follows: .

[0136] Elements in the cumulative distance matrix DT If its time step and time step Time difference constraint not met Then let it Invalid value;

[0137] For satisfying the time difference constraint Each element The assignment method is as follows:

[0138] Set the initial boundary conditions as follows:

[0139]

[0140]

[0141] when at the same time ,at the same time When, the following recursive formula is used to determine Value:

[0142]

[0143] Step A36: Calculate the optimal path based on the cumulative distance matrix DT;

[0144] Specifically, the plan uses the element in the lower left corner. Starting from the top right element Optimal path to destination ,in , representing the k-th element in the optimal path. This represents the number of elements in the optimal path, and This optimal path satisfies the following constraints:

[0145] Boundary conditions: , That is, the selected path must start from the bottom left corner and end at the top right corner;

[0146] Continuity and monotonicity: If the previous step of the path ,Next step Must meet , This ensures that W moves monotonically over time and can only align with its adjacent elements;

[0147] Step A37: Calculate the grid cells based on the planned optimal path. Rainfall value time series Time series of flow rates to the outlet section Minimum cumulative distance DTW ( , ):

[0148]

[0149] Step A38, calculate grid cells Rainfall value time series With the watershed Time series of outlet cross-section flow Similarity features :

[0150]

[0151] Similarity features The higher the score, the higher the value of the grid cell. Rainfall value time series and watershed The more similar the time series of the outlet cross-section flow, the better the performance of the grid cell. Hydrological response characteristics.

[0152] Step A4, according to each grid cell With the watershed The distance between the exit sections and the grid cells The DEM digital elevation is used to calculate the value of each raster cell. Cumulative confluence distance characteristics ;

[0153] This step is specifically as follows:

[0154] Step A41, the watershed Each grid cell in A grid cell has 8 adjacent cells. The cells directly above, below, to the left, and to the right of the grid cell are called orthogonal neighbors; the cells at the top left, top right, bottom left, and bottom right of the grid cell are called intersecting neighbors.

[0155] Step A42, the watershed The location of the outlet section is set as the outlet section station of the watershed; based on the DEM digital elevation of the watershed, raster cells are used. The recursive form of the cumulative confluence distance characteristic is:

[0156]

[0157]

[0158] in: For grid cells The infinitesimal distance of the flow direction, Raster resolution; Represents grid cells The cumulative confluence distance characteristic of its downstream adjacent grid cells in the direction of flow;

[0159] Grid cell The direct form of the cumulative confluence distance characteristic is:

[0160]

[0161] in: Expression grid unit Each grid cell passed through on the path to the outlet section station o of the watershed , Indicates the individual grid cells that have passed through. The infinitesimal distance of the flow direction.

[0162] Step A5, construct the watershed Three-dimensional temporal feature matrix :

[0163]

[0164] Three-dimensional temporal feature matrix Each element in For the aforementioned watershed Three-dimensional temporal feature matrix Normalization is performed to eliminate the influence of dimensions, resulting in a normalized three-dimensional time series feature matrix. Each of its elements ;

[0165] Step A6, Set the weight matrix , These represent the weights for rainfall values, similarity features, and cumulative confluence distance features, respectively. ,and ;

[0166] Step A7, for the normalized three-dimensional temporal feature matrix With weight matrix Perform the Hadamard product to obtain the clustering feature matrix. Among them, the clustering feature matrix Each element in is ;

[0167] Step A8: Use the TimeSeriesKMeans clustering method to cluster the feature matrix. Clustering the elements in the data yields the following results: A cluster;

[0168] For each cluster, calculate the rainfall value of each element within the cluster at the same time step. The mean value is used to obtain the rainfall characteristic time series for each cluster;

[0169] In this step, the TimeSeriesKMeans clustering method is used to cluster the feature matrix. When clustering the elements in the dataset, the core idea of ​​this clustering method is to minimize the intra-cluster distance and maximize the inter-cluster difference. Euclidean distance is typically used to measure the difference between samples, as shown in the following formula:

[0170]

[0171] in: Represents the clustering feature matrix elements in With elements The Euclidean distance.

[0172] Step A9: Set different weight matrices γ, and repeat steps A6 to A8 to obtain the watershed. Multiple time series of rainfall characteristics that characterize different spatial properties of rainfall.

[0173] Specifically, various configurations of different weight matrices γ are set to adjust the contribution ratios of rainfall value, similarity, and confluence distance in clustering. The TimeSeriesKMeans algorithm is used to cluster rainfall grids under three strategies: rainfall intensity-dominated, similarity-dominated, and confluence distance-dominated. For each strategy, the average rainfall time series within each cluster region is calculated, and key feature sequences characterizing different spatial characteristics of rainfall are extracted, such as features of high-value areas, similarity areas, and rapid confluence areas. These are then used to construct a multi-weighted rainfall spatial feature sequence for training regional flood forecasting models.

[0174] The flood forecasting regional model constructed in this invention is a multi-source feature fusion flood forecasting regional model, which can be called a spatiotemporal clustering regional LSTM model. It deeply fuses meteorological forcing data and static features from the entire watershed with extracted multi-dimensional clustered rainfall feature sequences as input features for the regional LSTM model. Using historical measured flood data as the output target, a dataset including training, validation, and test sets is constructed. During the model training phase, a regional model collaborative loss function is introduced. After standardizing the input features and target variables, the final regional LSTM model is obtained through training. This model can effectively capture the spatiotemporal distribution differences of watershed rainfall and the nonlinear characteristics of hydrological response. In this invention, the regional model collaborative loss function is:

[0175]

[0176]

[0177] Where: B represents the total number of watersheds in the flood forecast area. The outlet cross-section discharge time series of watershed b is shown below. ; The simulated flow time series is for the outlet section of watershed b. ; Let be the normalized standard deviation of the outlet flow of watershed b; Let be the loss coefficient for watershed b; .

[0178] In this invention, when training the flood forecasting regional model, the Nash efficiency coefficient (NSE) corresponding to each watershed is used as an evaluation index for forecast accuracy.

[0179]

[0180] in: Let be the Nash efficiency coefficient of watershed b; For watershed b at time step The outlet cross-sectional flow rate; For watershed b at time step Simulated flow rate at the outlet section; Let be the average discharge at the outlet section of watershed b. .

[0181] The present invention also provides a flood area forecasting system based on rainfall spatiotemporal location perception, comprising:

[0182] The flood forecasting area division module is used to divide the flood forecasting area into multiple watersheds based on the DEM elevation data of the flood forecasting area; each watershed has a unique outlet section;

[0183] The typical historical rainfall scenario acquisition module is used to acquire multiple typical historical rainfall scenarios; under each typical historical rainfall scenario, it acquires the meteorological forcing dataset, static feature dataset, and rainfall feature time series obtained by multidimensional clustering of each watershed, as well as the outlet section flow time series of the watershed.

[0184] The flood forecasting regional model construction and training module is used to construct the flood forecasting regional model for the flood forecasting region. Under each typical historical rainfall scenario, sample data from all watersheds are input simultaneously. For each watershed, the sample data includes meteorological forcing datasets, static feature datasets, and rainfall feature time series after multidimensional clustering as model inputs, and the outlet section flow time series corresponding to the watershed as the output target. The loss function adopts the regional model collaborative loss function constructed by considering the simulated flow of the outlet section of all watersheds. The flood forecasting regional model is collaboratively trained to obtain the trained flood forecasting regional model.

[0185] The flood forecasting module is used to forecast floods in the flood forecasting area using a trained flood forecasting area model.

[0186] In summary, this invention proposes a flood regional forecasting method and system based on rainfall spatiotemporal location awareness. By incorporating monotonic constrained dynamic time warping with prior knowledge of rainfall-outlet section flow, it effectively identifies and quantifies key temporal features between each grid rainfall sequence and the watershed outlet section flow response. This significantly enhances the model's ability to perceive time lags, intensity matching, and response consistency in the rainfall-driven mechanism. Furthermore, TimeSeriesKMeans spatiotemporal clustering analysis is used to achieve accurate perception of rainfall spatiotemporal characteristics. Simultaneously, a regional LSTM model with strong generalization performance is constructed based on watershed spatiotemporal rainfall classification, meteorological forcing data, and static feature data. Using the watershed global average loss function as the loss function, the weights and threshold parameters of the regional LSTM model are updated, achieving accurate flood prediction at the regional scale. In addition to identifying rainfall spatiotemporal features, rainfall differences and confluence information are introduced, providing strong technical support for regionalized and intelligent flood early warning and water resource management. The trained regional LSTM model can also provide technical support for flood prediction in unmeasured areas. This invention has a rigorous design and is easy to implement, and is expected to play an important role in water resource management, flood control, and disaster reduction.

[0187] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings. However, the present invention can be implemented in many different ways as defined and covered by the claims. In this embodiment, a flood area forecasting method and system based on rainfall spatiotemporal location perception is provided to solve the problems of insufficient perception capability of existing flood forecasting models for rainfall spatiotemporal distribution characteristics and difficulty in effectively integrating multi-source static attributes and dynamic hydrological response information of watersheds.

[0188] This study focuses on the CAMELS-DE dataset from Germany, which covers 1582 watersheds within Germany and provides long-term, high-precision diurnal hydrological and meteorological observation data. The dataset includes measured flood discharge series, meteorological forcing datasets (rasterized rainfall, temperature, solar radiation), and static feature data (geographic, hydrological, land cover, soil, and human activities), with watershed areas ranging from 5 to 14004 km². 2 The data spans from 1951 to 2020, providing good spatial representativeness and temporal continuity. The study uses the period from October 1, 1970 to December 31, 1999 as the training period, October 1, 1965 to September 30, 1970 as the validation period, and January 1, 2000 to December 31, 2020 as the testing period. The study constructs a lumped LSTM model and a regional LSTM model for comparative analysis: the former is validated by constructing a lumped LSTM model to demonstrate its modeling performance in a single watershed, while the latter is validated by constructing a regional LSTM model to demonstrate its overall predictive ability at a regional scale. The mean and median NSE values ​​are used as indicators to evaluate the overall model performance, comprehensively reflecting the model's average performance and robustness in large-scale watershed groups.

[0189] Table 1 shows the forecast evaluation indicators of the original lumped LSTM model and the regional LSTM model constructed in this invention for the entire watershed and a single watershed:

[0190] Table 1:

[0191]

[0192] As can be seen from Table 1, the prediction accuracy of the regional LSTM model constructed in this invention is higher than that of the original lumped LSTM model.

[0193] like Figure 2 This is a similarity analysis result diagram of the dynamic time warping method based on monotonicity constraints in this invention; as shown... Figure 3 This is a high-dimensional raster rainfall sequence spatiotemporal clustering result diagram based on the TimeSeriesKMeans method in this invention; as shown... Figure 4 This is a histogram of NSE (Network Estimates and Optimizations) results from the regional LSTM prediction in this invention; Figure 4 In this paper, the spatiotemporal clustering region LSTM is the region LSTM model of this invention; the lumped LSTM is a model in the prior art; thus, it is proved that the prediction accuracy of the region LSTM model of this invention is higher than that of the original lumped LSTM model.

[0194] This invention discloses a flood regional forecasting method and system based on rainfall spatiotemporal location awareness. By extracting spatiotemporal features from a watershed's raster rainfall dataset, a dynamic time warping (m-DTW) calculation method considering the monotonic constraint of rainfall-flood flow is proposed to quantify the temporal matching degree between each raster rainfall sequence and the flow at the watershed outlet. The three-dimensional weighted information of raster rainfall sequences, similarity features, and cumulative confluence distance features is used as feature input for the TimeSeriesKMeans clustering method to obtain spatiotemporal rainfall sequence clusters and identify key rainfall sub-regions with similar hydrological response characteristics. Furthermore, a regional LSTM is constructed and trained based on the rainfall spatiotemporal awareness results, hydro-meteorological dynamic forcing dataset, and static dataset to ultimately achieve accurate prediction of the watershed outlet flow process. Compared with existing technologies, the regional flood forecasting method based on rainfall spatiotemporal location awareness constructed in this invention can significantly improve the regional forecasting model's ability to perceive spatiotemporal differences in watershed rainfall, effectively enhancing the accuracy and reliability of flood forecasts, and also providing result references for flood forecasting in unmeasured areas.

[0195] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A flood area prediction method based on rainfall spatiotemporal position awareness, characterized in that, include: Based on the DEM elevation data of the flood forecast area, the flood forecast area is divided into multiple watersheds; Each watershed has a unique outlet section; Obtain multiple typical historical rainfall scenarios; under each typical historical rainfall scenario, obtain the meteorological forcing dataset, static feature dataset, and rainfall feature time series obtained by multidimensional clustering of each watershed, as well as the outlet cross-sectional flow time series of the watershed; A flood forecasting regional model is constructed for the flood forecasting area. For each typical historical rainfall scenario, sample data from all watersheds are simultaneously input. For each watershed, the sample data includes meteorological forcing datasets, static feature datasets, and rainfall feature time series after multidimensional clustering as model inputs. The outlet section flow time series corresponding to the watershed is used as the output target. The loss function is the regional model collaborative loss function constructed considering the simulated outlet section flow of all watersheds. The flood forecasting regional model is then collaboratively trained to obtain a trained flood forecasting regional model. The regional model collaborative loss function is: ； ； Where: B represents the total number of watersheds in the flood forecast area. The outlet cross-section discharge time series of watershed b is shown below. ; The simulated flow time series is for the outlet section of watershed b. ; Let be the normalized standard deviation of the outlet flow of watershed b; Let be the loss coefficient for watershed b; ; A pre-trained flood forecasting area model is used to forecast floods in the aforementioned flood forecasting area. The rainfall feature time series obtained from multidimensional clustering of each watershed were acquired using the following method: Step A1, for the watershed , B represents the total number of watersheds in the flood forecast area. Rasterization is performed to obtain N raster cells; the typical historical rainfall scenario has T time steps. Step A2: Under the typical historical rainfall scenario, obtain each grid cell. At each time step Rainfall observations , , According to grid units The watershed The degree of coverage affects rainfall observations. The correction was performed to obtain the corrected rainfall value. Thus, grid cells are obtained. Rainfall value time series ; Step A3: Considering the time-delay response characteristics of rainfall and watershed outlet cross-section flow, calculate the value of each grid cell. Rainfall value time series With the watershed Time series of outlet cross-section flow Similarity features ; Step A4, according to each grid cell With the watershed The distance between the exit sections and the grid cells The DEM digital elevation is used to calculate the value of each raster cell. Cumulative confluence distance characteristics ; Step A5, construct the watershed Three-dimensional temporal feature matrix : ； Three-dimensional temporal feature matrix Each element in For the aforementioned watershed Three-dimensional temporal feature matrix Normalization is performed to obtain the normalized three-dimensional temporal feature matrix. Each of its elements ; Step A6, Set the weight matrix , These represent the weights for rainfall values, similarity features, and cumulative confluence distance features, respectively. ,and ; Step A7, for the normalized three-dimensional temporal feature matrix With weight matrix Perform the Hadamard product to obtain the clustering feature matrix. Among them, the clustering feature matrix Each element in is ; Step A8: Use the TimeSeriesKMeans clustering method to cluster the feature matrix. Clustering the elements in the data yields the following results: A cluster; For each cluster, calculate the rainfall value of each element within the cluster at the same time step. The mean value is used to obtain the rainfall characteristic time series for each cluster; Step A9: Set different weight matrices γ, and repeat steps A6 to A8 to obtain the watershed. Multiple time series of rainfall characteristics that characterize different spatial properties of rainfall.

2. The flood area forecasting method based on rainfall spatiotemporal location perception according to claim 1, characterized in that, The meteorological forcing dataset includes rasterized precipitation, air temperature, evapotranspiration, and solar radiation; the static feature dataset includes watershed area, average elevation, frequency of heavy precipitation, proportion of clay soil, proportion of sandy soil, proportion of loam soil, proportion of artificial land cover, proportion of agricultural land, proportion of forest and semi-natural areas, proportion of wetlands, proportion of water bodies, annual average precipitation, precipitation seasonality index, proportion of snowfall, frequency of light precipitation, duration of heavy precipitation and duration of light precipitation, as well as the calculated standard deviation of daily precipitation in the watershed, daily average global radiation, minimum air temperature, maximum air temperature, and potential evapotranspiration index.

3. The flood area forecasting method based on rainfall spatiotemporal location perception according to claim 1, characterized in that, Step A2 is as follows: Step A21, obtain the watershed Raster rainfall matrix : ； Step A22, according to the watershed Boundary vector data and each raster cell Spatial relationships, to obtain the watershed raster coverage matrix ;in, Representative watershed grid cells The coverage rate, if the watershed The boundary vector data completely covers the raster cells. Then the coverage The value is 1; if there is partial coverage, then the coverage rate is 1. =Grid unit Covered area / grid unit The area; Step A23, calculate the raster rainfall matrix. With raster coverage matrix The Hadamard product yields the corrected rainfall matrix. : ； Rainfall value matrix Each element in the grid is a grid cell. At time step Corrected rainfall values .

4. The flood area forecasting method based on rainfall spatiotemporal location perception according to claim 1, characterized in that, Step A3 is as follows: Step A31, the watershed Time series of outlet cross-section flow Simplified representation: Its time step The corresponding element is ; the basin Each grid cell Rainfall value time series Simplified representation: Its time step The corresponding element is ;in, ; ; Step A32, calculate the watershed First-order monotonic sequence of outlet cross-sectional flow and the basin Each grid cell Rainfall values ​​first-order monotonic sequence The formula is as follows: ； ； in: , These are boundary conditions; Step A33, calculate the watershed Second-order monotonic sequence of outlet cross-sectional flow and the basin Each grid cell Second-order monotonic sequence of rainfall values The formula is as follows: ； ； in: , , , , which are boundary conditions; Step A34, according to the watershed The time series of outlet cross-sectional flow, the first-order monotonic series of outlet cross-sectional flow, and the second-order monotonic series of outlet cross-sectional flow, at each time step A three-dimensional feature vector of the outlet cross-sectional flow is constructed. This leads to the three-dimensional characteristic matrix of the outlet cross-sectional flow. ; According to the watershed Each grid cell The time series of rainfall values, the first-order monotonic series of rainfall values, and the second-order monotonic series of rainfall values, at each time step , constructing grid cells Rainfall three-dimensional feature vector ; and then the grid cells are constructed. Rainfall three-dimensional feature matrix ; The following formula is used to calculate the three-dimensional characteristic matrix of the outlet cross-sectional flow. any time step Corresponding element With the three-dimensional feature matrix of rainfall any time step Corresponding element Euclidean distance : ； in, The smaller the value, the higher the degree of similarity; ; ; Step A35, Three-dimensional feature matrix based on outlet cross-sectional flow and grid cells Rainfall three-dimensional feature matrix Based on the similarity of each element, construct the cumulative distance matrix DT: ； in: Three-dimensional characteristic matrix representing the outlet cross-sectional flow any time step Corresponding element With the three-dimensional feature matrix of rainfall any time step Corresponding element The cumulative distance, , The calculation method is as follows: Based on the range of the time-delay response between basin rainfall and outlet cross-section flow, a time step is set. and time step Time difference constraints between: ;in, The time-delay response threshold; Elements in the cumulative distance matrix DT If its time step and time step Time difference constraint not met Then let it Invalid value; For satisfying the time difference constraint Each element The assignment method is as follows: Set the initial boundary conditions as follows: ； ； when at the same time ,at the same time When, the following formula is used to determine Value: ； Step A36: Calculate the optimal path based on the cumulative distance matrix DT; Specifically, the plan uses the element in the lower left corner. Starting from the top right element Optimal path to destination ,in , representing the k-th element in the optimal path. This represents the number of elements in the optimal path, and This optimal path satisfies the following constraints: Boundary conditions: , ; Continuity and monotonicity: If the previous step of the path ,Next step Must meet , This ensures that W moves monotonically over time and can only align with its adjacent elements; Step A37: Calculate the grid cells based on the planned optimal path. Rainfall value time series Time series of flow rates to the outlet section Minimum cumulative distance DTW ( , ): ； Step A38, calculate grid cells Rainfall value time series With the watershed Time series of outlet cross-section flow Similarity features : ； Similarity features The higher the score, the higher the value of the grid cell. Rainfall value time series and watershed The more similar the time series of the outlet cross-section flow, the better the performance of the grid cell. Hydrological response characteristics.

5. A flood area forecasting method based on rainfall spatiotemporal location perception according to claim 1, characterized in that, Step A4 is as follows: Step A41, the watershed Each grid cell in A grid cell has 8 adjacent cells. The cells directly above, below, to the left, and to the right of the grid cell are called orthogonal neighbors; the cells at the top left, top right, bottom left, and bottom right of the grid cell are called intersecting neighbors. Step A42, the watershed The location of the outlet section is used to set the outlet section station of the watershed; grid unit The recursive form of the cumulative confluence distance characteristic is: ； ； in: For grid cells The infinitesimal distance of the flow direction, Raster resolution; Represents grid cells The cumulative confluence distance characteristic of its downstream adjacent grid cells in the direction of flow; Grid cell The direct form of the cumulative confluence distance characteristic is: ； in: Expression grid unit Each grid cell passed through on the path to the outlet section station o of the watershed , Indicates the individual grid cells that have passed through. The infinitesimal distance of the flow direction.

6. The flood area forecasting method based on rainfall spatiotemporal location perception according to claim 1, characterized in that, The TimeSeriesKMeans clustering method was used to analyze the clustering feature matrix. When clustering the elements in the dataset, Euclidean distance is used to measure the differences between the elements: ； in: Represents the clustering feature matrix elements in With elements The Euclidean distance.

7. A flood area forecasting method based on spatiotemporal location perception of rainfall according to claim 1, characterized in that, When training the flood forecasting regional model, the Nash efficiency coefficient (NSE) for each watershed is used as an evaluation index for forecast accuracy. ； in: Let be the Nash efficiency coefficient of watershed b; For watershed b at time step The outlet cross-sectional flow rate; For watershed b at time step Simulated flow rate at the outlet section; Let be the average discharge at the outlet section of watershed b. .

8. A flood area forecasting system based on rainfall spatiotemporal location awareness for implementing the flood area forecasting method based on rainfall spatiotemporal location awareness as described in any one of claims 1-7, characterized in that, include: The flood forecasting area division module is used to divide the flood forecasting area into multiple watersheds based on the DEM elevation data of the flood forecasting area; Each watershed has a unique outlet section; The typical historical rainfall scenario acquisition module is used to acquire multiple typical historical rainfall scenarios; under each typical historical rainfall scenario, it acquires the meteorological forcing dataset, static feature dataset, and rainfall feature time series obtained by multidimensional clustering of each watershed, as well as the outlet section flow time series of the watershed. The flood forecasting regional model construction and training module is used to construct the flood forecasting regional model for the flood forecasting region. Under each typical historical rainfall scenario, sample data from all watersheds are input simultaneously. For each watershed, the sample data includes meteorological forcing datasets, static feature datasets, and rainfall feature time series after multidimensional clustering as model inputs, and the outlet section flow time series corresponding to the watershed as the output target. The loss function adopts the regional model collaborative loss function constructed by considering the simulated flow of the outlet section of all watersheds. The flood forecasting regional model is collaboratively trained to obtain the trained flood forecasting regional model. The flood forecasting module is used to forecast floods in the flood forecasting area using a trained flood forecasting area model.

Images