Accumulated water depth prediction method fusing spatial-temporal characteristics
By establishing a node relationship matrix with weights and using long and short-term memory networks and other technical means, the timing and spatial characteristics are extracted, and the shortcomings of the water accumulation prediction method in the existing technology are solved in capturing time and spatial characteristics, and the depth of water accumulation is improved Accuracy of prediction.
Patent Information
- Application Number
- CN202410979070.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-22
- Publication Date
- 2024-11-22
AI Technical Summary
The existing water accumulation prediction methods have limited ability to capture time and space characteristics, and cannot fully consider the correlation between the two, resulting in limited accuracy of the prediction results.
A water accumulation depth prediction method that integrates space-time characteristics is adopted. By obtaining the water accumulation depth time series data and precipitation time series data of the water accumulation monitoring station, a node relationship matrix with weight is established, and a water accumulation depth prediction model is constructed, using long Technical means such as short-term memory networks and Chebishev polynomials extract timing and spatial features and perform feature fusion to improve prediction accuracy.
Effectively combining the characteristics of water accumulation and precipitation enriches the input information of the prediction model, improves the accuracy of water accumulation prediction, and can capture the temporal and spatial characteristics of the change in the depth of water accumulation.
Smart Images

Figure CN119005490A_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the technical field of meteorological management, and in particular relates to a method for predicting water depth by integrating temporal and spatial characteristics. Background Art
[0002] With global warming and accelerated urbanization, the frequency and intensity of extreme precipitation events have increased significantly. Therefore, in order to reduce the harm of urban waterlogging to urban construction, environmental safety and social stability, it is of great significance to predict the depth of urban waterlogging. Existing waterlogging prediction methods mainly include empirical models and statistical methods, which have limited ability to consider temporal and spatial characteristics in waterlogging prediction. They are often based on historical data and statistical laws and cannot capture complex temporal and spatial characteristics, resulting in limited accuracy of prediction results. In addition, when extracting temporal and spatial features, waterlogging prediction methods based on machine learning and traditional convolutional neural networks often process them separately and cannot fully consider the correlation between the two. Summary of the invention
[0003] Purpose of the invention: In order to solve the problems existing in the above-mentioned prior art, the present invention provides a method for predicting water depth by integrating temporal and spatial characteristics.
[0004] Technical solution: The present invention discloses a method for predicting water depth by integrating temporal and spatial features, which specifically includes the following steps:
[0005] Step 1: Obtain the water depth time series data R of each waterlogging monitoring station i , precipitation time series data H i , the longitude and latitude of the waterlogging monitoring station and the distance between the ith waterlogging monitoring station and other waterlogging monitoring stations; i = 1, 2, ..., I; I represents the total number of waterlogging monitoring stations;
[0006] Step 2: To ensure that the waterlogging is caused by rainfall, the original waterlogging depth time series data is preprocessed based on the sliding window method;
[0007] Step 3: Establish an N×N-dimensional node relationship matrix A' with weights based on the relationship between waterlogging monitoring stations; form a feature matrix X with the preprocessed waterlogging depth time series data and the corresponding precipitation time series data. The dimension of the feature matrix X is [B, N, T, F], where B is the number of samples, T is the size of the sliding window, and F is the number of features;
[0008] Step 4: Construct a water depth prediction model, use the feature matrix X and the node relationship matrix A' as the water depth prediction model, and train the water depth prediction model;
[0009] Step 5: Use the trained waterlogging depth prediction model to predict the waterlogging depth.
[0010] Furthermore, the pre-processing in step 2 includes the following steps:
[0011] (1) If there is no rainfall in the current period t, there is no water accumulation in the period t-1, but there is water accumulation in the period t, the water depth in the period t is calculated as 0 cm;
[0012] (2) If there is no rainfall in the current period t, the water depth is a centimeters, and if there is no rainfall in the period t-1, the water depth is b centimeters. If b>a, then let b=a;
[0013] (3) If there is no rainfall in period t-1 and the water depth is less than 3 cm, and there is no rainfall in period t and the water depth is less than 3 cm, then the water depth in period t is set to 0 cm.
[0014] Furthermore, the method for constructing the node relationship matrix A' in step 3 is specifically as follows:
[0015] Step 3.1: Calculate the correlation between the water depth time series data of any two waterlogging monitoring stations. Take the waterlogging monitoring station as a node. If the correlation between the i-th node and the j-th node is lower than the preset threshold, then set the values of A[i][j] corresponding to the two nodes in the relationship matrix A to 0, otherwise set the value of A[i][j] to 1;
[0016] Step 3.2: Calculate the weight coefficient of A[i][j];
[0017] Step 3.3: Multiply the weight coefficient of A[i][j] by the value of A[i][j] to get the node relationship matrix A'.
[0018] Furthermore, in step 3.2, the weight coefficient W of A[i][j] is calculated according to the following formula:
[0019] W=W1+W2
[0020] Where W1 = 1 / elevation, W2 = exp(-distance ij / r), elevation represents the absolute elevation of the waterlogging monitoring station; distance ij represents the actual distance between the i-th waterlogging monitoring station and the j-th waterlogging monitoring station; r represents the average distance between all waterlogging monitoring stations.
[0021] Furthermore, the water depth prediction model in step 4 processes the input data as follows:
[0022] Step 4.1: Use the water accumulation time series feature extraction module to extract the time series features of the feature matrix X. The output feature of the water accumulation time series feature extraction module is temperate_At, and the dimension of temperate_At is [B, N, T*F];
[0023] Step 4.2: reconstruct the dimension of temperate_At into [B, N, T, F], and input the reconstructed temperate_At into the water accumulation depth spatial feature extraction module; the water accumulation depth spatial feature extraction module outputs the feature spatial_At with a dimension size of [B, N, N];
[0024] Step 4.3: Reconstruct the dimension of tempereral_At into [B, F*T, N], and then input the reconstructed tempereral_At and the feature spatial_At into the water depth feature fusion module to fuse the two features;
[0025] Step 4.4: Calculate the Laplace matrix and the second-order Chebyshev polynomial for the node relationship matrix A', perform matrix multiplication on the result of the second-order Chebyshev polynomial and the output of the water depth feature fusion module; input the result of the matrix multiplication and the feature matrix X into the residual module, and then use the ReLU activation function to perform nonlinear processing on the result of the residual module to obtain the final output.
[0026] Furthermore, the water accumulation time series feature extraction module in step 4.1 adopts a long short-term memory network.
[0027] Furthermore, the water depth spatial feature extraction module in step 4.2 calculates the input feature using the following formula:
[0028] spatial_At=softmax(W5·σ((temperal_AtW1W2)(W3temperal_At) T +W4))
[0029] Among them, W1 is the time step weight, W2 is a two-dimensional matrix with a size of [T, F], which is used to map the input features to each node and realize the conversion from feature space to node space, W3 is a one-dimensional matrix with a size of T, which is used to represent the weight of each channel of the input feature, so that the model can select and adjust at the feature level, W4 is a three-dimensional matrix with a size of [1, N, N], which is used to represent the similarity between time steps, W5 is a two-dimensional matrix with a size of [N, N], which is used to represent the mutual influence between time steps, σ is a nonlinear activation function, and softmax is the softmax function.
[0030] Beneficial effects: This method effectively combines the characteristics of waterlogging and precipitation, enriches the input information of the prediction model, and can capture the temporal and spatial characteristics of changes in waterlogging depth. It also makes full use of the extracted features for prediction, thereby improving the accuracy of waterlogging prediction. BRIEF DESCRIPTION OF THE DRAWINGS
[0031] Figure 1 It is a schematic diagram of the process of the present invention.
[0032] Figure 2 This is a distribution map of waterlogging monitoring stations of the present invention.
[0033] Figure 3 It is a schematic diagram of extracting spatiotemporal features of the water accumulation depth prediction model integrating spatiotemporal features of the present invention.
[0034] Figure 4 It is a schematic diagram of the structure of the water accumulation depth prediction model of the present invention.
[0035] Figure 5 The figure is a comparison chart between the predicted results using the prediction model of the present invention and the actual results. DETAILED DESCRIPTION
[0036] like Figure 1 The method for predicting waterlogging depth by integrating spatiotemporal features of the present invention is shown in the following steps: a sliding window of waterlogging and precipitation time series of a waterlogging monitoring station is set for training to obtain a waterlogging depth prediction model; and the waterlogging depth prediction model is applied to obtain the waterlogging depth prediction result for the next hour at the waterlogging monitoring stations in the region.
[0037] Step 1: Obtain the water depth time series data R of the waterlogging monitoring station i , precipitation time series data H i The latitude and longitude of the waterlogging monitoring station and the distance between them, where i represents any waterlogging monitoring station. At the same time, to ensure that the waterlogging is caused by rainfall, the original waterlogging data is preprocessed by the sliding window method. The distribution map of the 15 waterlogging monitoring stations to be predicted in this embodiment is as follows Figure 2 shown.
[0038] Step 2: According to the adjacent relationship between the waterlogging monitoring stations, an N×N weighted node relationship matrix A' is established, where N is the number of waterlogging monitoring stations.
[0039] Step 3: Convert the water depth time series data into R i and precipitation time series data H i The input historical data is divided by the sliding window method, where i represents any waterlogging monitoring station.
[0040] Step 4: Based on the training set, the waterlogging time series data R of each site i is iand precipitation time series data H i The constructed feature matrix X and the node relationship matrix A' with weights are taken as input, i represents any waterlogging monitoring station, and the waterlogging depth prediction model integrating spatiotemporal features is input for training and parameter tuning is performed by calculating the loss and back propagation algorithm to obtain the optimal waterlogging depth prediction model.
[0041] Step 5: Evaluate the accuracy of the obtained water depth prediction model. The evaluation indicators include MAE, MAPE, and RMSE.
[0042] In step 1, the data preprocessing method is as follows:
[0043] (1) If there is no rainfall in the current period t, no water accumulation in the period t-1, and there is water accumulation in the current period, the water depth in the current period t is calculated as 0 cm.
[0044] (2) If there is no rainfall in the current period t, the water depth is a centimeters. If there is no rainfall in the period t-1, the water depth is b centimeters. If b>a, then let b=a.
[0045] (3) If there is no rainfall in the current period t-1 and the water depth is less than 3 cm, and there is no rainfall in the time period t and the water depth is less than 3 cm, then the water depth in the time period t is set to 0 cm.
[0046] In step 2, the method for constructing the node relationship matrix A' with weights is as follows:
[0047] First, the correlation between the water depth time series of any two waterlogging monitoring stations is calculated. The degree of mutual influence between different stations is quantified by calculating the Pearson correlation coefficient, and the connection of the node relationship matrix is set according to these correlation coefficients. For node pairs with correlation less than the preset value, the corresponding elements in the node relationship matrix A[i][j] are set to 0. Otherwise, the corresponding elements are set to 1. Next, the weight coefficient W of the node relationship matrix A[i][j] is defined. This weight coefficient takes into account two factors: the elevation of the waterlogging monitoring station and the distance between the waterlogging monitoring stations. Among them, the first part of W, W1, is defined as 1 / elevation, and elevation represents the absolute elevation of the waterlogging monitoring station; the second part of W, W2, is defined as exp(-distance / r), and distance represents the actual distance between waterlogging monitoring stations i and j, and r represents the average distance between all waterlogging monitoring stations. Finally, the weight coefficient W is obtained by adding W1 and W2; the node relationship matrix A' with weights can be obtained by multiplying W and the node relationship matrix.
[0048] Step 3 is as follows:
[0049] First, we select three hours as the window size based on the water accumulation data of the monitoring station. Starting from the starting point of the time series, we slide the window to the end of the sequence, each time for three hours. That is, we use the water accumulation depth data of the first three hours to predict the water accumulation depth of the next hour. Then we divide the training set and the test set, where 80% of the sample data is used to construct the training set and 20% of the sample data is used to construct the test set.
[0050] The waterlogging time series data R of all waterlogging monitoring stations in step 4 i and precipitation time series data H i The dimension of the feature matrix X is [B, N, T, F], where B is the number of samples input each time the network is trained, N is the number of waterlogging monitoring stations, T is the size of the sliding window, and F is the two features of waterlogging depth and precipitation; Figure 3 As shown in the figure, the feature matrix X composed of the waterlogging and precipitation data of each waterlogging monitoring station and the node relationship matrix A' with weights are input into the waterlogging depth prediction model for training. The node relationship between the stations is used to extract features in the spatial dimension. At the same time, the node relationship between the stations at different times will be passed to the next time through the network layer to further extract features from the time dimension.
[0051] like Figure 4 As shown in Figure 1, the water depth prediction model includes a Chebyshev convolution module, a residual module, and a ReLu activation function module. The processing flow of the Chebyshev convolution module is as follows:
[0052] (1) The water accumulation time series feature extraction module extracts features through a long short-term memory network. The input dimension of this module is T, the hidden layer dimension is T, and the number of network layers is 2. The feature matrix X is input into the water accumulation time series feature extraction module to obtain temperate_At. The data dimension before input is [B, N, T, F], and the data dimension after feature extraction is [B, N, T*F]. In order to reorganize the data in the tensor, the dimension of temperate_At is reconstructed into [B, N, T, F].
[0053] (2) The water depth spatial feature extraction module generates the node attention coefficient through five learnable weight matrices W1, W2, W3, W4 and W5:
[0054] spatial_At=softmax(W5·σ((temperal_AtW1W2)(W3temperal_At) T +W4))
[0055] Among them, W1 is a one-dimensional matrix with a size of F. This matrix can be used to represent the bias term of each node. The input tensor temperate_At is first multiplied by the time step weight W1 to obtain temperate_AtW1. This is a weighted operation in the time dimension, and the result dimension is [B, N, F]. Different importance is given to each node in the calculation; W2 is a two-dimensional matrix with a size of [T, F]. Temperal_AtW1 is multiplied by W2 to obtain temperate_AtW1W2. This step is further weighted in the feature dimension, and the result dimension is [B, N, T]. W2 is used to map the input features to each node to realize the conversion from feature space to node space; W3 is a one-dimensional matrix with a size of T. Temperal_At is multiplied by W3, and the result dimension is [B, N, T]. It is used to represent the weight of each channel of the input feature, helping the model to select and adjust at the feature level; W4 is a three-dimensional matrix of size [1,N,N], which is used to represent the similarity between time steps and helps the model capture the timing characteristics of time series data; W5 is a two-dimensional matrix of size [N,N], which is used to represent the mutual influence between time steps and helps the model understand the dynamic changes of time series data. The output result is spatial_At, and the dimension size is [B,N,N].
[0056] (3) The dimension of tempereral_At is reconstructed into [B, F*T, N], and then it is input into the water accumulation depth feature fusion module together with spatial_At. The output data dimension of the water accumulation depth feature fusion module is [B, N, T, F]. The Laplace matrix and the second-order Chebyshev polynomial are calculated according to the node relationship matrix A', and then the polynomial calculation result is matrix multiplied with the time_conv_output obtained by the feature fusion module.
[0057] The result of matrix multiplication and the original input feature X are input into the residual module, and then nonlinearly processed by the ReLU activation function to obtain the final output.
[0058] The present invention calculates the loss through the mean square error loss function, and its calculation formula is as follows:
[0059] Where n is the number of samples, is the sample prediction value, and y is the true value.
[0060] The calculation methods of the main evaluation indicators MAE, MAPE, and RMSE in the present invention are as follows:
[0061] (1)MAE
[0062] MAE is the mean absolute error, which is the average of the absolute values of the differences between the predicted values and the actual values. It is used to measure the accuracy of the regression model. The smaller the value, the better.
[0063] Where n is the number of samples input into the model in each round, is the predicted value of the sample, y i1 is the true value of the sample, and i1 is an arbitrary sample.
[0064] (2)MAPE
[0065] MAPE is the mean absolute percentage error, which is the average value of the ratio of the absolute value of the difference between the predicted value and the actual value to the actual value. It is used to measure the accuracy of the regression model. The smaller the value, the better.
[0066] Where n is the number of samples input into the model in each round, is the predicted value of the sample, y i1 is the true value of the sample, and i1 is an arbitrary sample.
[0067] (3)RMSE
[0068] RMSE is the root mean square error, which is the square root of MSE. It is used to measure the accuracy of the regression model. The smaller the value, the better.
[0069] Where n is the number of samples input into the model in each round, is the predicted value of the sample, y i1 is the true value of the sample, and i1 is an arbitrary sample.
[0070] The hourly waterlogging depth time series data and precipitation time series data of the 15 waterlogging monitoring stations in this embodiment on August 15, 2020 are shown in Tables 1 and 2.
[0071] Table 1
[0072] 118 119 121 123 124 125 126 127 128 129 130 131 132 133 134 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 2 0 1 1 2 4 0 1 0 1 0 0 0 0 0 0 3 0 0 1 0 4 0 1 0 1 0 0 0 0 0 0 4 0 0 1 0 5 0 1 0 1 0 0 0 0 0 0 5 0 0 0 0 4 0 0 0 0 0 0 0 0 0 0 6 0 0 1 3 4 0 1 0 0 0 0 0 0 0 0 7 0 0 0 1 3 0 0 0 0 0 0 0 0 0 0 8 0 0 0 0 4 0 1 0 0 0 0 0 0 0 0 9 0 1 0 4 5 0 2 0 0 0 0 0 0 0 0 10 0 1 3 1 4 0 1 0 1 1 0 0 0 1 1 11 0 0 1 0 4 0 1 0 1 1 0 0 0 1 0 12 0 0 1 1 4 0 1 0 1 0 0 0 0 0 0 13 0 0 0 0 4 0 0 0 0 0 0 0 0 0 0 14 0 0 1 0 3 0 1 0 0 0 0 0 0 0 0 15 0 0 1 0 3 0 1 0 0 0 0 0 0 0 0 16 0 0 0 0 4 0 0 0 0 0 0 0 0 0 0 17 0 0 1 0 4 0 0 0 1 0 0 0 0 0 0 18 0 0 0 0 5 0 0 0 0 0 0 0 0 0 0 19 0 0 0 0 2 0 0 0 0 0 0 0 0 0 0 20 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 21 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 22 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 23 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 24 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
[0073] Table 2
[0074]
[0075]
[0076] According to the adjacent relationship between all waterlogging monitoring stations, a 15×15 node relationship matrix A' with weights is established as shown in Table 3:
[0077] Table 3
[0078] 118 119 121 123 124 125 126 127 128 129 130 131 132 133 134 118 0.0 0.4 0.2 0.7 0.3 0.5 0.3 0.5 0.2 0.2 0.2 0.2 0.2 0.3 0.3 119 0.4 0.0 0.2 0.4 0.2 0.7 0.2 0.4 0.2 0.4 0.3 0.2 0.2 0.3 0.3 121 0.2 0.2 0.0 0.3 0.6 0.2 0.5 0.4 0.9 0.3 0.5 0.6 0.7 0.7 0.6 123 0.7 0.4 0.3 0.0 0.3 0.5 0.5 0.7 0.3 0.2 0.3 0.2 0.3 0.4 0.4 124 0.3 0.2 0.6 0.3 0.0 0.3 0.7 0.5 0.6 0.2 0.3 0.4 0.5 0.7 0.7 125 0.5 0.7 0.2 0.5 0.3 0.0 0.3 0.5 0.3 0.4 0.4 0.3 0.3 0.4 0.4 126 0.3 0.2 0.5 0.5 0.7 0.3 0.0 0.6 0.5 0.2 0.3 0.3 0.4 0.6 0.7 127 0.5 0.4 0.4 0.7 0.5 0.5 0.6 0.0 0.4 0.3 0.4 0.3 0.3 0.5 0.6 128 0.2 0.2 0.9 0.3 0.6 0.3 0.5 0.4 0.0 0.3 0.5 0.7 0.8 0.8 0.7 129 0.2 0.4 0.3 0.2 0.2 0.4 0.2 0.3 0.3 0.0 0.6 0.5 0.4 0.4 0.3 130 0.2 0.3 0.5 0.3 0.3 0.4 0.3 0.4 0.5 0.6 0.0 0.8 0.6 0.6 0.5 131 0.2 0.2 0.6 0.2 0.4 0.3 0.3 0.3 0.7 0.5 0.8 0.0 0.9 0.6 0.5 132 0.2 0.2 0.7 0.3 0.5 0.3 0.4 0.3 0.8 0.4 0.6 0.9 0.0 0.6 0.5 133 0.3 0.3 0.7 0.4 0.7 0.4 0.6 0.5 0.8 0.4 0.6 0.6 0.6 0.0 0.8 134 0.3 0.3 0.6 0.4 0.7 0.4 0.7 0.6 0.7 0.3 0.5 0.5 0.5 0.8 0.0
[0079] The sliding window is set to 3 hours, and the results are shown in Table 4. Taking the waterlogging monitoring station 119 as an example, the leftmost column is time, the middle two columns are waterlogging and precipitation data, and the rightmost column is the sliding window division. The 1st, 2nd, and 3rd hour data are used to predict the 4th hour data, and so on. Finally, the waterlogging and precipitation time series data are divided into 80% training sets and 20% test sets.
[0080] Table 4
[0081]
[0082]
[0083] The prediction results output by this embodiment are as follows: Figure 5 As shown, Figure 5 The figure below shows the predicted and true values of water depth at a certain station in the test set for 24 hours a day for five consecutive days. Figure 5 (a) shows the comparison between the predicted result and the real result of the waterlogging observation station with station number 118; (b) shows the comparison between the predicted result and the real result of the waterlogging observation station with station number 121; (c) shows the comparison between the predicted result and the real result of the waterlogging observation station with station number 124; gray represents the real value, black represents the predicted value, the vertical axis represents the waterlogging depth per hour in centimeters (cm), and the horizontal axis represents time. Figure 5 It can be intuitively seen that the prediction accuracy of the model proposed in this embodiment is good.
[0084] It should also be noted that the various specific technical features described in the above specific embodiments can be combined in any suitable manner without contradiction. In order to avoid unnecessary repetition, the present invention will not further describe various possible combinations.
Claims
1. A method for predicting water depth by integrating spatiotemporal features, characterized in that: The specific steps include: Step 1: Obtain the water depth time series data R of each waterlogging monitoring station i , precipitation time series data H i , the longitude and latitude of the waterlogging monitoring station and the distance between the ith waterlogging monitoring station and other waterlogging monitoring stations; i = 1, 2, ..., I; I represents the total number of waterlogging monitoring stations; Step 2: To ensure that the waterlogging is caused by rainfall, the original waterlogging depth time series data is preprocessed based on the sliding window method; Step 3: Establish an N×N-dimensional node relationship matrix A' with weights based on the relationships between waterlogging monitoring stations; The preprocessed water depth time series data and the corresponding precipitation time series data are combined into a feature matrix X, the dimension of which is [B, N, T, F], where B is the number of samples, T is the size of the sliding window, and F is the number of features; Step 4: Construct a water depth prediction model, use the feature matrix X and the node relationship matrix A' as the water depth prediction model, and train the water depth prediction model; Step 5: Use the trained waterlogging depth prediction model to predict the waterlogging depth.
2. The method for predicting water depth by integrating spatiotemporal features according to claim 1, characterized in that: The pretreatment in step 2 comprises the following steps: (1) If there is no rainfall in the current period t, there is no water accumulation in the period t-1, but there is water accumulation in the period t, the water depth in the period t is calculated as 0 cm; (2) If there is no rainfall in the current period t, the water depth is a centimeters, and if there is no rainfall in the period t-1, the water depth is b centimeters. If b>a, then let b=a; (3) If there is no rainfall in period t-1 and the water depth is less than 3 cm, and there is no rainfall in period t and the water depth is less than 3 cm, then the water depth in period t is set to 0 cm.
3. The method for predicting water depth by integrating spatiotemporal features according to claim 1, characterized in that: The method for constructing the node relationship matrix A' in step 3 is specifically as follows: Step 3.1: Calculate the correlation between the water depth time series data of any two waterlogging monitoring stations. Take the waterlogging monitoring station as a node. If the correlation between the i-th node and the j-th node is lower than the preset threshold, then set the values of A[i][j] corresponding to the two nodes in the relationship matrix A to 0, otherwise set the value of A[i][j] to 1; Step 3.2: Calculate the weight coefficient of A[i][j]; Step 3.3: Multiply the weight coefficient of A[i][j] by the value of A[i][j] to get the node relationship matrix A'.
4. The method for predicting water depth by integrating spatiotemporal features according to claim 3, characterized in that: In step 3.2, the weight coefficient W of A[i][j] is calculated according to the following formula: W=W1+W2 Where W1 = 1 / elevation, W2 = exp(-distance ij / r), elevation represents the absolute elevation of the waterlogging monitoring station; distance ij represents the actual distance between the i-th waterlogging monitoring station and the j-th waterlogging monitoring station; r represents the average distance between all waterlogging monitoring stations.
5. The method for predicting water depth by integrating spatiotemporal features according to claim 1, characterized in that: The water depth prediction model in step 4 processes the input data as follows: Step 4.1: Use the water accumulation time series feature extraction module to extract the time series features of the feature matrix X. The output feature of the water accumulation time series feature extraction module is temperate_At, and the dimension of temperate_At is [B, N, T*F]; Step 4.2: reconstruct the dimension of temperate_At into [B, N, T, F], and input the reconstructed temperate_At into the water accumulation depth spatial feature extraction module; the water accumulation depth spatial feature extraction module outputs the feature spatial_At with a dimension size of [B, N, N]; Step 4.3: Reconstruct the dimension of tempereral_At into [B, F*T, N], and then input the reconstructed tempereral_At and the feature spatial_At into the water depth feature fusion module to fuse the two features; Step 4.4: Calculate the Laplace matrix and the second-order Chebyshev polynomial for the node relationship matrix A', perform matrix multiplication on the result of the second-order Chebyshev polynomial and the output of the water depth feature fusion module; input the result of the matrix multiplication and the feature matrix X into the residual module, and then use the ReLU activation function to perform nonlinear processing on the result of the residual module to obtain the final output.
6. The method for predicting water depth by integrating spatiotemporal features according to claim 5, characterized in that: The water accumulation time series feature extraction module in step 4.1 adopts a long short-term memory network.
7. The method for predicting water depth by integrating spatiotemporal features according to claim 5, characterized in that: The water depth spatial feature extraction module in step 4.2 calculates the input feature using the following formula: spatial_At=softmax(W5·σ((temperal_AtW1W2)(W3temperal_At) T +W4)) Among them, W1 is the time step weight, W2 is a two-dimensional matrix with a size of [T, F], which is used to map the input features to each node and realize the conversion from feature space to node space, W3 is a one-dimensional matrix with a size of T, which is used to represent the weight of each channel of the input feature, so that the model can select and adjust at the feature level, W4 is a three-dimensional matrix with a size of [1, N, N], which is used to represent the similarity between time steps, W5 is a two-dimensional matrix with a size of [N, N], which is used to represent the mutual influence between time steps, σ is a nonlinear activation function, and softmax is the softmax function.
