ConvLSTM model-based urban rainstorm water depth prediction method and system

Abstract

The invention discloses an urban rainstorm water depth prediction method and system based on a ConvLSTM model, and the method comprises the steps: carrying out the data collection of a to-be-simulated region, and constructing an urban rainfall flood model of the to-be-simulated region; performing simulation prediction and data preprocessing on the rainstorm event data information of the to-be-simulated area based on the urban rainfall flood model of the to-be-simulated area to obtain normalized rainfall spatio-temporal characteristic data of the to-be-simulated area; constructing a ConvLSTM model based on a convolutional neural network and a long short-term memory network; and performing water depth prediction on the normalized rainfall spatial-temporal characteristic data of the to-be-simulated area based on a ConvLSTM model to obtain a rainstorm water depth prediction result of the to-be-simulated area. According to the embodiment of the invention, the accuracy of real-time waterlogging disaster forecasting is improved by considering the correlation of the adjacent space grids and the influence of the space-time structure of rainfall on the flood process. The urban rainstorm water depth prediction method and system based on the ConvLSTM model can be widely applied to the technical field of urban rainstorm waterlogging simulation.

Description

Technical Field

[0001] The present invention relates to the technical field of urban rainstorm waterlogging simulation, and in particular to a method and system for predicting urban rainstorm water depth based on a ConvLSTM model. Background Art

[0002] Most of the current urban flood control and drainage planning is based on the assumption of uniform rainfall in order to simplify the design and calculation. However, in actual situations, heavy rains usually show obvious spatial heterogeneity, so the rainfall shows the characteristics of uneven rainfall distribution in space. Existing studies have shown that this significant spatiotemporal characteristic of rainfall may aggravate urban waterlogging disasters. Therefore, when facing urban waterlogging events that may be caused by heavy rains, the spatiotemporal characteristics of heavy rains should be taken into consideration; the impact of waterlogging can be greatly reduced by early warning of rain and floods and formulating appropriate response measures. However, waterlogging caused by heavy rains usually occurs suddenly and has a large impact range. In urban areas, due to the variety of surface cover and complex distribution, these characteristics bring certain difficulties to the early warning and prediction of waterlogging. Existing research is also committed to improving the timeliness and accuracy of flood forecasting. At present, early warning and prediction of waterlogging are mainly based on flood models of physical equations. Physical models are mainly divided into hydrological models, hydrodynamic models, and hydrological and hydrodynamic coupling models. They achieve more detailed simulation of flood inundation by solving a large number of complex physical equations. However, physical model building requires high-precision terrain data and rich hydrological data, which are often difficult to obtain. At the same time, solving physical equations takes a long time. These problems have brought certain limitations to the physical model, which in turn leads to unsatisfactory results in real-time flood disaster forecasting. Summary of the invention

[0003] In order to solve the above technical problems, the purpose of the present invention is to provide a method and system for predicting urban rainstorm water depth based on a ConvLSTM model, which improves the accuracy of real-time urban waterlogging disaster forecasting by considering the correlation between adjacent spatial grids and the impact of the spatiotemporal structure of rainfall on the flood process.

[0004] The first technical solution adopted by the present invention is: a method for predicting urban rainstorm water depth based on a ConvLSTM model, comprising the following steps:

[0005] Collect data for the area to be simulated and build an urban rain and flood model for the area to be simulated;

[0006] Based on the urban rain and flood model of the area to be simulated, the rainstorm event data information of the area to be simulated is simulated and predicted and data preprocessed to obtain the normalized spatiotemporal characteristic data of rainfall in the area to be simulated;

[0007] Construct the ConvLSTM model based on convolutional neural network and long short-term memory network;

[0008] Based on the ConvLSTM model, the water depth prediction of the normalized spatiotemporal characteristic data of rainfall in the simulated area is carried out to obtain the prediction result of the heavy rain water depth in the simulated area.

[0009] Furthermore, the step of collecting data for the area to be simulated and constructing an urban rain and flood model for the area to be simulated specifically includes:

[0010] Obtain the land use type, elevation and drainage network data of the area to be simulated, and construct a one-dimensional hydrological model of the area to be simulated;

[0011] Obtain rainstorm data information of the area to be simulated and input it into the one-dimensional hydrological model of the area to be simulated for data derivation, so as to obtain the position coordinates and overflow information of the overflow pipe point under the rainfall conditions of the area to be simulated;

[0012] Acquire the elevation information of the area to be simulated and perform a two-dimensional evolution simulation based on the position coordinates and overflow information of the overflow pipe point under the rainfall conditions in the area to be simulated, and construct a two-dimensional hydrodynamic model of the area to be simulated;

[0013] The one-dimensional hydrological model of the area to be simulated is coupled with the two-dimensional hydrodynamic model of the area to be simulated to construct an urban stormwater model of the area to be simulated.

[0014] Furthermore, the step of simulating and predicting the rainstorm event data information of the simulated area based on the urban rain and flood model of the simulated area and preprocessing the data to obtain the normalized spatiotemporal characteristic data of rainfall in the simulated area specifically includes:

[0015] Obtain data information on rainstorm events with different return periods in the simulated area and with different moving directions and paths of the rainstorm center;

[0016] The spatially variable rainfall model based on a fixed spatial structure is used to scale the rainstorm event data information to obtain rainfall information with different spatial and temporal characteristics in the simulated area.

[0017] Inputting rainfall information with different spatiotemporal characteristics of the area to be simulated into the urban rain and flood model of the area to be simulated to perform rainfall simulation prediction, and obtaining the predicted spatiotemporal characteristic data of rainfall in the area to be simulated, wherein the predicted spatiotemporal characteristic data of rainfall in the area to be simulated includes the flooding range, flooding depth and hourly evolution process of flooding under different spatiotemporal rainstorm scenarios in the area to be simulated;

[0018] The predicted spatiotemporal characteristic data of rainfall in the area to be simulated are normalized to obtain normalized spatiotemporal characteristic data of rainfall in the area to be simulated.

[0019] Furthermore, the expression of the spatially variable rainfall model based on a fixed spatial structure is specifically as follows:

[0020]

[0021] R(x,y,μ1,μ2,σ,t)=γ(μ1,μ2)f(x,y,μ1,μ2,σ)r(t)

[0022]

[0023] In the above formula, R(x,y,μ1,μ2,σ,t) represents the rainfall intensity at position (x,y) at time t in the spatially variable rainfall model with fixed spatial structure, (μ1,μ2) represents the position coordinates of the rainfall center, r(t) represents the rainfall intensity at time t under spatially uniform rainfall conditions, A represents the basin area, f(x,y,μ1,μ2,σ) is the distribution function of spatial rainfall intensity represented by a two-dimensional truncated Gaussian distribution function, σ represents the standard deviation of the Gaussian distribution, and γ(μ1,μ2) represents the scaling function of the two-dimensional truncated Gaussian distribution, the purpose of which is to ensure that the integral of the Gaussian distribution over the domain is equal to 1.

[0024] Furthermore, the expression for normalizing the predicted spatiotemporal characteristic data of rainfall in the area to be simulated is specifically as follows:

[0025]

[0026] In the above formula, X n represents the normalized temporal and spatial characteristic data of rainfall in the simulated area, X represents the sequence value, and X min Indicates the minimum value of the sequence value, X max Indicates the maximum value of the sequence.

[0027] Furthermore, the ConvLSTM model specifically includes a two-dimensional convolutional long short-term memory network layer, a two-dimensional convolutional neural network layer and a fully connected neural network layer, wherein the output end of the two-dimensional convolutional long short-term memory network layer is connected to the input end of the two-dimensional convolutional neural network layer, and the output end of the two-dimensional convolutional neural network layer is connected to the input end of the fully connected neural network layer.

[0028] Further, the step of performing water depth prediction on the normalized spatiotemporal characteristic data of rainfall in the area to be simulated based on the ConvLSTM model to obtain a rainstorm water depth prediction result for the area to be simulated specifically includes:

[0029] The normalized spatiotemporal characteristic data of rainfall in the area to be simulated is input into the ConvLSTM model;

[0030] Based on the two-dimensional convolutional long short-term memory network layer of the ConvLSTM model, the normalized spatiotemporal characteristic data of rainfall in the area to be simulated are processed to capture the spatiotemporal characteristics and obtain the spatiotemporal characteristic map of rainfall in the area to be simulated.

[0031] Based on the two-dimensional convolutional neural network layer of the ConvLSTM model, the spatial information feature extraction process is performed on the spatiotemporal feature map of rainfall in the simulated area to obtain the spatiotemporal feature map of rainfall in the simulated area after convolution;

[0032] Based on the fully connected neural network layer of the ConvLSTM model, the spatiotemporal characteristic map of rainfall in the simulated area after convolution is processed by feature integration to obtain the prediction result of heavy rain depth in the simulated area.

[0033] Furthermore, the expression of the two-dimensional convolutional long short-term memory network layer based on the ConvLSTM model is specifically as follows:

[0034]

[0035] In the above formula, σ represents the sigmoid activation function, tanh(·) represents the hyperbolic tangent activation function, and x t represents the input of the current time step, h t-1 represents the hidden state of the previous time step, c t-1 represents the memory unit of the previous time step, f t 、i t 、c t , o t Respectively represent the forget gate, input gate, memory unit and output gate, * represents the convolution operator, represents the Hadamard product.

[0036] Furthermore, the expression of the two-dimensional convolutional neural network layer based on the ConvLSTM model is specifically as follows:

[0037]

[0038] In the above formula, represents the input of the kth neuron in the lth convolution layer, represents the bias of the kth neuron in the lth convolution layer, represents the output of the i-th neuron in the (l-1)th layer of convolution, represents the convolution kernel from the i-th neuron in the (l-1)-th layer of convolution to the k-th neuron in the l-th layer of convolution, and Conv2D(·) represents a two-dimensional convolution operation.

[0039] The second technical solution adopted by the present invention is: a city rainstorm water depth prediction system based on ConvLSTM model, comprising:

[0040] The first module is used to collect data in the area to be simulated and build an urban rain and flood model for the area to be simulated;

[0041] The second module is used to simulate and predict the rainstorm event data information of the simulated area based on the urban rain and flood model of the simulated area and preprocess the data to obtain the normalized spatiotemporal characteristic data of rainfall in the simulated area;

[0042] The third module is used to build a ConvLSTM model based on convolutional neural network and long short-term memory network;

[0043] The fourth module is used to predict the water depth of the normalized spatiotemporal characteristic data of rainfall in the simulated area based on the ConvLSTM model to obtain the prediction result of the heavy rain water depth in the simulated area.

[0044] The beneficial effects of the method and system of the present invention are as follows: the present invention collects data for the area to be simulated, constructs an urban rain and flood model for the area to be simulated, further simulates and predicts the rainstorm event data information for the area to be simulated based on the urban rain and flood model, and preprocesses the data, so as to obtain normalized spatiotemporal characteristic data of rainfall in the area to be simulated, fully considers the influence of the spatiotemporal structure of rainfall on the flood process and the complex spatiotemporal structure of rainfall, and then constructs a ConvLSTM model based on a convolutional neural network and a long short-term memory network, predicts the water depth of the normalized spatiotemporal characteristic data of rainfall in the area to be simulated, and considers the correlation of adjacent spatial grids, so as to realize the rapid and accurate simulation of the evolution process of waterlogging under complex spatiotemporal rainfall conditions, thereby realizing the rapid and accurate simulation of the scope, depth and evolution process of waterlogging based on rainfall forecast input conditions, and finally improving the accuracy of real-time waterlogging disaster forecast. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 It is a flow chart of the steps of a method for predicting urban rainstorm water depth based on a ConvLSTM model of the present invention;

[0046] Figure 2 It is a structural block diagram of a system for predicting urban rainstorm depth based on a ConvLSTM model of the present invention;

[0047] Figure 3 It is a schematic diagram of a one-dimensional hydrological model SWMM of a study area provided by a specific embodiment of the present invention;

[0048] Figure 4 It is a schematic diagram of the movement path of the rainstorm center in the study area of ​​the rainstorm scenario provided by the specific embodiment of the present invention;

[0049] Figure 5It is a schematic diagram of the correlation between the spatial difference map of the flooding depth and the simulated water depth of a certain flood-prone area in a Chicago-type flooding event with a return period of 100 years provided by a specific embodiment of the present invention;

[0050] Figure 6 It is a schematic diagram of the correlation between the spatial difference map of the flooding depth and the simulated water depth of a certain flood-prone area in a Chicago-type flooding event with a return period of 80 years provided by a specific embodiment of the present invention;

[0051] Figure 7 It is a schematic diagram of the correlation between the spatial difference map of the flooding depth and the simulated water depth of a certain flood-prone area in a Chicago-type flooding event with a return period of 50 years provided by a specific embodiment of the present invention;

[0052] Figure 8 It is a schematic diagram of the correlation between the spatial difference map of the flooding depth and the simulated water depth of a certain flood-prone area in a Chicago-type flooding event with a return period of 20 years provided by a specific embodiment of the present invention;

[0053] Fig. 9 It is a schematic diagram of the correlation between the spatial difference map of the flooding depth and the simulated water depth of a certain flood-prone area in a Chicago-type flooding event with a return period of 15 years provided by a specific embodiment of the present invention;

[0054] Fig.10 It is a schematic diagram of the correlation between the spatial difference map of the flooding depth and the simulated water depth of a certain flood-prone area in a Chicago-type flooding event with a return period of 10 years provided by a specific embodiment of the present invention;

[0055] Fig.11 It is a schematic diagram of the correlation between the spatial difference map of the flooding depth and the simulated water depth of a certain flood-prone area in a Chicago-type flood with a return period of 2 years provided by a specific embodiment of the present invention;

[0056] Fig.12 It is a schematic diagram of the correlation between a spatial difference map of the inundation depth of a certain flood-prone area in a Chicago-type flood with a return period of one year and a simulated water depth provided by a specific embodiment of the present invention. DETAILED DESCRIPTION

[0057] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. The step numbers in the following embodiments are only provided for the convenience of explanation and description, and the order between the steps is not limited in any way. The execution order of each step in the embodiment can be adaptively adjusted according to the understanding of those skilled in the art.

[0058] Reference Figure 1 The present invention provides a method for predicting urban rainstorm water depth based on a ConvLSTM model, the method comprising the following steps:

[0059] S100, collecting data for the area to be simulated, and constructing an urban rain and flood model for the area to be simulated;

[0060] Specifically, the land use type, elevation and drainage network data of the area to be simulated are obtained to construct a one-dimensional hydrological model of the area to be simulated; the rainstorm data information of the area to be simulated is obtained and input into the one-dimensional hydrological model of the area to be simulated for data derivation to obtain the position coordinates and overflow volume information of the overflow pipe point under the rainfall conditions of the area to be simulated; the elevation information of the area to be simulated is obtained and combined with the position coordinates and overflow volume information of the overflow pipe point under the rainfall conditions of the area to be simulated to perform a two-dimensional evolution simulation to construct a two-dimensional hydrodynamic model of the area to be simulated; the one-dimensional hydrological model of the area to be simulated is coupled with the two-dimensional hydrodynamic model of the area to be simulated to construct an urban stormwater model of the area to be simulated.

[0061] In this embodiment, an urban stormwater model is constructed based on the basic data of the study area. Specifically, the land use type, elevation and drainage network data of the study area are collected, and a coupled stormwater simulation based on the one-dimensional hydrological model SWMM and the two-dimensional hydrodynamic model LISFLOOD-FP is constructed. In the coupled stormwater simulation, the land use type, elevation and drainage network data are used to construct the SWMM model; the rainstorm data is input into the SWMM model, and the SWMM model derives the location coordinates and overflow information of the overflow pipe point under rainfall conditions, which are input into the LISFLOOD-FP model and combined with the elevation information to perform a two-dimensional evolution simulation of flooding and accumulation of water. Finally, the LISFLOOD-FP model is used to output the flooding range, flooding depth and flooding evolution process information; the constructed composite flood simulation is parameterized and verified by adjusting the model parameters until the simulation accuracy meets the requirements. Figure 3 As shown, taking a certain river basin in a certain city as an example (the area is about 116.26km 2 ), and construct an urban waterlogging model.

[0062] S200, based on the urban rain and flood model of the area to be simulated, simulate and predict the rainstorm event data information of the area to be simulated and preprocess the data to obtain normalized spatiotemporal characteristic data of rainfall in the area to be simulated;

[0063] Specifically, data information of rainstorm events with different recurrence periods and different moving directions and paths of the rainstorm center in the area to be simulated is obtained; the rainstorm event data information is scaled by a spatially variable rainfall model based on a fixed spatial structure to obtain rainfall information with different spatiotemporal characteristics in the area to be simulated; the rainfall information with different spatiotemporal characteristics in the area to be simulated is input into an urban rain and flood model of the area to be simulated to perform rainfall simulation prediction to obtain predicted spatiotemporal characteristic data of rainfall in the area to be simulated, wherein the predicted spatiotemporal characteristic data of rainfall in the area to be simulated includes the inundation range, inundation depth and hourly evolution process of inundation under different spatiotemporal rainstorm scenarios in the area to be simulated; data normalization is performed on the predicted spatiotemporal characteristic data of rainfall in the area to be simulated to obtain normalized spatiotemporal characteristic data of rainfall in the area to be simulated.

[0064] In this embodiment, rainfall scenarios with different spatiotemporal characteristics are designed, a spatially variable rainfall model based on a fixed spatial structure (SVR-SS) is used to scale the rainfall in the region, and a spatiotemporal rainstorm-waterlogging dataset is generated using the constructed coupled rain-flood simulation model.

[0065] It should be noted that different spatiotemporal characteristics refer to rainstorm events with different recurrence periods and different moving directions and paths of the rainstorm center. Different recurrence periods include 1, 2, 5, 10, 15, 20, 30, 50, 80, and 100-year recurrence periods; different moving directions and paths of the rainstorm center refer to the rainstorm center passing through the study area horizontally from left to right or from right to left, passing through the study area vertically from top to bottom or from bottom to top, passing through the study area obliquely from top left to bottom right or from top right to bottom left or from bottom left to top right or from bottom right to top left, such as Figure 4 shown.

[0066] The spatially variable rainfall model with fixed spatial structure (SVR-SS) refers to a method that uses the density function of a two-dimensional truncated Gaussian distribution to represent the spatial structure of rainfall while ensuring that the total rainfall in the basin is the same. Its control equation is as follows:

[0067]

[0068] R(x,y,μ1,μ2,σ,t)=γ(μ1,μ2)f(x,y,μ1,μ2,σ)r(t)

[0069]

[0070] In the above formula, R(x,y,μ1,μ2,σ,t) represents the rainfall intensity at position (x,y) at time t in the spatially variable rainfall model with fixed spatial structure, (μ1,μ2) represents the position coordinates of the rainfall center, r(t) represents the rainfall intensity at time t under spatially uniform rainfall conditions, A represents the basin area, f(x,y,μ1,μ2,σ) is the distribution function of spatial rainfall intensity represented by a two-dimensional truncated Gaussian distribution function, σ represents the standard deviation of the Gaussian distribution, and γ(μ1,μ2) represents the scaling function of the two-dimensional truncated Gaussian distribution, the purpose of which is to ensure that the integral of the Gaussian distribution over the domain is equal to 1.

[0071] In addition, it should be noted that the spatiotemporal rainstorm-waterlogging database with different characteristics refers to the input of rainfall processes with different spatiotemporal characteristics into the constructed coupled rain-flood model to simulate the flooding range, flooding depth and hourly evolution of flooding under different spatiotemporal rainstorm scenarios.

[0072] Furthermore, the generated data set is normalized and divided into a training set, a validation set and a test set. In this embodiment, the normalization method adopts the min-max normalization method to normalize the rainstorm waterlogging input features and the flooding depth sequence data respectively. The normalization is calculated by the following formula, which is expressed as follows:

[0073]

[0074] In the above formula, X n represents the normalized temporal and spatial characteristic data of rainfall in the simulated area, X represents the sequence value, and X min Indicates the minimum value of the sequence value, X max Indicates the maximum value of the sequence.

[0075] Finally, the normalized spatiotemporal rainstorm-waterlogging database was divided into training set, validation set and test set in a ratio of 8:1:1. In order to ensure the generalization ability of the model, the divided events were adjusted by random selection and manual adjustment to ensure that the test set contained rainfall events with different return periods and different rainstorm movement paths.

[0076] S300, based on convolutional neural network and long short-term memory network, build ConvLSTM model;

[0077] Specifically, the ConvLSTM model specifically includes a two-dimensional convolutional long short-term memory network layer, a two-dimensional convolutional neural network layer and a fully connected neural network layer, wherein the output end of the two-dimensional convolutional long short-term memory network layer is connected to the input end of the two-dimensional convolutional neural network layer, and the output end of the two-dimensional convolutional neural network layer is connected to the input end of the fully connected neural network layer.

[0078] In this embodiment, the ConvLSTM algorithm is an integrated deep learning algorithm based on a convolutional neural network and a long short-term memory network. The rapid prediction of waterlogging in the present invention uses a spatiotemporal rainfall process as input, and the ConvLSTM network used refers to a two-dimensional convolutional long short-term memory neural network (ConvLSTM 2D). The main structure of the constructed ConvLSTM network includes a two-dimensional convolutional long short-term memory network layer (ConvLSTM 2D), a two-dimensional convolutional neural network layer (Conv 2D) and a fully connected neural network layer.

[0079] The Keras library in Python was used to build a convolutional long short-term memory network (ConvLSTM) model. The convolutional long short-term memory layer of the ConvLSTM model was designed to be 3 layers, and the tanh function was selected as the activation function. The convolution kernel size was set to 7×7, and the number of convolution kernels was 14, 28, and 56 respectively. A two-dimensional convolution layer was connected after the convolutional long short-term memory layer for further feature extraction. The convolution kernel size was set to 7×7, and the number of convolution kernels was set to 1. The fully connected layer was designed to be 1 layer, with 1024 neurons, and the output layer corresponded to the water depth of different grids.

[0080] There are multiple parameters that need to be set and optimized when building the ConvLSTM model. In this study, the parameters of the ConvLSTM model were optimized by manual adjustment. The optimized parameter settings are shown in Table 1:

[0081]

[0082] S400, based on the ConvLSTM model, the normalized spatiotemporal characteristic data of rainfall in the area to be simulated is used to predict water depth, and a prediction result of heavy rain water depth in the area to be simulated is obtained.

[0083] Specifically, the normalized spatiotemporal characteristic data of rainfall in the area to be simulated are input into the ConvLSTM model; based on the two-dimensional convolutional long short-term memory network layer of the ConvLSTM model, the spatiotemporal characteristic capture processing is performed on the normalized spatiotemporal characteristic data of rainfall in the area to be simulated, and the spatiotemporal characteristic map of rainfall in the area to be simulated is obtained; based on the two-dimensional convolutional neural network layer of the ConvLSTM model, the spatial information feature extraction processing is performed on the spatiotemporal characteristic map of rainfall in the area to be simulated, and the spatiotemporal characteristic map of rainfall in the area to be simulated after convolution is obtained; based on the fully connected neural network layer of the ConvLSTM model, the feature integration processing is performed on the spatiotemporal characteristic map of rainfall in the area to be simulated after convolution, and the rainstorm water depth prediction result of the area to be simulated is obtained.

[0084] In this embodiment, it should be noted that the independent variable of the ConvLSTM network model is the rainfall spatiotemporal sequence process (the rainfall duration is 180 minutes, and the rainfall every 5 minutes is used as a feature, with a total of 36 features; the spatial resolution of rainfall is 8m, and the spatial rainfall scaling factor has a total of 2340×1632 features), and the dependent variable is the water depth of different spatial grids (a total of 345255 grids, with a grid resolution of 8m×8m). The design spatiotemporal rainstorm-waterlogging scenario is adjusted according to the ratio of 8:1:1 by random selection and manual adjustment. The data set is divided into a training set, a validation set, and a test set. A total of 64 samples are used for model training, and a total of 8 samples are used for validation during the training process.

[0085] The convolutional long short-term memory network layer can effectively capture the spatiotemporal characteristics of the input multi-dimensional feature grid data by combining the convolution operation with the long short-term memory network LSTM structure. It uses the inner product of the sliding window data and the convolution kernel to extract the data weights of different time steps and spatial positions, thereby generating a new feature map and retaining the time dependency.

[0086] The convolution layer focuses on feature extraction of the input data. Through the inner product operation of the sliding window and the convolution kernel, it extracts data features at different locations, forms a feature map, and then extracts spatial information.

[0087] The fully connected layer connects all the neural units in the feature map, passes through the hidden layer, and generates the final result by the output layer. This process integrates the extracted features to form the final prediction.

[0088] In the ConvLSTM network, the previous two-dimensional convolutional long short-term memory network layer is forward propagated to the neuron input of the next layer. The process is as follows:

[0089]

[0090] In the above formula, σ represents the sigmoid activation function, tanth(·) represents the hyperbolic tangent activation function, and x t represents the input of the current time step, h t-1 represents the hidden state of the previous time step, c t-1 represents the memory unit of the previous time step, f t 、i t 、c t , o t Respectively represent the forget gate, input gate, memory unit and output gate, * represents the convolution operator, represents the Hadamard product.

[0091] A 2D convolutional layer is added after the ConvLSTM layer to further extract features. The convolution operation of the convolutional neural network layer is as follows:

[0092]

[0093] In the above formula, represents the input of the kth neuron in the lth convolution layer, represents the bias of the kth neuron in the lth convolution layer, represents the output of the i-th neuron in the (l-1)th layer of convolution, represents the convolution kernel from the i-th neuron in the (l-1)-th layer of convolution to the k-th neuron in the l-th layer of convolution, and Conv2D(·) represents a two-dimensional convolution operation.

[0094] Finally, the output of the two-dimensional convolutional layer is used as the input of the fully connected layer to complete the construction of the model.

[0095] The spatiotemporal varying rainstorm sequence data in the test set are input into the constructed ConvLSTM model, and the changes in the flooded water depth of each spatial grid are output to achieve rapid simulation of the flooding range, flooding depth and flooding evolution process of rainstorm waterlogging.

[0096] Eight different test rainstorms were used as specific examples to test the ConvLSTM model's prediction effect on the spatiotemporal changes in water depth under different rainstorms. The eight test rainstorms were 100-year mobile path type 3, 80-year mobile path type 1, 50-year mobile path type 2, 20-year mobile path type 7, 15-year mobile path type 8, 10-year mobile path type 5, 2-year mobile path type 4, and 1-year mobile path type 6. Under each rainstorm, the ConvLSTM model predicts a total of three time steps (a total of 3 hours, with a prediction every hour) of spatial flooding depth. Since there are a large number of spatial grid maps predicted for each rainstorm, and the flooding situation in the third hour (just after the rain ends) is the most serious, the water depth grid map under the third hour is selected here to display and illustrate the spatial prediction effect. The ConvLSTM model predicted the temporal changes of the flooded water depth of 345,255 grids in total. The simulated values ​​of the coupled rain-flood model were compared with the predicted values ​​of the ConvLSTM model to evaluate the water depth prediction effect of the ConvLSTM model.

[0097] Under the 100-year moving path type 3 rainstorm, the difference and correlation (t=3h) between the water depth predicted by the ConvLSTM model of each grid and the simulated water depth of the urban waterlogging model based on SWMM+LISFLOOD-FP are shown in Figure 2. Figure 5 (a) and Figure 5 (b), where Figure 5 (a) The spatial difference map of flood depth in a flood-prone area under a 100-year Chicago-type (r=0.29) moving path 3 test rainstorm predicted by the convolutional long short-term memory network (ConvLSTM) model and simulated by the urban waterlogging model (SWMM+LISFLOOD-FP) (t=3h); Figure 5(b) Correlation between the water depth predicted by the CNN model and the simulated water depth of the urban waterlogging model for each grid. The spatial flooding water depth predicted by the ConvLSTM model is close to the result of the urban waterlogging model, and the flood-prone location coincides with the maximum water depth difference between the two is small (basically less than 0.1m). The water depth predicted by the ConvLSTM model has a high correlation with the water depth simulated by the urban waterlogging model, showing a significant linear correlation (P<0.001). At t=3h, the correlation coefficient R2 predicted by the ConvLSTM model is 0.951, the mean absolute error MAE is 0.051, and the mean square error MSE is 0.034. The results show that the difference between the water depth predicted by the ConvLSTM model and the simulated water depth of the urban waterlogging model over time is small, and the ConvLSTM model can better capture the trend of grid water depth over time.

[0098] Under the 80-year moving path type 1 rainstorm, the difference and correlation (t=3h) between the water depth predicted by the ConvLSTM model of each grid and the simulated water depth of the urban waterlogging model based on SWMM+LISFLOOD-FP are shown in Figure 2. Figure 6 (a) and Figure 6 (b) Figure 6 (a) The spatial difference map of flood depth in a flood-prone area under the test rainstorm of Chicago type (r = 0.29) with a return period of 80 years and moving path 1, predicted by the convolutional long short-term memory network (ConvLSTM) model and simulated by the urban waterlogging model (SWMM + LISFLOOD-FP) (t = 3h); Figure 6 (b) Correlation between the water depth predicted by the CNN model and the simulated water depth of the urban waterlogging model for each grid. The spatial flooding water depth predicted by the ConvLSTM model is close to the result of the urban waterlogging model, and the flood-prone location coincides with the maximum water depth difference between the two is small (basically less than 0.1m). The water depth predicted by the ConvLSTM model has a high correlation with the water depth simulated by the urban waterlogging model, showing a significant linear correlation (P<0.001). At t=3h, the correlation coefficient R2 predicted by the ConvLSTM model is 0.967, the mean absolute error MAE is 0.028, and the mean square error MSE is 0.015. The results show that the difference between the water depth predicted by the ConvLSTM model and the simulated water depth of the urban waterlogging model over time is small, and the ConvLSTM model can better capture the trend of grid water depth over time.

[0099] Under the 50-year moving path type 2 rainstorm, the difference and correlation (t=3h) between the water depth predicted by the ConvLSTM model of each grid and the simulated water depth of the urban waterlogging model based on SWMM+LISFLOOD-FP are shown in Figure 2. Figure 7 (a) and Figure 7 (b), where Figure 7(a) The spatial difference map of flood depth in a flood-prone area under the test rainstorm of Chicago type (r = 0.29) with moving path 2 (t = 3h) predicted by the convolutional long short-term memory network (ConvLSTM) model and simulated by the urban waterlogging model (SWMM + LISFLOOD-FP); Figure 7 (b) Correlation between the water depth predicted by the CNN model of each grid and the simulated water depth of the urban waterlogging model. The spatial flooding water depth predicted by the ConvLSTM model is close to the result of the urban waterlogging model, and the flood-prone locations are consistent. The maximum water depth difference between the two is small (basically less than 0.1m). The water depth predicted by the ConvLSTM model has a high correlation with the water depth simulated by the urban waterlogging model, showing a significant linear correlation (P<0.001). At t=3h, the correlation coefficient R predicted by the ConvLSTM model is 2 The average absolute error (MAE) is 0.958, the mean absolute error (MAE) is 0.037, and the mean square error (MSE) is 0.018. The results show that the difference between the water depth predicted by the ConvLSTM model and the simulated water depth of the urban waterlogging model over time is small, and the ConvLSTM model can better capture the trend of grid water depth over time.

[0100] Under the 20-year moving path type 7 torrential rain, the difference and correlation (t=3h) between the water depth predicted by the ConvLSTM model of each grid and the simulated water depth of the urban waterlogging model based on SWMM+LISFLOOD-FP are shown in Figure 2. Figure 8 (a) and Figure 8 (b), where Figure 8 (a) The spatial difference map of flood depth in a flood-prone area under a 20-year Chicago-type (r=0.29) moving path 7 test rainstorm predicted by the convolutional long short-term memory network (ConvLSTM) model and simulated by the urban waterlogging model (SWMM+LISFLOOD-FP) (t=3h); Figure 8 (b) Correlation between the water depth predicted by the CNN model of each grid and the simulated water depth of the urban waterlogging model. The spatial flooding water depth predicted by the ConvLSTM model is close to the result of the urban waterlogging model, and the flood-prone locations are consistent. The maximum water depth difference between the two is small (basically less than 0.1m). The water depth predicted by the ConvLSTM model has a high correlation with the water depth simulated by the urban waterlogging model, showing a significant linear correlation (P<0.001). At t=3h, the correlation coefficient R predicted by the ConvLSTM model is 2 The average absolute error (MAE) is 0.94, the mean absolute error (MAE) is 0.028, and the mean square error (MSE) is 0.015. The results show that the difference between the water depth changes over time predicted by the ConvLSTM model and the simulated water depth changes by the urban waterlogging model is small, and the ConvLSTM model can better capture the trend of grid water depth changes over time.

[0101] Under the 15-year return path type 8 rainstorm, the difference and correlation (t = 3h) between the water depth predicted by the ConvLSTM model of each grid and the simulated water depth of the urban waterlogging model based on SWMM + LISFLOOD-FP are shown in Figure 2. Fig. 9 (a) and Fig. 9 (b), where Fig. 9 (a) The spatial difference map of flood depth in a flood-prone area under a 15-year Chicago-type (r=0.29) moving path 8 test rainstorm predicted by the convolutional long short-term memory network (ConvLSTM) model and simulated by the urban waterlogging model (SWMM+LISFLOOD-FP) (t=3h); Fig. 9 (b) Correlation between the water depth predicted by the CNN model of each grid and the simulated water depth of the urban waterlogging model. The spatial flooding water depth predicted by the ConvLSTM model is close to the result of the urban waterlogging model, and the flood-prone locations are consistent. The maximum water depth difference between the two is small (basically less than 0.1m). The water depth predicted by the ConvLSTM model has a high correlation with the water depth simulated by the urban waterlogging model, showing a significant linear correlation (P<0.001). At t=3h, the correlation coefficient R predicted by the ConvLSTM model is 2 The average absolute error (MAE) is 0.964, the mean absolute error (MAE) is 0.026, and the mean square error (MSE) is 0.019. The results show that the difference between the water depth changes over time predicted by the ConvLSTM model and the simulated waterlogging model is small, and the ConvLSTM model can better capture the trend of grid water depth changes over time.

[0102] Under the 10-year moving path type 5 torrential rain, the difference and correlation (t=3h) between the water depth predicted by the ConvLSTM model of each grid and the simulated water depth of the urban waterlogging model based on SWMM+LISFLOOD-FP are shown in Figure 2. Fig.10 (a) and Fig.10 (b), where Fig.10 (a) The spatial difference map of flood depth in a flood-prone area under a 10-year Chicago-type (r=0.29) moving path 5 test rainstorm predicted by the convolutional long short-term memory network (ConvLSTM) model and simulated by the urban waterlogging model (SWMM+LISFLOOD-FP) (t=3h); Fig.10 (b) Correlation between the water depth predicted by the CNN model of each grid and the simulated water depth of the urban waterlogging model. The spatial flooding water depth predicted by the ConvLSTM model is close to the result of the urban waterlogging model, and the flood-prone locations are consistent. The maximum water depth difference between the two is small (basically less than 0.1m). The water depth predicted by the ConvLSTM model has a high correlation with the water depth simulated by the urban waterlogging model, showing a significant linear correlation (P<0.001). At t=3h, the correlation coefficient R predicted by the ConvLSTM model is 2The average absolute error (MAE) is 0.969, the mean absolute error (MAE) is 0.020, and the mean square error (MSE) is 0.014. The results show that the difference between the water depth changes over time predicted by the ConvLSTM model and the simulated water depth changes by the urban waterlogging model is small, and the ConvLSTM model can better capture the trend of grid water depth changes over time.

[0103] Under the 2-year moving path type 4 torrential rain, the difference and correlation (t=3h) between the water depth predicted by the ConvLSTM model of each grid and the simulated water depth of the urban waterlogging model based on SWMM+LISFLOOD-FP are shown in Figure 2. Fig.11 (a) and Fig.11 (b), where Fig.11 (a) The spatial difference map of flood depth in a flood-prone area under a two-year Chicago-type (r=0.29) mobile path 4 test rainstorm predicted by the convolutional long short-term memory network (ConvLSTM) model and simulated by the urban waterlogging model (SWMM+LISFLOOD-FP) (t=3h); Fig.11 (b) Correlation between the water depth predicted by the CNN model of each grid and the simulated water depth of the urban waterlogging model. The spatial flooding water depth predicted by the ConvLSTM model is close to the result of the urban waterlogging model, and the flood-prone locations are consistent. The maximum water depth difference between the two is small (basically less than 0.1m). The water depth predicted by the ConvLSTM model has a high correlation with the water depth simulated by the urban waterlogging model, showing a significant linear correlation (P<0.001). At t=3h, the correlation coefficient R predicted by the ConvLSTM model is 2 The average absolute error (MAE) is 0.924, the mean absolute error (MAE) is 0.023, and the mean square error (MSE) is 0.010. The results show that the difference between the water depth predicted by the ConvLSTM model and the waterlogging model simulation over time is small, and the ConvLSTM model can better capture the trend of grid water depth over time.

[0104] Under the 6-type rainstorm with a return period of one year, the difference and correlation (t=3h) between the water depth predicted by the ConvLSTM model of each grid and the simulated water depth by the urban waterlogging model based on SWMM+LISFLOOD-FP are shown in Figure 2. Fig.12 (a) and Fig.12 (b) Fig.12 (a) is a spatial difference map of flood depths predicted by the Convolutional Long Short-Term Memory Network (ConvLSTM) model and simulated by the urban waterlogging model (SWMM+LISFLOOD-FP) in a flood-prone area of ​​the embodiment under a one-year Chicago-type (r=0.29) and moving path 6 test rainstorm (t=3h); Fig.12(b) Correlation between the water depth predicted by the CNN model and the simulated water depth of the urban waterlogging model for each grid. The spatial flooding water depth predicted by the ConvLSTM model is close to the result of the urban waterlogging model, and the flood-prone location coincides with the maximum water depth difference between the two is small (basically less than 0.1m). The water depth predicted by the ConvLSTM model has a high correlation with the water depth simulated by the urban waterlogging model, showing a significant linear correlation (P<0.001). At t=3h, the correlation coefficient R2 predicted by the ConvLSTM model is 0.925, the mean absolute error MAE is 0.022, and the mean square error MSE is 0.005. The results show that the difference between the water depth predicted by the ConvLSTM model and the simulated water depth of the urban waterlogging model over time is small, and the ConvLSTM model can better capture the trend of grid water depth over time.

[0105] From the perspective of simulation duration, the ConvLSTM model can complete the calculation of the spatial flooding depth and evolution of the study area (116.26km2) within 2 seconds, and can output hourly urban flooding inundation maps for the next 3 hours for each rainstorm. The urban flooding simulation based on SWMM+LISFLOOD-FP basically takes 1 hour. The computing efficiency of the former is more than 1,800 times higher than that of the latter.

[0106] Reference Figure 2 , an urban rainstorm water depth prediction system based on ConvLSTM model, including:

[0107] The first module 201 is used to collect data in the area to be simulated and construct an urban rain and flood model for the area to be simulated;

[0108] The second module 202 is used to simulate and predict the rainstorm event data information of the simulated area based on the urban rain and flood model of the simulated area and preprocess the data to obtain the normalized spatiotemporal characteristic data of rainfall in the simulated area;

[0109] The third module 203 is used to build a ConvLSTM model based on a convolutional neural network and a long short-term memory network;

[0110] The fourth module 204 is used to predict the water depth of the normalized spatiotemporal characteristic data of rainfall in the area to be simulated based on the ConvLSTM model to obtain the prediction result of the rainstorm water depth in the area to be simulated.

[0111] The contents of the above method embodiments are all applicable to the present system embodiments. The functions specifically implemented by the present system embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0112] The above is a specific description of the preferred implementation of the present invention, but the invention is not limited to the embodiments. Those skilled in the art may make various equivalent modifications or substitutions without violating the spirit of the present invention. These equivalent modifications or substitutions are all included in the scope defined by the claims of this application.

Claims

1. A method for predicting urban rainstorm water depth based on ConvLSTM model, characterized in that: The following steps are involved: Collect data for the area to be simulated and build an urban rain and flood model for the area to be simulated; Based on the urban rain and flood model of the area to be simulated, the rainstorm event data information of the area to be simulated is simulated and predicted and data preprocessed to obtain the normalized spatiotemporal characteristic data of rainfall in the area to be simulated; Construct the ConvLSTM model based on convolutional neural network and long short-term memory network; Based on the ConvLSTM model, the water depth prediction of the normalized spatiotemporal characteristic data of rainfall in the simulated area is carried out to obtain the prediction result of the heavy rain water depth in the simulated area.

2. According to the method for predicting urban rainstorm depth based on ConvLSTM model in claim 1, it is characterized in that: The step of collecting data from the area to be simulated and constructing an urban rain and flood model for the area to be simulated specifically includes: Obtain the land use type, elevation and drainage network data of the area to be simulated, and construct a one-dimensional hydrological model of the area to be simulated; Obtain rainstorm data information of the area to be simulated and input it into the one-dimensional hydrological model of the area to be simulated for data derivation, so as to obtain the position coordinates and overflow information of the overflow pipe point under the rainfall conditions of the area to be simulated; Acquire the elevation information of the area to be simulated and perform a two-dimensional evolution simulation based on the position coordinates and overflow information of the overflow pipe point under the rainfall conditions in the area to be simulated, and construct a two-dimensional hydrodynamic model of the area to be simulated; The one-dimensional hydrological model of the area to be simulated is coupled with the two-dimensional hydrodynamic model of the area to be simulated to construct an urban stormwater model of the area to be simulated.

3. According to the method for predicting urban rainstorm depth based on ConvLSTM model in claim 2, it is characterized in that: The step of performing simulation prediction and data preprocessing on the rainstorm event data information of the simulated area based on the urban rain and flood model of the simulated area to obtain normalized spatiotemporal characteristic data of rainfall in the simulated area specifically includes: Obtain data information on rainstorm events with different return periods in the simulated area and with different moving directions and paths of the rainstorm center; The spatially variable rainfall model based on a fixed spatial structure is used to scale the rainstorm event data information to obtain rainfall information with different spatial and temporal characteristics in the simulated area. Inputting rainfall information with different spatiotemporal characteristics of the area to be simulated into the urban rain and flood model of the area to be simulated to perform rainfall simulation prediction, and obtaining the predicted spatiotemporal characteristic data of rainfall in the area to be simulated, wherein the predicted spatiotemporal characteristic data of rainfall in the area to be simulated includes the flooding range, flooding depth and hourly evolution process of flooding under different spatiotemporal rainstorm scenarios in the area to be simulated; The predicted spatiotemporal characteristic data of rainfall in the area to be simulated are normalized to obtain normalized spatiotemporal characteristic data of rainfall in the area to be simulated.

4. According to claim 3, a method for predicting urban rainstorm depth based on ConvLSTM model, characterized in that: The expression of the spatially variable rainfall model based on a fixed spatial structure is specifically as follows: R(x,y,μ1,μ2,σ,t)=γ(μ1,μ2)f(x,y,μ1,μ2,σ)r(t) In the above formula, R(x,y,μ1,μ2,σ,t) represents the rainfall intensity at position (x,y) at time t in the spatially variable rainfall model with fixed spatial structure, (μ1,μ2) represents the position coordinates of the rainfall center, r(t) represents the rainfall intensity at time t under spatially uniform rainfall conditions, A represents the basin area, f(x,y,μ1,μ2,σ) is the distribution function of spatial rainfall intensity represented by a two-dimensional truncated Gaussian distribution function, σ represents the standard deviation of the Gaussian distribution, and γ(μ1,μ2) represents the scaling function of the two-dimensional truncated Gaussian distribution.

5. According to claim 4, a method for predicting urban rainstorm depth based on ConvLSTM model, characterized in that: The expression for normalizing the spatial and temporal characteristic data of the rainfall in the predicted area to be simulated is specifically as follows: In the above formula, X n represents the normalized temporal and spatial characteristic data of rainfall in the simulated area, X represents the sequence value, and X min Indicates the minimum value of the sequence value, X max Indicates the maximum value of the sequence.

6. According to claim 5, a method for predicting urban rainstorm depth based on ConvLSTM model is characterized in that: The ConvLSTM model specifically includes a two-dimensional convolutional long short-term memory network layer, a two-dimensional convolutional neural network layer and a fully connected neural network layer, wherein the output end of the two-dimensional convolutional long short-term memory network layer is connected to the input end of the two-dimensional convolutional neural network layer, and the output end of the two-dimensional convolutional neural network layer is connected to the input end of the fully connected neural network layer.

7. The method for predicting urban rainstorm depth based on ConvLSTM model according to claim 6, characterized in that: The step of performing water depth prediction on the normalized spatiotemporal characteristic data of rainfall in the area to be simulated based on the ConvLSTM model to obtain the prediction result of the rainstorm water depth in the area to be simulated specifically includes: The normalized spatiotemporal characteristic data of rainfall in the area to be simulated is input into the ConvLSTM model; Based on the two-dimensional convolutional long short-term memory network layer of the ConvLSTM model, the normalized spatiotemporal characteristic data of rainfall in the area to be simulated are processed to capture the spatiotemporal characteristics and obtain the spatiotemporal characteristic map of rainfall in the area to be simulated. Based on the two-dimensional convolutional neural network layer of the ConvLSTM model, the spatial information feature extraction process is performed on the spatiotemporal feature map of rainfall in the simulated area to obtain the spatiotemporal feature map of rainfall in the simulated area after convolution; Based on the fully connected neural network layer of the ConvLSTM model, the spatiotemporal characteristic map of rainfall in the simulated area after convolution is processed by feature integration to obtain the prediction result of heavy rain depth in the simulated area.

8. According to claim 7, a method for predicting urban rainstorm depth based on ConvLSTM model, characterized in that: The expression of the two-dimensional convolutional long short-term memory network layer based on the ConvLSTM model is specifically as follows: In the above formula, σ represents the sigmoid activation function, tanh(·) represents the hyperbolic tangent activation function, and x t represents the input of the current time step, h t-1 represents the hidden state of the previous time step, c t-1 represents the memory unit of the previous time step, f t 、i t 、c t , o t They represent the forget gate, input gate, memory unit and output gate respectively, * represents the convolution operator, represents the Hadamard product.

9. The method for predicting urban rainstorm depth based on ConvLSTM model according to claim 8, characterized in that: The expression of the two-dimensional convolutional neural network layer based on the ConvLSTM model is specifically as follows: In the above formula, represents the input of the kth neuron in the lth convolution layer, represents the bias of the kth neuron in the lth convolution layer, represents the output of the i-th neuron in the (l-1)th convolution layer, represents the convolution kernel from the i-th neuron in the (l-1)-th layer of convolution to the k-th neuron in the l-th layer of convolution, and Conv2D(·) represents a two-dimensional convolution operation.

10. A city rainstorm water depth prediction system based on ConvLSTM model, characterized in that: Includes the following modules: The first module is used to collect data in the area to be simulated and build an urban rain and flood model for the area to be simulated; The second module is used to simulate and predict the rainstorm event data information of the simulated area based on the urban rain and flood model of the simulated area and preprocess the data to obtain the normalized spatiotemporal characteristic data of rainfall in the simulated area; The third module is used to build a ConvLSTM model based on convolutional neural network and long short-term memory network; The fourth module is used to predict the water depth of the normalized spatiotemporal characteristic data of rainfall in the simulated area based on the ConvLSTM model to obtain the prediction result of the heavy rain water depth in the simulated area.

Images