A rainwater system simulation method for urban waterlogging early warning

By applying deep learning technology in urban flooding early warning, establishing a simulation model for rainfall production and pipeline network convergence processes, the problem of high cognitive requirements for water physical processes in the existing technology is solved, and efficient and accurate under complex pipeline structures is achieved Water accumulation warning prediction provides a flexible and timely warning system.

CN114139433BActive Publication Date: 2024-12-06TONGJI UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202110862977.4
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-07-29
Publication Date
2024-12-06
Estimated Expiration
2041-07-29

AI Technical Summary

Technical Problem

The existing technology has high cognition requirements for hydrological physical processes in urban flooding warning prediction, and lacks deep learning-based stormwater system simulation methods, making it difficult to effectively apply to water accumulation warning prediction under complex pipeline structures.

Method used

Using the deep learning-based rainwater system simulation method, a simulation model for rainfall production flow and a one-dimensional pipeline convergence process are established. Through feasibility analysis, data set design and model construction, the external inflow flow of the pipeline network and the node water accumulation flow are simulated. , provide real-time warning prediction.

Benefits of technology

It realizes efficient and accurate water accumulation warning and prediction under complex pipeline structures, which is flexible and timely, can handle large data volume, is widely applicable, and the model can be continuously updated to improve training accuracy.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114139433B_ABST
    Figure CN114139433B_ABST
Patent Text Reader

Abstract

The present invention provides a rainwater system simulation method for urban waterlogging early warning, which is based on deep learning and consists of two deep learning models: a rainfall runoff process simulation model and a one-dimensional pipe network confluence process simulation model. The rainfall runoff process simulation model is used to simulate the external inflow of the pipe network, such as sub-catchment area runoff and rainfall infiltration, and serves as the boundary condition for the training of the one-dimensional pipe network confluence process simulation model; the one-dimensional pipe network confluence process simulation model is used to simulate the water accumulation flow at the pipe network node, and at the same time has the function of comparing the network architecture of the deep learning model. The modeling of the present invention is simple, highly flexible and time-effective, and the model can be updated by supplementing data with high accuracy; it can realize the simulation of the water accumulation process under pipe network structures of different complexities, and has broad application prospects.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of urban rainwater system simulation, and relates to a technical method for real-time early warning and prediction of urban waterlogging, particularly a technical method based on deep learning theory involving simulation of rainfall formation and its runoff generation process and one-dimensional pipe network confluence process. Background Art

[0002] With the continuous acceleration of urbanization, urban waterlogging has attracted more and more attention because it seriously threatens the safety of life and property of residents and affects the normal operation of cities. How to alleviate urban waterlogging and reduce the losses caused by waterlogging has become an urgent problem for municipal departments to solve. At the same time, the continuous development of modern scientific and technological means such as the Internet of Things, mobile Internet, big data analysis and intelligent algorithms has provided reliable technical support for the prevention and control of urban waterlogging, making scientific research fields related to urban drainage technology such as real-time monitoring and systematic management of urban drainage pipe networks and research on urban waterlogging early warning and prediction technologies gradually receive attention.

[0003] The earliest research on urban waterlogging warning and prediction technology was based on the urban stormwater management model platform. The collected topographic maps, terrain information, drainage network data, historical rainfall data and other information in the study area were used to build models in the urban stormwater management model software, and subsequent urban waterlogging simulation and warning prediction in the study area were carried out on the basis of the modeling. This method started early, has in-depth research, and has mature practical applications, but its application has always been constrained by problems such as limited knowledge of hydrology and hydrodynamics. The development of cutting-edge technology deep learning architecture has made it possible to apply prediction methods based on time series data analysis in the field of hydrology. Therefore, how to apply data-driven models that meet low requirements for the cognition of hydrological physical processes to urban waterlogging warning and prediction research is one of the challenges faced in scientific research and practical applications.

[0004] In the field of urban drainage system research, some researchers have introduced deep learning theory to study the prediction of system sewage flow. However, there are few rainwater system simulation methods based on deep learning theory specifically for urban waterlogging warning prediction in the currently reported literature and patents. Therefore, there is an urgent need to design and construct corresponding methods for research. Summary of the invention

[0005] In view of the above problems, the purpose of the present invention is to provide a rainwater system simulation method based on deep learning for urban waterlogging early warning and prediction. The method has simple modeling, strong practicality, high efficiency and accuracy, and is suitable for early warning and prediction of urban waterlogging under various pipe network structures with different degrees of complexity.

[0006] The present invention provides the following technical solutions: a rainwater system simulation method for urban waterlogging early warning, the method is based on deep learning, and the method establishes two deep learning models, namely a rainfall runoff process simulation model and a one-dimensional pipe network confluence process simulation model; wherein the rainfall runoff process simulation model is used to simulate the external inflow of the pipe network such as sub-catchment area runoff and rainfall infiltration, and serves as the boundary condition for training the one-dimensional pipe network confluence process simulation model; the one-dimensional pipe network confluence process simulation model is used to simulate the water accumulation flow at the pipe network node.

[0007] Furthermore, the rainfall runoff simulation model includes a feasibility analysis module, a data set design module, and a model building module;

[0008] The feasibility analysis module is used to perform feasibility analysis on the external inflow of the pipe network node based on rainfall;

[0009] The data set design module is used to design various data such as rainfall and simulated hydraulic data required for constructing input and output models;

[0010] The model building module is used to design the deep learning model structure, train the model and perform evaluation.

[0011] Furthermore, the one-dimensional pipe network confluence process simulation model includes a model building module and a model evaluation module;

[0012] The model building module is used to design the model input matrix form, select the deep learning model architecture, design the model hyperparameters, and perform model training;

[0013] The model evaluation module is used to evaluate the model effect and compare the model architecture.

[0014] The beneficial effects of the present invention are:

[0015] 1. The rainwater system simulation method based on deep learning theory provided by the present invention can simulate the external inflow of the pipeline network in the rainfall runoff process simulation model, and simulate the water accumulation flow at the pipeline network node in the one-dimensional pipeline network confluence process simulation model, while also having the deep learning model network architecture comparison function.

[0016] 2. The rainwater system simulation method provided by the present invention is applied to the research on real-time early warning and prediction of urban waterlogging. It has the characteristics of simple modeling, flexibility and practicality, high efficiency and accuracy, and wide applicability. Compared with flood risk maps or flood control maps with only a limited number of calculation examples, this method has higher flexibility. Compared with the method based on urban rainwater model, this method is more efficient and faster. Hydraulic model simulation requires higher real-time computing resources, while this method can separate real-time computing resources to obtain higher timeliness.

[0017] 3. The rainwater system simulation method provided by the present invention establishes a deep learning network framework based on rainfall data and external inflow data of the pipe network, thereby establishing a unified pipe network waterlogging early warning and prediction system, laying the foundation for processing increasing amounts of data and more and more data features.

[0018] 4. The rainwater system simulation method provided by the present invention compares several deep learning network models, analyzes the deep learning network architecture that is more in line with the rainwater system simulation characteristics and can mine the deep characteristics of the data, and provides data support for the applicability of the deep learning network in the rainwater system simulation problem.

[0019] 5. The rainwater system simulation method provided by the present invention can continuously supplement new rainfall data to update the model during the model training process, thereby improving the training accuracy of the model.

[0020] 6. The rainwater system simulation method provided by the present invention is based on a data-driven model, and the complexity of the rainwater system pipe network model has limited influence on it, that is, it can realize the simulation of the water accumulation process under different pipe network structures, thereby expanding the scope of application of the method. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] The present invention will be described in more detail below based on embodiments and with reference to the accompanying drawings, wherein:

[0022] Figure 1 A schematic diagram of the flow chart of the rainwater system simulation method provided by the present invention;

[0023] Figure 2 This is the hydraulic model diagram of the TJU pipe network;

[0024] Figure 3 It is a comparison chart of the correlation between rainfall intensity and external inflow flow at node 34;

[0025] Figure 4 This is the structure diagram of the LSTM model for simulating the rainfall runoff process;

[0026] In the figure: LSTM: long short-term memory neural network; RMSE: root mean square error; MAE: mean absolute error; CC: Pearson correlation coefficient; NSE: Nash efficiency coefficient.

[0027] Figure 5 This is a comparison chart of the test results of the LSTM model simulating the rainfall runoff process and the actual results;

[0028] in, Figure 5 (A): Training sample of external inflow of waterlogging at node 32; Figure 5 (B) Node 32 external inbound traffic test sample.

[0029] Figure 6This is the structure diagram of the LSTM model for simulating the one-dimensional pipe network confluence process;

[0030] Figure 7 This is a comparison chart of the test results and actual results of the LSTM model simulation of the one-dimensional pipe network confluence process;

[0031] in, Figure 7 (A): Training sample of water flow at node 16; Figure 7 (B): Node 42 water flow test sample.

[0032] Figure 8 This is a comparison chart between the test results and the actual results of the MLP model simulation of the one-dimensional pipe network confluence process;

[0033] in, Figure 8 (A): Training sample of water flow at node 16; Figure 8 (B): Node 42 water flow test sample.

[0034] Fig. 9 This is a comparison chart of the test results of the CNN model simulating the one-dimensional pipe network confluence process and the actual results;

[0035] in, Fig. 9 (A): Training sample of water flow at node 16; Fig. 9 (B): Node 42 water flow test sample. DETAILED DESCRIPTION

[0036] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0037] Example 1

[0038] The present embodiment provides a rainwater system simulation method for urban waterlogging early warning, which is based on deep learning. The method establishes two deep learning models, namely a rainfall runoff process simulation model and a one-dimensional pipe network confluence process simulation model. The rainfall runoff process simulation model is used to simulate external inflows of the pipe network, such as sub-catchment area runoff and rainfall infiltration, and serves as a boundary condition for training the one-dimensional pipe network confluence process simulation model. The one-dimensional pipe network confluence process simulation model is used to simulate the water accumulation flow at the pipe network nodes.

[0039] The rainfall runoff process simulation model includes a feasibility analysis module, a data set design module, and a model building module;

[0040] Feasibility analysis module, used to perform feasibility analysis on the external inflow of pipe network nodes based on rainfall;

[0041] The data set design module is used to design various data such as rainfall and simulated hydraulic data required to construct input and output models;

[0042] The model building module is used to design the deep learning model structure, train the model and evaluate it.

[0043] One-dimensional pipe network confluence process simulation model, including model building module and model evaluation module;

[0044] The model building module is used to design the model input matrix form, select the deep learning model architecture, design the model hyperparameters, and perform model training;

[0045] The model evaluation module is used to evaluate the model effect and compare the model architecture.

[0046] Example 2

[0047] See also Figure 1 and Figure 2 Taking the TJU pipe network as an example, the present invention illustrates the implementation of urban waterlogging warning using the model constructed by the rainwater system simulation method of Example 1.

[0048] During implementation, when establishing a simulation model for the rainfall runoff process, a feasibility analysis should be conducted first:

[0049] In the first step, the input and output of the runoff model are determined as rainfall and node external inflow.

[0050] The second step is to analyze the correlation between rainfall and external inflow at the node, see Figure 3 :Use the time dimension as the feature dimension and calculate the rainfall data R(s,t) (sample × time) and the external inflow data L of a single node respectively n The covariance matrix of (s, t) (sample × time) is obtained by principal component analysis (PCA) on the two covariance matrices to obtain new R'(s) (sample) and L'(s) = {L1'(s)...L n '(s)...L N '(s)}(samples) are combined into a new feature matrix F(s) = {R'(s), L'(s)}, and its correlation coefficient matrix C(N+1) is calculated. From the correlation coefficient analysis, it can be seen that the correlation coefficient between rainfall and the external inflow of each node can generally reach above 0.9, which shows that the runoff model established by rainfall and the external inflow of each node is feasible.

[0051] Then design the data set:

[0052] The first step is to construct rainfall data: Step 1: Taking Shanghai as an example, refer to the "Shanghai Heavy Rain Intensity Formula and Design Rainfall Type" and local historical rainfall data to determine the construction parameters of the return period and other rainfall types, and construct single-peak and double-peak rainfall scenarios; Step 2: Construct a random Gaussian white noise of the same length as the rainfall data; Step 3: Divide the constructed rainfall data into three sections before and after the rain peak position from the time dimension at 50% of the peak rainfall size, and arrange the constructed noise according to the principle of adding positive noise to the first and last sections and negative noise to the middle section; Step 4: Add processed Gaussian white noise with a noise percentage characteristic of 0% to 100% to the constructed rainfall data.

[0053] In the second step, the existing hydraulic model is used for one-dimensional simulation to obtain the total lateral inflow of each node and the accumulated water flow data of the node under the corresponding rainfall events.

[0054] Then the model is built:

[0055] The first step is to determine the input and output of the model: with a general understanding of the hydrological mechanism process, the input of the model is determined to be the rainfall data R(t) of a certain rainfall event, and the output is the total side inflow L of the node under the rainfall event. n (t).

[0056] The second step is to design the model structure. Figure 4 , the model architecture is determined to be 2 layers of LSTM + 1 layer of fully connected output, the model parameters include the number of hidden layer units 16, batch size BATCHSIZE = 64, learning rate LR = 0.001, optimizer Adam, loss function MSELoss, EPOCH = 200, the data set is normalized by Z-score for model training and testing of the model effect, see Figure 5 The model evaluation results are shown in Table 1:

[0057] Table 1 Evaluation results of rainfall runoff simulation model

[0058]

[0059] Then, in the stage of establishing a one-dimensional pipe network confluence process simulation model, the first step is to build the model:

[0060] The first step is to determine the model input and output: with a general understanding of the hydrodynamic process, the input of the model is determined to be the total lateral inflow L(t) of each node in the pipeline network under a certain rainfall event, and the output is the accumulated water flow P(t) of each node in the pipeline network under this rainfall event.

[0061] The second step is to determine the matrix form of the model input: Step 1: Combined with the cross-correlation analysis, calculate the lag time from the entry of rainfall runoff into the pipe network to the occurrence of waterlogging at the node, that is, the time delay value T(N,N) between the two time series of input quantity I(N,T) and output quantity O(N,T).

[0062]

[0063]

[0064]

[0065] Step 2: According to the delay value T(n) of each node corresponding to the input quantity in T(N,N) i ,N), determine the delay interval of the node as ΔT(i);

[0066] T(n i ,N)=[t(n i ,n1) ... t(n i ,n j ) ... t(n i ,n N )] T (4)

[0067] ΔT(i)=[t1 i …t j i ... t n i ] (5)

[0068] Step 3: Determine the input vector form M(N,t) of the model at time t:

[0069] M(N,t)=[m(n1,t) ... m(n i ,t) ... m(n N ,t)] (6)

[0070] m(n i ,t)=[I(n i ,tt n i ) ... I(n i ,tt j i ) ... I(n i ,t-t1 i )] (7)

[0071] The third step is to select the model architecture: design the model structure under the three network architectures of LSTM, MLP, and CNN respectively. Figure 6The LSTM model architecture is 3 layers of LSTM + 2 layers of fully connected output; the MLP model architecture is 5 layers of fully connected output; the CNN model architecture includes 3 layers of CNN + 2 layers of fully connected output. After the data set is divided, it is normalized by Z-score and the model is trained;

[0072] The fourth step is to design model hyperparameters. Step 1: Taking the LSTM model as an example, select the model hyperparameters that need to be optimized, including the learning rate LR, batch size BATCHSIZE, and the number of hidden layer units, and set the range of each hyperparameter to establish a hyperparameter set:

[0073] learning rate LR = [0.001 0.002 0.005 0.01 0.02 0.05 0.1];

[0074] Batch size BATCHSIZE = [8 16 64 128 256];

[0075] Number of hidden layer units = [8 16 64 128 256];

[0076] Step 2: Take the average of the loss function values ​​of the model under three trainings as the model training result under a set of hyperparameter combinations; Step 3: Repeat the test and use the hyperparameter optimization library Hyperopt to select the optimal model hyperparameter combination.

[0077] The final LSTM model hyperparameter combination is determined as follows: learning rate LR = 0.001, batch size BATCHSIZE = 64, and the number of hidden layer units is 256 and 128. The MLP model and CNN model parameters are adjusted based on the LSTM model training results: the model's learning rate LR and batch size BATCHSIZE remain unchanged, the number of hidden layer units of the MLP model is 256, 256, 256, 256, 128, and the number of hidden layer units of the CNN model is 8, 32, 64, 256, 128.

[0078] Then comes the evaluation of the model:

[0079] The first step is to establish the evaluation system of the model: determine the evaluation indicators of the model as RMSE, MAE, CC and NSE, RMSE and MAE verify whether the neural network training converges, CC and NSE evaluate the accuracy of the model prediction.

[0080] The second step is to compare the three network architectures of LSTM, MLP and CNN under the established model evaluation system. Figure 7 , Figure 8 , Fig. 9 The model evaluation results are shown in Table 2:

[0081] Table 2 Evaluation results of the one-dimensional pipe network confluence process simulation model

[0082]

[0083]

[0084] Judging from various indicators, the training effects of the three network architectures are LSTM>CNN>MLP, and it is finally determined that the LSTM model is more suitable for rainwater system simulation problems.

[0085] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that the technical solutions described in the above embodiments may still be modified, or some or all of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the scope of the technical solutions of the embodiments of the present invention.

[0086] In addition, those skilled in the art will appreciate that, although some embodiments herein include certain features included in other embodiments but not other features, the combination of features of different embodiments is meant to be within the scope of the present invention and to form different embodiments. For example, in the above claims, any one of the claimed embodiments may be used in any combination. The information disclosed in this background technology section is intended only to deepen the understanding of the overall background technology of the present invention and should not be regarded as an admission or in any form of implication that the information constitutes prior art known to those skilled in the art.

Claims

1. A rainwater system simulation method for urban waterlogging early warning, the method is based on deep learning, It is characterized in that The method establishes two deep learning models, namely a rainfall runoff process simulation model and a one-dimensional pipe network confluence process simulation model; wherein the rainfall runoff process simulation model is used to simulate the external inflow of the pipe network including sub-catchment area runoff and rainfall infiltration, and serves as the boundary condition for training the one-dimensional pipe network confluence process simulation model; the one-dimensional pipe network confluence process simulation model is used to simulate the water accumulation flow at the pipe network node; The rainfall runoff simulation model includes a feasibility analysis module, a data set design module, and a model building module; The feasibility analysis module is used to perform feasibility analysis on the external inflow of the pipe network node based on rainfall; The data set design module is used to design the rainfall and simulated hydraulic data required for constructing the input and output model; The model building module is used to design the deep learning model structure, train the model and evaluate it; The one-dimensional pipe network confluence process simulation model includes a model building module and a model evaluation module; The model building module is used to design the model input matrix form, select the deep learning model architecture, design the model hyperparameters, and perform model training; The model evaluation module is used to evaluate the model effect and compare the model architecture; The method specifically comprises: In the stage of establishing the rainfall runoff simulation model, the feasibility analysis of calculating the external inflow of the pipe network nodes based on rainfall is first carried out. Specifically: The first step is to determine the input and output of the rainfall runoff simulation model as rainfall and node external inflow respectively; The second step is to analyze the correlation between rainfall and node external inflow. Specifically, the correlation coefficient between rainfall and node external inflow is calculated. If the correlation coefficient is greater than the preset threshold, it indicates that it is feasible to establish a rainfall runoff process simulation model based on rainfall and node external inflow. After that, the rainfall data, node external inflow data and node flood flow data required for the model are constructed. Specifically: The first step is to construct rainfall data. Specifically, step 1: refer to the rainfall reference materials and historical rainfall data of the city to be simulated, determine various rain type construction parameters, construct single and double peak rainfall scenarios, and construct rainfall data for multiple rainfall sessions; step 2: construct random Gaussian white noise of equal length to the rainfall data; step 3: divide the constructed rainfall data into three sections before and after the rain peak position from the time dimension at 50% of the peak value of rainfall, and arrange the constructed noise according to the principle of adding positive noise to the first and last sections and negative noise to the middle section; step 4: add processed Gaussian white noise with a noise percentage characteristic of 0% to 100% to the constructed rainfall data; In the second step, a one-dimensional simulation is performed using the hydraulic model to obtain the total side inflow data of each node and the water accumulation flow data of the node under the corresponding rainfall events; The next step is to build a simulation model for the rainfall runoff process: The first step is to determine the input and output of the model: the input of the rainfall runoff simulation model is determined as the rainfall data R(t) of a certain event, and the output is the total side inflow data L of the node under the rainfall event. n (t); The second step is to design the model architecture: determine the model architecture as 2-layer LSTM + 1-layer fully connected output, the model parameters include the number of hidden layer units 16, batch size BATCHSIZE = 64, learning rate LR = 0.001, optimizer Adam, loss function MSELoss, EPOCH = 200; The third step is to perform Z-score standardization on the dataset including rainfall data under multiple rainfall events and total side inflow data of each node, and use the Z-score standardized dataset to train the designed model and test the model effect; After that, in the stage of establishing a one-dimensional pipe network confluence process simulation model, a one-dimensional pipe network confluence process simulation model is first built, specifically: The first step is to determine the input and output of the one-dimensional pipe network confluence process simulation model: the input of the model is determined to be the total side inflow data L(t) of each node in the pipe network under a certain rainfall event, and the output is the accumulated water flow data P(t) of each node in the pipe network under the rainfall event; The second step is to determine the matrix form of the model input: Step 1: Combined with the cross-correlation analysis, calculate the lag time from the time when the rainfall runoff enters the pipe network to the time when water accumulates at the node, that is, the time delay value T(N,N) between the two time series of input quantity I(N,T) and output quantity O(N,T); Step 2: According to the delay value T(n) of each node corresponding to the input quantity in T(N,N) i ,N), determine the delay interval of the node as ΔT(i); T(n i ,N)=[t(n i ,n 1 )...t(n i ,n j )...t(n i ,n N )] T ; ΔT(i)=[t 1 i …t j i …t n i ]; Step 3: Determine the input vector form M(N,t) of the model at time t: M(N,t)=[m(n 1 ,t)…m(n i ,t)…m(n N ,t)]; m(n i ,t)=[I(n i ,t-t n i )…I(n i ,t-t j i )…I(n i ,t-t 1 i )]; The third step is to select the model architecture: design the model architecture under the three network architectures of LSTM, MLP, and CNN respectively. The LSTM model architecture is 3 layers of LSTM + 2 layers of fully connected output; the MLP model architecture is 5 layers of fully connected output; the CNN model architecture includes 3 layers of CNN + 2 layers of fully connected output; The fourth step is to calculate and assign the optimal hyperparameter combination of the three model architectures: LSTM, MLP, and CNN; The fifth step is to perform Z-score standardization on the data set including the total side inflow data of each node under multiple rainfall events and the water accumulation flow data of each node. The Z-score standardized data set is used to train the assigned model and evaluate the model effect. The LSTM with the best effect is selected as the one-dimensional pipe network confluence process simulation model. When using the trained rainfall runoff process simulation model and the one-dimensional pipe network confluence process simulation model to simulate the rainwater system, the rainfall data under the rainfall event in the city to be simulated is input into the rainfall runoff process simulation model, which is processed by the rainfall runoff process simulation model and outputs the total lateral inflow data of each node under the rainfall event. Then, the total lateral inflow data of each node under the rainfall event is input into the one-dimensional pipe network confluence process simulation model, which is processed by the one-dimensional pipe network confluence process simulation model and outputs the accumulated water flow data of each node under the rainfall event.