A fast prediction method for urban waterlogging distribution based on improved convolutional neural network

Abstract

The present invention relates to a method for rapid prediction of urban waterlogging distribution based on an improved convolutional neural network, comprising: 1. constructing a rapid prediction model data set: the data set is a waterlogging influencing factor-waterlogging characterization factor data set, and the waterlogging influencing factors include rainfall factors, terrain factors, pipe network length factors, and distance factors from the river system; 2. building a rapid prediction model based on a convolutional neural network and performing training and testing. In the process of building the model, different forms of waterlogging influencing factors are first encoded separately, then fused, and finally decoded after fusion; in the process of training and testing, an improved loss function is used; 3. model prediction and model prediction result analysis. This method accurately realizes the direct conversion of rainfall and flood elements (flooding distribution, maximum flooding depth), and improves the speed and accuracy of urban flood forecasting.

Description

Technical Field

[0001] The present invention belongs to the field of urban flood control and smart water conservancy, and in particular relates to a method for quickly predicting urban waterlogging distribution based on an improved convolutional neural network. Background Art

[0002] With climate change and urbanization, extreme heavy precipitation events are frequent, and the sensitivity and vulnerability of cities to heavy precipitation events have increased significantly, resulting in cities facing severe waterlogging risks. Therefore, a rapid prediction method for urban waterlogging waterlogging distribution is needed to improve my country's urban waterlogging prevention and disaster reduction capabilities and effectively reduce urban flood disasters. Using real-time monitoring of short-term rainfall data, urban waterlogging depth and waterlogging distribution can be quickly predicted, and urban waterlogging risk areas can be assessed, so that measures can be taken quickly to alleviate the losses caused by waterlogging. In the future, the rapid prediction system for urban waterlogging waterlogging distribution can be combined with online weather forecast services to notify citizens in waterlogging risk areas in advance to take preventive measures. With the development of computer technology, two-dimensional hydrological / hydrodynamic models have been widely used in urban hydrological processes and flood dynamics simulations, but numerical simulation models based on physical mechanisms still require a long computing time and are difficult to meet the needs of rapid prediction of urban waterlogging. This problem is more prominent when numerical models are applied to areas with complex infrastructure and urban pipe networks. This emphasizes the necessity of developing rapid flood prediction tools that can be used for urban flood risk management and urban planning. Although artificial intelligence methods are gradually being applied in flood prediction, most current studies only consider the relationship between rainfall and inundation, without fully considering the influencing factors of urban waterlogging, resulting in low waterlogging prediction accuracy and poor visualization. In addition, the loss function used in traditional convolutional neural network-based regression prediction tasks is not conducive to accurately predicting the distribution and depth of urban waterlogging. Therefore, it is necessary to design a new loss function to improve the accuracy of rapid urban waterlogging prediction. Summary of the invention

[0003] The purpose of the present invention is to overcome the shortcomings of the prior art and to provide a method for rapid prediction of urban waterlogging distribution based on an improved convolutional neural network, which can more accurately realize the direct conversion of rainfall and flood elements (flood distribution, maximum flood depth) and improve the speed and accuracy of urban flood forecasting.

[0004] The above-mentioned object of the present invention is achieved by the following technical solutions:

[0005] A method for rapid prediction of urban waterlogging distribution based on an improved convolutional neural network, characterized in that it comprises the following steps:

[0006] Step 1: Construct a rapid prediction model data set: This data set is a waterlogging influencing factor-waterlogging characterization factor data set. Waterlogging influencing factors include rainfall factor, terrain factor, pipe network length factor, and distance factor from the river system. The rainfall factor includes designed rainfall and measured rainfall. Select the maximum flooding depth as the waterlogging characterization factor, use PCSWMM to build an urban waterlogging numerical simulation model, simulate the flooding of the application area under different rainfall conditions, and obtain the waterlogging characterization factor data set.

[0007] Step 2: Building a fast prediction model based on convolutional neural network includes:

[0008] 2.1. Sample the designed rainfall and measured rainfall data in step 1 every five minutes;

[0009] 2.2. Process the image data of waterlogging influencing factors and waterlogging characterization factors into a unified rectangular grid format and minimize them to the application area, wherein the image data of waterlogging influencing factors include terrain factor image data, pipe network length factor image data and distance factor image data from the river system, and the waterlogging characterization factor data is the maximum flooding depth data. The waterlogging influencing factor data and waterlogging characterization factor data in the grid are extracted based on the orthogonal grid and divided into a training set and a test set. A multi-channel encoder is used to perform multi-channel image encoding for different surface features of waterlogging influencing factors, and the information after the prediction image encoding is decoded by a decoder to construct a fast prediction model;

[0010] 2.3. Train and test the constructed model;

[0011] In the process of training and testing, an improved loss function is used. Specifically, the loss function is designed from the two aspects of waterlogging data set imbalance and information loss. In view of the situation that the area of ​​the flooded area is much smaller than the area of ​​the non-flooded area, the root mean square error is used as the basis to provide a greater weight for the error loss in the area with a larger flood depth, so as to compensate for the problem of reduced prediction accuracy caused by the relative lack of non-flooded and deep flooded waterlogging data. The loss function of this part is designed as follows:

[0012]

[0013] Where: n is the number of grids; y i Predict the value of the i-th grid for the model; Simulate the value of the i-th grid for the numerical model;

[0014] The perceptual loss function is introduced as the second part of the loss function to solve the problem of inaccurate prediction of the maximum flooding depth extreme value due to excessive smoothing and loss of image pixels caused by minimizing the average loss of the loss function; the expression of the perceptual loss function is as follows:

[0015]

[0016] Where: is the expected distribution of the difference between the predicted value and the simulated value; w, h, d are the width, height and depth of the feature image respectively; ||·||2 is the Euclidean distance.

[0017] Step 3: Perform model prediction and analyze the model prediction results: the evaluation indicators used include Nash efficiency coefficient NSE, peak signal-to-noise ratio PSNR, and maximum flooding depth extreme value relative error Δy p , the performance of the model is evaluated in terms of robustness and generalization.

[0018] Further: In step 1, the rainfall factor is based on the historical rainfall data of the application area, with the Huff rainfall type and the Chicago rainfall type commonly used in China as the basic rainfall type, and a total of 105 design rainfalls with different recurrence periods and rainfall durations in the area are obtained. At the same time, 9 typical rainfalls in the application area are selected as the measured rainfall.

[0019] Further: the surface features of different waterlogging influencing factors in step 2.2 include waterlogging influencing factor features in image form and waterlogging influencing factor features in tensor form; the waterlogging influencing factors in image form include terrain factors, pipe network length factors, and distance factors from the river system; the multi-channel encoder used by the waterlogging influencing factors in image form is a deep neural network structure containing six layers of convolution, arranged in a pyramidal manner; the input size of the waterlogging influencing factors in image form is 2048×2048×5, and after six layers of convolution and three layers of pooling operations, the size of the waterlogging influencing factors in image form after encoding is updated to 32×32×256; the waterlogging influencing factors in tensor form are rainfall factors; the tensor form The encoding of waterlogging impact factors is realized through a feedforward neural network. The feedforward neural network uses neurons in the hidden layer to output the rainfall sequence as a z×1 vector in the fully connected layer and reshape it into an m×n×k structure. The rainfall sequence is a 36-dimensional vector before processing. The size of the waterlogging impact factor in the form of the encoded tensor is 32×32×256. Based on this, the fully connected layer is set to 2048 nodes, and the rainfall sequence is finally reshaped into a 32×32×2 structure; the feature decoding of the waterlogging impact factor after fusion is a deep neural network structure containing six deconvolution layers, arranged in an inverted pyramid; all convolution layers use SAME padding to keep the size of the output feature map unchanged and the activation function is the RELU function.

[0020] The present invention has the advantages and positive effects:

[0021] 1. When constructing a rapid prediction model data set, the present invention not only uses rainfall factors as waterlogging influencing factors, but also includes terrain factors, pipe network length factors, and distance factors from the river system. Compared with the prior art, the present invention fully considers many influencing factors of urban waterlogging, which is conducive to more accurate identification of waterlogging distribution areas and improves the accuracy and visualization of waterlogging prediction.

[0022] 2. When building a rapid prediction model, the present invention first encodes different forms of waterlogging influencing factors separately, then fuses them, and finally decodes the fused factors to establish a mapping relationship between rainfall, topography, pipe network length, distance from the river system and maximum flooding depth.

[0023] 3. During the training and testing process, the present invention adopts an improved loss function to make up for the problem of reduced prediction accuracy caused by the relative lack of non-flooded and deeply flooded water accumulation data, and also improves the accuracy of the prediction of the maximum flooded water depth extreme value.

[0024] In summary, the rapid prediction method for urban waterlogging waterlogging distribution based on improved convolutional neural network proposed in the present invention can quickly and timely predict the urban waterlogging area and the maximum waterlogging depth according to real-time monitoring of rainstorm information. Compared with the prediction results of traditional convolutional neural networks, it has the advantages of higher prediction accuracy of urban waterlogging waterlogging distribution, faster response speed and more accurate prediction, and has great application value in urban flood prevention and disaster reduction and smart city risk emergency management. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 This is a framework diagram of rapid prediction of urban waterlogging distribution based on improved convolutional neural network of the present invention;

[0026] Figure 2a is a schematic diagram of prediction results of an embodiment of the present invention;

[0027] Figure 2b It is a schematic diagram of the measured results of the embodiment of the present invention. DETAILED DESCRIPTION

[0028] The structure of the present invention is further described below with reference to the accompanying drawings and by way of examples. It should be noted that the present examples are descriptive rather than restrictive.

[0029] A fast prediction method for urban waterlogging distribution based on improved convolutional neural network, see Figure 1-2bThe invention is to establish the relationship between the influencing factors of urban waterlogging and the characterization factors of urban waterlogging through numerical simulation, statistical analysis and deep learning methods, and to quickly predict the spatial distribution of urban waterlogging and the water depth of urban waterlogging. This rapid prediction method mainly includes: the construction of a rapid prediction model data set, the construction and training of a rapid prediction model, the design of a loss function, model prediction, and the analysis of model prediction results.

[0030] Before establishing a rapid prediction model, it is necessary to construct a waterlogging influencing factor-waterlogging characterization factor data set, and use this data set as a training-testing database for improving convolutional neural networks. Urban waterlogging is often affected by multiple factors. The present invention selects rainfall factors, terrain factors, pipe network length factors, and distance factors from the river system as waterlogging influencing factors. The rainfall factor selects four types of Huff rain types and Chicago rain types as design rainfall rain types based on the historical rainfall data of the application area. According to the dimensionless rainfall time distribution cumulative curves of these five design rainfall types, the design rainfall distribution in each time interval is obtained, and the design rainfall at different rainfall durations under the return period of 2, 3, 5, 10, 20, 50, and 100 years is multiplied by the rain type distribution, that is, the design rainstorm process with different rainfall durations under different return periods in the study area is obtained, and 105 design rainfalls are obtained. At the same time, 9 typical rainfall processes are selected to form a rainfall database. Taking the rapid forecast of waterlogging in Xinbadali, Tianjin as an example, the terrain, pipe network and river distribution data of Xinbadali, Tianjin were processed based on the ArcGIS platform to obtain the terrain factor, pipe network factor and distance factor data from the river system. The maximum flooding depth was selected as the waterlogging characterization factor, and the PCSWMM was used to build an urban waterlogging numerical simulation model to simulate the flooding of the application area under different rainfall conditions to obtain the waterlogging characterization factor data set. After obtaining the waterlogging characterization factor data set, in order to improve the efficiency and accuracy of modeling, the variance inflation factor and XGboost algorithm were used to screen the waterlogging influencing factors to determine the waterlogging influencing factors that were finally involved in the modeling.

[0031] The construction and training of the rapid prediction model can be broken down into three main steps. First, the above-mentioned designed rainfall and measured rainfall data are sampled every five minutes; then, all raster layers are processed to minimize coverage of the entire application area, and the waterlogging influencing factor data and waterlogging characterization data in the grid are extracted based on the orthogonal grid and divided into training sets and test sets. The encoder is used to perform multi-channel image encoding for different surface features, and the decoder is used to decode the information encoded by the predicted image. Finally, the constructed model is trained and tested.

[0032] The above-mentioned different surface features include waterlogging impact factor features in image form and waterlogging impact factor features in tensor form. Specifically, the waterlogging impact factor features in image form and the waterlogging impact factor features in tensor form are encoded respectively, the two forms of encoding are fused at the joint layer, and the combined waterlogging impact factor features are subjected to multi-channel data decoding. Among them, the waterlogging impact factor in image form includes terrain factors, pipe network length factors, and distance factors from the river system. The waterlogging impact factor multi-channel encoder is a deep neural network structure containing six layers of convolution, arranged in a pyramid arrangement. The input size of the waterlogging impact factor is 2048×2048×5. After six layers of convolution and three layers of pooling operations, the size of the waterlogging impact factor after encoding is updated to 32×32×256. The waterlogging influencing factor in tensor form is the rainfall factor. The feature encoding is realized by a feedforward neural network. The feedforward neural network uses neurons in the hidden layer to output the rainfall sequence as a suitable z×1 vector in the fully connected layer, and then reshapes it into an m×n×k structure. The rainfall sequence is a 36-dimensional vector before processing. The size of the encoded image waterlogging influencing factor is 32×32×256. Based on this, the fully connected layer is set to 2048 nodes, and the rainfall sequence is finally reshaped into a 32×32×2 structure. The feature decoding of the waterlogging influencing factor after fusion is a deep neural network structure containing six deconvolution layers, arranged in an inverted pyramid. All convolutional layers use SAME padding to keep the size of the output feature map unchanged and the activation function is the RELU function.

[0033] The model training process is to optimize a certain defined loss function. A suitable loss function helps to improve the accuracy of model prediction. The present invention designs the loss function from two aspects: imbalance of waterlogging data set and information loss. In view of the problem that the area of ​​the flooded area is much smaller than the area of ​​the non-flooded area, based on the root mean square error, a greater weight is given to the error loss in the area with a larger flooding depth, to make up for the problem of reduced prediction accuracy caused by the relative lack of non-flooded and deep flooded waterlogging data. This part of the loss function is designed as follows:

[0034]

[0035] Where: n is the number of grids; y i Predict the value of the i-th grid for the model; Simulate the value of the ith grid for the numerical model

[0036] Since the traditional loss functions MSE and MAE only consider the difference between the predicted maximum flooding depth value and the actual maximum flooding depth value in model training, the predicted image will be over-smoothed due to the loss function minimizing the average loss, resulting in inaccurate prediction of the maximum flooding depth extreme value. The present invention introduces the perceptual loss function as the second part of the loss function to solve the problem of inaccurate prediction of the maximum flooding depth extreme value caused by the loss function minimizing the average loss and over-smoothing, resulting in loss of image pixels. The expression of the perceptual loss function is as follows:

[0037]

[0038] Where: is the expected distribution of the difference between the predicted value and the simulated value; w, h, d are the width, height and depth of the feature image respectively; ||·||2 is the Euclidean distance.

[0039] The present invention adopts Nash efficiency coefficient (NSE), peak signal-to-noise ratio (PSNR), maximum submergence depth extreme value relative error Δy p The performance of different prediction models is evaluated from the aspects of robustness and generalization by using evaluation indicators such as Δy. In summary, compared with the convolutional neural network model with standard MSE as the loss function, the NSE value of the prediction result of the model proposed in the present invention increases from 0.82 to 0.9, an increase of 9.8%; the PSNR calculation result also increases from 29.8 to 34.7, an increase of 16.4%; Δy p The index decreased by 21.4%, as shown in Table 1, which shows that the convolutional neural network model with improved loss function has a higher accuracy in predicting extreme values. In summary, the rapid prediction model of urban waterlogging distribution based on improved convolutional neural network has good reliability.

[0040] Table 1 Model prediction performance evaluation results

[0041]

[0042] Although the embodiments and drawings of the present invention are disclosed for illustrative purposes, those skilled in the art will appreciate that various substitutions, changes and modifications are possible without departing from the spirit and scope of the present invention and the appended claims. Therefore, the scope of the present invention is not limited to the contents disclosed in the embodiments and drawings.

Claims

1. A method for rapid prediction of urban waterlogging distribution based on improved convolutional neural network, characterized in that: The steps include: Step 1: Construct a rapid prediction model data set: This data set is a waterlogging influencing factor-waterlogging characterization factor data set. Waterlogging influencing factors include rainfall factor, terrain factor, pipe network length factor, and distance factor from the river system. The rainfall factor includes designed rainfall and measured rainfall. Select the maximum flooding depth as the waterlogging characterization factor, use PCSWMM to build an urban waterlogging numerical simulation model, simulate the flooding of the application area under different rainfall conditions, and obtain the waterlogging characterization factor data set. Step 2: Building a fast prediction model based on convolutional neural network includes: 2.

1. Sample the designed rainfall and measured rainfall data in step 1 every five minutes; 2.

2. Process the image data of waterlogging influencing factors and waterlogging characterization factors into a unified rectangular grid format and minimize them to the application area, wherein the image data of waterlogging influencing factors include terrain factor image data, pipe network length factor image data and distance factor image data from the river system. Extract the waterlogging influencing factor data and waterlogging characterization factor data in the grid based on the orthogonal grid and divide them into a training set and a test set. Use a multi-channel encoder to perform multi-channel image encoding for different surface features of waterlogging influencing factors, decode the information after the prediction image is encoded by a decoder, and construct a fast prediction model. 2.

3. Train and test the constructed model; In the process of training and testing, an improved loss function is used. Specifically, the loss function is designed from the two aspects of waterlogging data set imbalance and information loss. In view of the situation that the area of ​​the flooded area is much smaller than the area of ​​the non-flooded area, the root mean square error is used as the basis to provide a greater weight for the error loss in the area with a larger flood depth, so as to compensate for the problem of reduced prediction accuracy caused by the relative lack of non-flooded and deep flooded waterlogging data. The loss function of this part is designed as follows: Where: n is the number of grids; y i Predict the value of the i-th grid for the model; Simulate the value of the i-th grid for the numerical model; The perceptual loss function is introduced as the second part of the loss function to solve the problem of inaccurate prediction of the maximum flooding depth extreme value due to excessive smoothing and loss of image pixels caused by minimizing the average loss of the loss function; the expression of the perceptual loss function is as follows: Where: is the expected distribution of the difference between the predicted value and the simulated value; w, h, d are the width, height and depth of the feature image respectively; ||·||2 is the Euclidean distance; Step 3: Perform model prediction and analyze the model prediction results: the evaluation indicators used include Nash efficiency coefficient NSE, peak signal-to-noise ratio PSNR, and maximum flooding depth extreme value relative error Δy p , the performance of the model is evaluated in terms of robustness and generalization.

2. The method for rapid prediction of urban waterlogging distribution based on improved convolutional neural network according to claim 1 is characterized in that: In step 1, the rainfall factor is based on the historical rainfall data of the application area. The Huff rainfall type and the Chicago rainfall type commonly used in China are used as the basic rainfall types. A total of 105 design rainfalls with different recurrence periods and rainfall durations in the area are obtained. At the same time, 9 typical rainfalls in the application area are selected as the measured rainfall.

3. The method for rapid prediction of urban waterlogging distribution based on improved convolutional neural network according to claim 1 is characterized in that: The surface features of different waterlogging influencing factors in step 2.2 include waterlogging influencing factor features in image form and waterlogging influencing factor features in tensor form; the waterlogging influencing factors in image form include terrain factors, pipe network length factors, and distance factors from the river system; the multi-channel encoder used by the waterlogging influencing factors in image form is a deep neural network structure containing six layers of convolution, arranged in a pyramidal manner; the input size of the waterlogging influencing factors in image form is 2048×2048×5, and after six layers of convolution and three layers of pooling operations, the size of the waterlogging influencing factors in image form after encoding is updated to 32×32×256; the waterlogging influencing factors in tensor form are rainfall factors; The encoding of waterlogging impact factors is realized through a feedforward neural network. The feedforward neural network uses neurons in the hidden layer to output the rainfall sequence as a z×1 vector in the fully connected layer and reshape it into an m×n×k structure. The rainfall sequence is a 36-dimensional vector before processing. The size of the waterlogging impact factor in the form of the encoded tensor is 32×32×256. Based on this, the fully connected layer is set to 2048 nodes, and the rainfall sequence is finally reshaped into a 32×32×2 structure; the feature decoding of the waterlogging impact factor after fusion is a deep neural network structure containing six deconvolution layers, arranged in an inverted pyramid; all convolution layers use SAME padding to keep the size of the output feature map unchanged and the activation function is the RELU function.

Images