Urban flood depth learning prediction method considering disaster environment clustering

Abstract

The application provides a kind of urban flood depth learning prediction method considering disaster environment clustering, steps include, S1 establishes the urban flood one-dimensional and two-dimensional coupling model based on PCSWMM, the regional inundation depth under different rainfall-tide combination is simulated using the model.S2 based on rainfall data, tide data, disaster environment data and inundation depth data to build a deep learning database.S3 using K-means clustering algorithm, according to the disaster environment characteristic index, the region is divided into different categories.S4 combined with disaster environment clustering and the cause of each category, select different input variables, establish different deep learning models, and apply bayesian algorithm to optimize the hyperparameters of deep learning model, complete the rapid prediction of urban flood.The method considers the clustering of disaster environment, and uses bayesian algorithm to optimize the deep learning model, which has strong reliability and effectively improves the prediction accuracy of urban flood.

Description

Technical Field

[0001] This invention relates to the field of hydrological prediction technology, specifically to a deep learning method for predicting urban flooding that considers disaster-prone environmental clustering. Background Technology

[0002] With global climate change and rapid urbanization, the frequency of urban flooding has increased significantly, causing enormous losses to people's lives and property. Due to the combined effects of heavy rainfall and high tides, coastal cities are more prone to flooding, and because their populations and property are concentrated in one area, the losses caused by floods are even more severe.

[0003] Traditional hydrodynamic models, while capable of predicting urban flooding, typically require sufficient data to calibrate their parameters, and simulating a flood can be time-consuming. In recent years, deep learning models have been widely applied in rapid urban flood prediction. They eliminate the need for complex model building, providing inundation data in just tens or even seconds, greatly aiding urban flood control and rapid response. However, deep learning training demands vast amounts of training data, such as observed rainfall-inundation depth data, which is scarce in urban areas, particularly in China, limiting the application of deep learning models. To address these issues, researchers have combined physical models with deep learning models. The physical model (i.e., the hydrodynamic model) simulates inundation depths under different rainfall conditions, providing sufficient training data for the deep learning model, while the deep learning model solves the problem of the long training time associated with physically driven models.

[0004] Besides the amount of training data affecting deep learning models, the hyperparameters within those models also influence prediction accuracy. For instance, optimizing hyperparameters in a Support Vector Machine (SVM) using a grid search method can reduce model error; similarly, optimizing a Long Short-Term Memory (LSTM) model using Particle Swarm Optimization (PSO) can improve model accuracy. However, traditional hyperparameter optimization methods suffer from drawbacks such as localized optimization and long optimization times.

[0005] Furthermore, the selection of input variables is a crucial step in building deep learning models. The input variables for deep learning models are generally determined by disaster-causing factors (such as rainfall) and disaster-prone environments (such as elevation and slope). Different disaster-prone environments lead to different causes of flooding in different urban areas. For example, some areas are flooded due to their low elevation, while others near rivers are flooded due to overflowing water. Existing deep learning models rarely consider the physical mechanisms of urban flooding (such as the causes of flooding), relying entirely on data-driven approaches, which may affect the accuracy and interpretability of the models. Summary of the Invention

[0006] This invention provides a deep learning method for predicting urban flooding that considers disaster-prone environmental clustering, aiming to at least solve one of the problems of existing technologies.

[0007] This invention combines deep learning models with disaster-prone environment clustering to propose a rapid urban flood prediction model. It uses the K-means clustering algorithm to cluster disaster-prone environments and then uses the Bayesian algorithm to optimize the hyperparameters of the deep learning model.

[0008] A deep learning-based method for predicting urban flooding that considers disaster-prone environmental clustering includes the following steps: S1 uses PCSWMM to establish a one- or two-dimensional coupled hydrological and hydrodynamic model, and uses this model to simulate the regional inundation depth under different rainfall-tide combinations.

[0009] S2 constructs a deep learning database based on rainfall data, tide data, disaster-prone environment data, and inundation depth data.

[0010] S3 uses the K-means clustering algorithm to divide the region into different categories based on disaster-prone environmental characteristics.

[0011] S4 selects different input variables based on different disaster-prone environmental categories and their causes of flooding, establishes different deep learning models, and applies the Bayesian algorithm to optimize the hyperparameters of the deep learning models to complete urban flood prediction.

[0012] In S1, a one-dimensional hydrodynamic model is established using river channel and pipeline network data; a two-dimensional surface inundation model is established using elevation data. The one-dimensional and two-dimensional models are connected through orifices, realizing the flow interaction between the one-dimensional hydrodynamic model and the two-dimensional surface inundation model. The data used to construct the PCSWMM model includes digital elevation model (DEM) data, pipeline network data, river network data, and structures. The pipeline network data includes pipes and manholes, and the river network data includes river cross-sections and cross-sectional spacing. The PCSWMM model constructed in this invention includes one-dimensional pipelines, one-dimensional manholes, sub-catchments, outlets, and a two-dimensional grid. This model is used to simulate the inundation depth of the invention area under different conditions.

[0013] In S2, design rainfall-tide data for different return periods in the invention area are calculated and loaded into the PCSWMM model to obtain inundation depth data for the invention area under different conditions. Design rainfall and tide data are determined by the Generalized Extreme Value (GEV) function, and the GEV function parameters are estimated using the maximum likelihood method. Then, the design rainfall and tide for multiple return periods are calculated using the GEV function. Considering the most severe flooding scenario, rainfall-tide events occurring under extreme weather conditions in the invention area are taken as typical rainfall-tide processes, and the same-scale amplification method is used to obtain design rainfall and tide processes for multiple return periods. In this invention, paired combinations of rainfall and tide processes for multiple return periods can obtain multiple rainfall-tide events. The data from these events are used as input conditions for the PCSWMM model to obtain inundation depths under different conditions. Disaster-prone environmental data for the invention area are obtained using basic data and a Geographic Information System (GIS). A deep learning database is constructed using rainfall data, tide data, disaster-prone environmental data, and inundation depth data.

[0014] In S3, the K-means clustering algorithm first determines the number of clusters. This invention uses SSE to determine the optimal number of clusters. The SSE function takes the value of the cluster number *k* at the inflection point as the optimal number of clusters. The K-means clustering algorithm randomly selects... K Using a sample of data points, and based on the similarity metric, all remaining data points are sequentially assigned to the subset with the highest similarity. Next, new cluster centers are recalculated. If the new cluster centers remain unchanged, the clustering process is complete; otherwise, the cluster centers are recalculated until they no longer change. The K-means clustering algorithm is used to divide the invention area into different categories based on the disaster-prone environment.

[0015] Based on the different disaster-prone environment characteristics of the grid with the optimal number of clusters obtained in S3, S4 establishes the corresponding GBDT model, namely the K-means-GBDT model. For each category of K-means-GBDT model, appropriate input variables are selected, and hyperparameters that have a greater impact on the deep learning model are selected. Then, the corresponding optimization range and interval are set, and the hyperparameters of the model are optimized using the Bayesian algorithm.

[0016] S41 selects different input variables based on the disaster-prone environment of different grid categories to build corresponding GBDT models, i.e., K-means-GBDT models. Then, appropriate input variables are selected for each category of K-means-GBDT model. The model in this step is built on the Scikit-Learn library in the Python 3.9 environment, and the dataset is divided into training and test sets in an approximately 7:3 ratio.

[0017] S42 Due to the varying amounts of data and input variables in different deep learning models, the optimal hyperparameters for each class of the K-means-GBDT model also differ. This model has many hyperparameters, but six are generally considered to have a significant impact: loss function, number of iterations, learning rate, subsampling ratio, maximum depth, and minimum impurity decrease. In this invention, these six hyperparameters are set within a specific optimization range, and a 10-fold cross-validation method is used to evaluate the model's training performance. RMSE is used to represent the cross-validation results. The Bayesian optimization is implemented using the Hyperopt library in Python 3.9.

[0018] As a preferred approach, the performance of the model is finally evaluated using absolute error, mean square error, root mean square error, and Nash efficiency coefficient. The model is then compared with the model in this invention using an unclassified optimized model to explore the accuracy of the model in this invention.

[0019] The present invention also provides a deep learning prediction system for urban flooding that considers disaster-prone environmental clustering, including a processor and a memory; the processor and the memory are communicatively connected, and the processor is used to read a computer program from the memory and execute it to implement the aforementioned method.

[0020] The beneficial effects of this invention are as follows: A comprehensive method combining K-means clustering and Bayesian optimization of a deep learning model is proposed to predict flood depth. This method considers the clustering of disaster-prone environments and uses a Bayesian algorithm to optimize the deep learning model, effectively improving prediction accuracy. The proposed model exhibits high flood prediction accuracy, significantly improving upon the unclassified GBDT model. This model demonstrates high reliability and can be used for rapid urban flood prediction. Attached Figure Description

[0021] To more clearly illustrate the technical solution of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 The framework of this invention; Figure 2 Location of the invention area; Figure 3Disaster-prone environmental factors: (a) elevation, (b) slope, (c) pipeline density, (d) distance from the river channel, (e) curvature; Figure 4 Rainfall and tide levels during Typhoon Rammasun; Figure 5 Design rainfall events with a return period of 5-100 years; Figure 6 The design tide level process with a recurrence period of 5-100 years; Figure 7 Maximum inundation depth under the combination of 100-year design rainfall and 100-year design tide level; Figure 8 Different number of clusters K SSE value; Figure 9 Different categories of grid distribution; Figure 10 GBDT model schematic diagram; Figure 11 The error of the maximum flood depth simulated by the K-means-GBDT model under the conditions of (a) R50-T50; (b) R50-T100; (c) R100-T5; (d) R100-T10; (e) R100-T20; (f) R100-T50; (g) R100-T100; Figure 12 The differences in maximum flooding depth among 50 sample grids in categories (a) 1, (b) 2, and (c) 3 simulated by the three models under the R100-T100 condition; Figure 13 The difference between the observed depth and the predicted depth at 8 observation points. Detailed Implementation

[0023] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0024] The following uses Haidian Island, located in the northern part of Haikou City, Hainan Province, China, as an example to introduce the deep learning prediction method for urban flooding that considers disaster-prone environment clustering, as described in this invention. The framework of this invention is as follows: Figure 1 As shown.

[0025] To make the method of the present invention clearer, Haidian Island and its data will be introduced first.

[0026] See Figure 2Haidian Island is adjacent to the Qiongzhou Strait. Due to its relatively flat terrain and unique geographical location, it is easily affected by rainfall and tides. During heavy rainfall, floodwaters usually flow into the pipe network first, and then into the Qiongzhou Strait through urban rivers. However, when extreme rainfall and high tides occur simultaneously, the rising tides prevent the floodwaters in the pipe network and rivers from being discharged in time, causing huge losses to the region.

[0027] The data consists of three parts. The first part is used to construct the PCSWMM model and includes digital elevation model (DEM) data, pipe network data, river network data, and structures. The second part is observational data of historical flood events used to calibrate the PCSWMM model. The final part is disaster-prone environment data, including DEM, slope (SL), pipe network density (DON), distance from the river channel (DR), and curvature (CU).

[0028] To establish a deep learning model for flood prediction in the invention area, this invention selected five disaster-prone environmental factors as part of the input variables for the deep learning model. These variables are as follows: Digital Elevation Model (DEM, m): The DEM reflects the topographic conditions of the invention area. It is the foundational data for the one-dimensional / two-dimensional coupled model and is also one of the disaster-prone environmental data. Provided by the Institute of Geographic Sciences and Natural Resources Research, Chinese Academy of Sciences. See [link / reference]. Figure 3 a.

[0029] Slope (SL, °): SL reflects the degree of topographic change and affects water flow direction, a major cause of flooding. It is obtained by analyzing DEM data using a Geographic Information System (GIS). See [link to GIS documentation]. Figure 3 b.

[0030] Pipeline density (DON, m / m) 2 ): DON represents the density of the underground drainage network in this area. In areas with low network density, water cannot be discharged in a timely manner, leading to flooding. This density is obtained through GIS density analysis of the network data; see [link / reference]. Figure 3 c.

[0031] Distance from the river channel (DR, m): DR is the distance from each grid cell to the river channel. Areas closer to the river may be more prone to flooding due to dam overflows or breaches, while areas farther from the river are safer. This distance is obtained through GIS distance analysis; see [link to GIS documentation]. Figure 3 d.

[0032] Curvature (CU): CU is a morphological measurement index of local terrain, describing changes in slope. Positive curvature indicates that the surface of a pixel is convex upwards; negative curvature indicates that the surface of a pixel is concave downwards; 0 indicates a flat surface. During downhill water flow, curvature affects the convergence or divergence of water; negative curvature tends to accumulate more water, increasing the risk of flooding. Calculated using GIS curvature analysis; see [link to relevant documentation]. Figure 3 e.

[0033] Specifically, the present invention provides a deep learning method for predicting urban flooding that considers disaster-prone environmental clustering, comprising the following steps: S1 establishes a one- or two-dimensional coupled hydrological and hydrodynamic model based on the PCSWMM model, and uses this model to simulate the regional inundation depth under different rainfall-tide combinations.

[0034] A one-dimensional hydrodynamic model was established using river channel and pipeline data; a two-dimensional surface inundation model was established using elevation data. The one-dimensional and two-dimensional models were connected by orifices, realizing the flow interaction between the one-dimensional hydrodynamic model and the two-dimensional surface inundation model. The PCSWMM model constructed in this invention includes 2042 one-dimensional pipes, 2075 one-dimensional manholes, 13 sub-catchments, 13 outlets, and 23580 two-dimensional grids.

[0035] The data used to construct the PCSWMM model includes a digital elevation model (DEM), pipe network data, river network data, and structures. The runoff calculation in the model uses the Manning formula:

[0036] In the formula For flow rate, m 3 / s; Width of the sub-catchment area, in meters; The Manning coefficient for the Earth's surface; The maximum depth of the depression is in meters. This represents the average slope of the sub-catchment area.

[0037] The Horton model is used as the infiltration model, and the infiltration formula is:

[0038] In the formula for Permeability at any given time, mm / h; To stabilize the infiltration rate, mm / h; The initial soil infiltration rate is expressed in mm / h. This is the infiltration attenuation coefficient, which is related to the soil physical properties. ; It refers to time, h.

[0039] The dynamic wave method is used in the flow transmission of pipelines and river networks. Its governing equations include the continuity equation and the momentum equation, which are described as follows:

[0040]

[0041] In the formula For flow rate, m 3 / s; The cross-sectional area of ​​the water passage is m. 2 t is time, s; It is the distance, in meters (m). The water depth is in meters (m). It is the acceleration due to gravity, m / s² 2 ; It is the friction slope.

[0042] To correct the PCSWMM model, the actual rainfall and tide events caused by Typhoon Rammasun on July 18, 2014 (see [link to PCSWMM model]) were used. Figure 4 Input model. Table 1 shows the data for each observation point (see...). Figure 2 The difference between the simulated maximum inundation depth and the measured maximum inundation depth was calculated. Since the maximum inundation depth simulated by the PCSWMM model is greatly affected by the accuracy of the topographic data and cannot completely reflect the inundation situation, the absolute error between the simulated and measured maximum inundation depths is less than 0.1m, and the relative error is also within the allowable range, indicating that the simulation results are reasonable. Therefore, it can be considered that the simulation results of the model are in good agreement with the actual situation, verifying the reliability of the constructed PCSWMM model.

[0043] Table 1. Differences between simulated and measured values

[0044] S2 constructs a deep learning database based on rainfall data, tide data, disaster-prone environment data, and inundation depth data.

[0045] The designed rainfall and tidal data were determined by the generalized extreme value (GEV) function, and the GEV function parameters were estimated using the maximum likelihood method. Location parameters were also considered in the rainfall distribution. = 122.51, scale parameter = 46.78, shape parameter =0.115. Location parameter in tidal distribution. = 2.349, scale parameter = 0.380, shape parameter = -0.048.

[0046] Then, the design rainfall and tidal levels for five return periods within a 5-100 year timeframe were calculated using the GEV function. Considering the most severe flooding scenario, the rainfall-tidal event of July 18, 2014, was taken as a typical rainfall-tidal process. The design rainfall and tidal processes for the 5-100 year return periods were obtained using the same scale amplification method. (See [link to relevant documentation]). Figure 5 and Figure 6 As shown.

[0047] In this invention, a pairwise combination of rainfall and tide levels with 5 return periods can be used to obtain 25 rainfall-tide events. The data of these events are used as input conditions for the PCSWMM model to simulate the inundation depth under different conditions.

[0048] The region was divided into 23,580 grids, but through multiple flood simulations, it was found that some grids were almost unaffected by flooding. To reduce computational complexity and more effectively predict inundation conditions in the invention area, this invention uses flood-prone grids as the prediction targets. When the region experienced extreme weather conditions combining the 100-year design rainfall and the 100-year design tide level (R100-T100), it experienced the most unfavorable inundation conditions. The maximum inundation depth of the grids in the invention area under these conditions was obtained using the PCSWMM model; see [link to PCSWMM model]. Figure 7 Grids with a maximum inundation depth greater than 0.15m under these conditions are considered objectively flood-susceptible grids. Therefore, the invention contains a total of 13,157 target grids. 328,925 inundation depth data points were obtained under 25 rainfall-tide events and used to establish a deep learning database.

[0049] The deep learning model database consists of rainfall data, tide data, disaster-prone environment data, and inundation depth data.

[0050] S3 uses the K-means clustering algorithm to divide the invention area into different categories based on the environmental characteristics of the disaster-prone environment.

[0051] First, the number of categories is determined. This invention uses SSE to determine the optimal number of clusters.

[0052] The formula for SSE is as follows:

[0053] In the formula It is the first Clusters, p A sample, The average of all data, K Indicates the number of clusters.

[0054] The K-means clustering algorithm should be randomly selected. KUsing a sample of data as an example, based on the similarity measure, all other data are sequentially assigned to the subset of data with the highest similarity. Next, new cluster centers are recalculated. If the new cluster centers do not change, the clustering process is complete; otherwise, the cluster centers are recalculated until they no longer change.

[0055] In this invention, Euclidean distance is used to measure the similarity of samples.

[0056] Suppose the dataset is classified as: Divide the dataset into K Clusters, each Include Data, cluster center is .

[0057]

[0058] The sum of the distances from all data points in the data to the cluster centers is:

[0059] A smaller value indicates less difference between data points within the same category; our goal is to make... .

[0060] For the distance between different categories, the distance should be maximized, and the distance function between different categories should be weighted.

[0061] For dataset X The mean of all data points in the dataset. It is a subset Our goal is to make maximum.

[0062] After calculation, the SSE values ​​for different cluster numbers are as follows: Figure 8 As shown. In Previously, SSE dropped rapidly, in Afterward, the decrease in SSE leveled off. The SSE function's performance in terms of cluster number... The inflection point occurs when the cluster number is 3, therefore the cluster number... This is the optimal solution.

[0063] Figure 9The grid distribution represents different clusters. The first type of cluster, located in the city center near rivers at relatively low elevations, is where the pipe network is concentrated. During extreme rainfall and high tides, the low elevation and poor drainage capacity of the pipe network and rivers can cause flooding. The second type of cluster is located on the city's outskirts at the highest elevations. Because these grids are close to the ocean, they are more susceptible to flooding due to high tides. The third type of cluster is located within the city, near rivers at low elevations, making it prone to flood accumulation and severe inundation.

[0064] Based on the different disaster-prone environment characteristics of the grid with the optimal number of clusters obtained in S3, S4 establishes the corresponding GBDT model, namely the K-means-GBDT model. For each category of K-means-GBDT model, appropriate input variables are selected, and hyperparameters that have a greater impact on the deep learning model are selected. Then, the corresponding optimization range and interval are set, and the hyperparameters of the model are optimized using the Bayesian algorithm.

[0065] S41 selects different input variables based on the disaster-prone environment of different grids, constructs K-means-GBDT models for different disaster-prone environments, and predicts the inundation depth of the region.

[0066] Since the three clusters of grids generated by the K-means clustering algorithm have different disaster-prone environmental characteristics, this invention establishes three different K-means-GBDT models based on these different attributes. Appropriate input variables are selected for each K-means-GBDT model according to the disaster-prone environment of the three grid categories.

[0067] The input variables include: The relatively low altitude and close river distances resulted in similar flood causes for the first and third types of clustered grids. Therefore, the input variables for these two types of K-means-GBDT models are consistent: 5 disaster-inducing environmental factors (DEM, SL, DON, DR, and CU), 3 rainfall factors (total rainfall, maximum 3-hour rainfall, and peak rainfall), and 3 tidal factors (peak tidal level, mean tidal level, and maximum 3-hour mean tidal level).

[0068] The second type of clustered grid is located near the ocean and is greatly affected by tides. To highlight this characteristic, this invention adds a "catastrophic effect" variable to the input variables of the K-means-GBDT model for this type. "." "It is an indicator used to quantify the degree of harm caused by rainfall and tides in coastal cities."

[0069] To more intuitively reflect the influence of tide level, this invention incorporates the following into the final calculation: Adjustments have been made:

[0070] In the formula This represents the degree of danger posed by the tide level. This indicates the level of danger posed by the rainfall.

[0071] Therefore, the input variables for this type of K-means-GBDT are five disaster-prone environmental factors, " "3 rainfall factors and 3 tidal factors.

[0072] The output variable of the K-means-GBDT models for all three categories is the maximum flood depth.

[0073] The datasets were designed for different combinations of rainfall-tide events, with 117,575 data points for the first model, 91,950 for the second, and 119,400 for the third. The datasets were split into training and test sets in approximately a 7:3 ratio. These models were built using the Scikit-Learn library in a Python 3.9 environment.

[0074] The structure diagram of the GBDT model is as follows: Figure 10 As shown, the steps of the GBDT algorithm are as follows: (1) First, initialize the weak learner:

[0075] (2) For t=1, 2, 3...., T: (a) For each sample Calculate the negative gradient:

[0076] (3) Take the value obtained in the previous step as the new true value, and then... As the next new decision tree The corresponding leaf node region is , , in Let T be the number of leaf nodes in decision tree T.

[0077] (4) Calculate the best-fit value, i.e., the residual, for each leaf node region:

[0078] (d) Update the model as follows:

[0079] (5) The models from each iteration are superimposed to obtain the final model, and the final expression is:

[0080] S42 Bayesian algorithm optimizes the hyperparameters of deep learning models.

[0081] Because the amount of data and input variables differ among the three types of deep learning models, the optimal hyperparameters for each category of the K-means-GBDT model also differ. The GBDT model has many hyperparameters, but it is generally believed that six hyperparameters have a significant impact on the model: loss function, number of iterations, learning rate, subsampling ratio, maximum depth, and minimum impurity decrease.

[0082] In this invention, the six hyperparameters are set within a certain optimization range (see Table 2), and the hyperparameter space is 6-dimensional, with a total of 273,600 hyperparameter combinations. This invention uses 10-fold cross-validation to evaluate the model's training performance and uses RMSE to represent the cross-validation results. Bayesian optimization is implemented using the Hyperopt library in Python 3.9.

[0083] Table 2 Optimization range of the 6 hyperparameters

[0084] The steps for Bayesian optimization of hyperparameters in a deep learning model are as follows: (1) Assume that the hyperparameters of Bayesian optimization are black-box functions:

[0085] Let A be the optimal hyperparameter set and A be the candidate hyperparameter space.

[0086] (2) Take a combination of several hyperparameters in the candidate hyperparameter space and find the observations in these cases.

[0087] (3) Based on a finite number of observations, we hypothesize the distribution of the objective function; this is called prior knowledge in Bayesian optimization. The tool for making assumptions about the function distribution is the probabilistic surrogate model. These probabilistic surrogate models typically come with assumptions that allow them to estimate the distribution of the objective function from a very small number of observations. These models are usually based on Gaussian processes (GPs), which can be specified by their mean function and covariance matrix function. We can then obtain the objective value of the hypothesized function.

[0088] (4) Use the acquisition function to select the next most "potential" one. This refers to the next most "potential" set of hyperparameters. The main purpose of using the acquisition function is to avoid local optima by weighing the trade-off between exploring regions where the objective function is very ambiguous and extracting regions where the objective function is expected to be optimal.

[0089] Repeat the above steps until the target value on the assumed distribution in the invention achieves optimal results, or all computational resources are exhausted.

[0090] After 100 Bayesian optimization iterations, the hyperparameters of the model were finally determined. Table 3 shows the results of the Bayesian algorithm for optimizing the hyperparameters. With the hyperparameters in the table, we obtained the minimum average RMSE of K-means-GBDT for each category as 0.0412, 0.0357, and 0.0343, respectively.

[0091] Table 3 Results of Bayesian optimization of hyperparameters for each category of K-means-GBDT model

[0092] The performance evaluation of the model of this invention is as follows. This invention selects four metrics to evaluate model performance: Mean Absolute Error (MAE), Mean Squared Error (MSE), Root Mean Squared Error (RMSE), and Nash Efficiency Coefficient (NSE). The smaller the values ​​of MAE, MSE, and RMSE, the higher the model's accuracy; the closer the NSE value is to 1, the more accurate the model. The calculation formulas for the four metrics are as follows:

[0093]

[0094]

[0095]

[0096] In the formula For sample size, For the true value of the sample, For the predicted value of the sample, This is the average of the true values.

[0097] Model performance of K-means-GBDT model

[0098] Figure 11The figure shows the error of the maximum inundation depth simulated by the K-means-GBDT model in the test set. As can be seen from the figure, most prediction errors are below 0.05 m. In different grid categories, the maximum inundation depth varies slightly in a small number of grids in categories 1 and 2, while the difference is smallest in category 3. Furthermore, this invention evaluated the performance of the K-means-GBDT models for different categories using MAE, MSE, RMSE, and NSE, and the evaluation results are shown in Table 4. The NSE of all K-means-GBDT models reached above 0.95, indicating that the K-means-GBDT models have good reliability.

[0099] Table 5 shows the average time for the K-means-GBDT model and the PCSWMM model to simulate an urban flooding event. Using the same computer, the K-means-GBDT model simulates an urban flooding event in an average of only 43 seconds, 160 times faster than the PCSWMM model. Therefore, the model proposed in this invention is suitable for rapid prediction of urban flooding.

[0100] Table 4. Accuracy of different categories of K-means-GBDT models and GBDT models

[0101] Table 5. Average time for simulating an urban flooding event using the K-means-GBDT model and the PCSWMM model.

[0102] Comparison results of K-means-GBDT model and GBDT model To better demonstrate the accuracy of the K-means-GBDT model, this invention uses an unclassified GBDT model as a control. Figure 12 (a), (b), and (c) represent the maximum flood depths simulated by the PCSWMM, K-means-GBDT, and GBDT models for 50 sample grids in each class (random sampling is used for each class), respectively, under the R100-T100 condition. In the figure, the maximum flood depth simulated by the K-means-GBDT model for each class is closer to that simulated by the PCSWMM model than that simulated by the GBDT model.

[0103] To further compare the accuracy of the K-means-GBDT model and the GBDT model, Table 4 presents the performance of different categories of K-means-GBDT and GBDT models. As can be seen from the table, the accuracy of each category of K-means-GBDT model is improved to varying degrees compared to the GBDT model. In particular, the NSE of the second category model improves from 0.8951 to 0.9580, indicating that the addition of "[missing information]" based on the disaster-prone environment in this invention... The K-means-GBDT model achieved good results. When simulating all grids susceptible to flooding, the accuracy of the K-means-GBDT model was significantly improved compared to the GBDT model, with the NSE increasing from 0.9361 to 0.9761. The results demonstrate that the K-means-GBDT model possesses good accuracy.

[0104] This invention substitutes the rainfall-tide data of Typhoon Rammasun into the K-means-GBDT model. Observation points 1, 2, 3, 4, and 6 are in grid type 1, observation points 5 and 8 are in grid type 2, and observation point 7 is in grid type 3. Figure 13 The figure shows the difference between the observed depth and the predicted depth at these eight observation points. As can be seen from the figure, the error between the observed depth and the predicted depth is very small. This further verifies the accuracy of the deep learning model established in this invention.

[0105] In summary, the results show that the K-means-GBDT model has high prediction accuracy. The NSE of the three classes of K-means-GBDT models are 0.963, 0.958, and 0.991, respectively. Compared with the GBDT model, their accuracy is effectively improved, especially when the catastrophic effect degree is added. When the accuracy of the K-means-GBDT model for the second category was significantly improved, the model demonstrated high reliability and could be used for rapid prediction of urban flooding.

[0106] The present invention also provides a rapid urban flood prediction system based on disaster-prone environment clustering and Bayesian optimization deep learning model, including a processor and a memory; the processor and the memory are communicatively connected, and the processor is used to read a computer program from the memory and execute it to implement the aforementioned method.

[0107] The technical solutions provided by this invention are not limited to the above embodiments. Any technical solutions formed by using the structure and method of this invention through transformation and substitution are within the protection scope of this invention.

Claims

1. A deep learning-based method for predicting urban flooding that considers disaster-prone environmental clustering, characterized in that, Includes the following steps: S1 establishes a one- or two-dimensional coupled hydrodynamic model based on the PCSWMM model, and uses this model to simulate the regional inundation depth under different rainfall-tide combinations. S2 constructs a deep learning database based on rainfall data, tide data, disaster-prone environment data, and inundation depth data; In S2, design rainfall-tide data for different return periods in the invention area are calculated and loaded into the PCSWMM model to obtain inundation depth data for the invention area under different conditions. Basic data and geographic information system are used to obtain disaster-prone environmental data for the invention area. A deep learning database is constructed using rainfall data, tide data, disaster-prone environmental data and inundation depth data. S3 uses the K-means clustering algorithm to divide the region into different categories based on disaster-prone environmental characteristic indicators; In S3, the K-means clustering algorithm is used to divide the invention area into different categories based on the disaster-prone environmental characteristic indicators, including DEM, slope SL, pipeline density DON, distance from the river DR, and curvature CU. The optimal number of clusters is determined using the sum of squared errors function, and then the invention area is divided into the corresponding categories; the optimal number of clusters is 3. S4 Based on different disaster-prone environmental categories and their causes of flooding, different input variables are selected to establish different deep learning models, and the Bayesian algorithm is applied to optimize the hyperparameters of the deep learning models. In S4, based on the different disaster-prone environment characteristics of the grid with the optimal number of clusters obtained in S3, a corresponding deep learning model, namely the K-means-GBDT model, is established. For each category of K-means-GBDT model, appropriate input variables are selected, and hyperparameters that have a greater impact on the deep learning model are selected. Then, the corresponding optimization range and interval are set, and the hyperparameters of the model are optimized using the Bayesian algorithm. Among them, the input variables of the first and third type K-means-GBDT models include 5 disaster-prone environmental factors, 3 rainfall factors and 3 tidal level factors; The input variables for the second type of K-means-GBDT model are 5 disaster-prone environmental factors, "N", 3 rainfall factors, and 3 tidal level factors; Among them, there are 5 disaster-inducing environmental factors, namely DEM, slope SL, pipeline density DON, distance from the river channel DR, and curvature CU; 3 rainfall factors, namely total rainfall, maximum 3-hour rainfall, and peak rainfall; 3 tidal level factors, namely peak tidal level, average tidal level, and maximum 3-hour average tidal level; N is the degree of disaster-causing effect. The hyperparameters that have a significant impact on deep models are: loss function, number of iterations, learning rate, subsampling ratio, maximum depth, and minimum impurity decrease.

2. The urban flood deep learning prediction method considering disaster-prone environment clustering as described in claim 1, characterized in that: In S1, a one-dimensional hydrodynamic model is established using river channel and pipeline network data, and a two-dimensional surface inundation model is established using elevation data. The one-dimensional and two-dimensional models are connected through orifices. This model is used to simulate the inundation depth of the invention area under different rainfall-tide combinations. The data used to construct the PCSWMM model include digital elevation model (DEM) data, pipeline network data, river network data, and structures. The pipeline network data includes pipes and manholes, and the river network data includes river cross-sections and cross-section spacing.

3. The urban flood deep learning prediction method considering disaster-prone environment clustering as described in claim 1, characterized in that: The GBDT algorithm is built on the Scikit-Learn library in the Python 3.9 environment. The dataset is divided into training and test sets in a 7:3 ratio. The Bayesian algorithm for optimizing hyperparameters is implemented in the Hyperopt library in the Python 3.9 environment.

4. A deep learning-based urban flood prediction system considering disaster-prone environment clustering, characterized in that: It includes a processor and a memory; the processor and the memory are communicatively connected, and the processor is configured to read a computer program from the memory and execute it to implement the method of any one of claims 1 to 3.

Images