Urban inland inundation water depth prediction method considering geographical similarity

Abstract

The invention relates to an urban inland inundation water depth prediction method considering geographical similarity, and aims to introduce a geographical similarity theory into urban inland inundation modeling and analyze a geographical modeling mechanism contained in an urban inland inundation process under a rainfall scenario. Constructing a geographic configuration data set; preprocessing data, and constructing a geographic similarity calculation unit; based on the raster data, a hydrological hydrodynamic model is adopted to simulate the earth surface confluence production process of the research area, and an urban inland inundation water depth data set is constructed; based on the urban inland inundation data set and a geographic similarity calculation unit, an automatic machine learning model library is adopted to construct a geographic similarity model; performing hyper-parameter optimization, and optimizing the geographic similarity model; and calculating the predicted water depth. The method has the beneficial effects that the geographic factors are integrated, the geographic similarity theory is introduced into urban inland inundation modeling, and the accuracy and timeliness of an urban inland inundation simulation result are improved through automatic machine learning.

Description

Technical Field

[0001] The present invention belongs to the technical field of urban waterlogging in the field of hydrology and water resources, and in particular relates to a method for predicting water depth of urban waterlogging taking into account geographical similarities. Background Art

[0002] Against the backdrop of increasing global climate change, the continuous advancement of urbanization, and frequent extreme rainfall, the frequent occurrence of urban waterlogging has become a phenomenon that cannot be ignored. This extreme event not only poses a serious threat to public health, but also causes huge property losses and poses a serious challenge to social stability and economic development.

[0003] In response to the problem of urban waterlogging caused by heavy rains, some scholars have used physical modeling methods such as process mechanism expression and geographic calculation to reveal the spatial distribution, spatiotemporal process and evolution law of urban waterlogging geographical phenomena. However, the physical driving method cannot meet the requirements of urban waterlogging simulation for computing speed and real-time performance. Some scholars have achieved rapid simulation of urban waterlogging based on data-driven methods, but this method ignores relevant knowledge about waterlogging and the simulation results lack interpretability. Moreover, these methods fail to effectively take into account the complex geographical elements of the city and cannot accurately express the impact of geographical factors in the simulated area on the flood process.

[0004] Therefore, the present invention proposes a method for predicting urban waterlogging depth while taking into account geographical similarities. Under the condition of taking into account geographical similarities, it comprehensively considers multiple factors such as geographical location, climate, topography, vegetation coverage, etc., and uses an automatic machine learning framework for modeling to achieve a complete expression and prediction of the occurrence and development process of urban waterlogging. Summary of the invention

[0005] In view of the problems existing in the above-mentioned technologies, the present invention provides a method for predicting urban waterlogging depth taking into account geographical similarities, with the aim of introducing the geographical similarity theory into urban waterlogging modeling and analyzing the geographical modeling mechanism implied in the urban waterlogging process under rainfall scenarios, in order to achieve a more accurate and reliable prediction of urban waterlogging depth.

[0006] In order to achieve the above-mentioned purpose of the invention, the technical solution adopted by the present invention is: a method for predicting urban waterlogging depth taking into account geographical similarities, which comprises: S1: Collect basic geographic information data, calculate geographic environmental factors, and construct geographic configuration data sets; S2: data preprocessing, constructing geographic similarity calculation units; S3: Based on raster data, the hydrological and hydrodynamic model is used to simulate the surface runoff generation and convergence process in the study area and construct an urban waterlogging depth dataset; S4: Based on the urban waterlogging dataset and the geographic similarity calculation unit, the geographic similarity model is constructed using the automatic machine learning model library; S5: Hyperparameter optimization to optimize the geographic similarity model; S6: Calculation of predicted water depth.

[0007] Furthermore, the S1 comprises the following steps: S11: Select the Nanjing urban area and obtain the basic geographic information data of the area from the National Catalogue Service For Geographic Information. S12: Based on digital elevation data, the geographic features of the region were extracted through ArcMap built-in calculation methods, and a geographic configuration dataset based on 15 geographic factors including land use type, soil type, aspect, slope, plan curvature, profile curvature, roughness, roughness number, terrain moisture index, sediment transport index, water flow dynamics index, terrain ruggedness index, terrain position index, normalized difference vegetation index, and terrain control index was constructed; S13: Potential explanatory variables are selected through Spearman correlation analysis method. The calculation formula is as follows: (1), in, is the Spearman correlation coefficient value, is the total sample size, Representative The rank difference between the data.

[0008] The collinearity between explanatory variables was tested using the variance inflation factor (VIF). Generally speaking, a VIF greater than 10 is considered high collinearity, and variables with a VIF value exceeding 10 will be removed until all selected explanatory variables are ultimately below 10, thereby determining a geographic configuration dataset that fully expresses the characteristics of the geographic environment.

[0009] Further, the S2 comprises the following steps: S21: Standardize and normalize each factor in the geographic configuration data set, and characterize the processed data. The calculation formula is as follows: (2), (3), (4), in, is the result after processing. It is a piece of data. is the minimum value in the data, is the maximum value in the data, is the mean of the sample data, is the standard deviation of the sample data, is the explanatory variable for a certain location The value of is the set of explanatory variable values ​​for all locations; S22: Based on the observed position and unknown location Assume that there is a similarity relationship and calculate the unknown position , and observation location , The formula for the similarity is as follows: (5), (6), In the formula, is the initial similarity value, It is based on covariates and Similarity of location, The decision based on the covariance level is made by comparing the similarity of the covariance levels of all covariances. and The similarity function of S23: Evaluate the similarity between the unknown position and the observed position by calculating the distance value. The calculation formula is as follows: (7), (8), (9), (10), (11), (12), In the above formula, represents the Chebyshev distance value, represents the Manhattan distance value, represents the Euclidean distance value, represents the Minkowski distance value, represents the total number of hypothesized categories, represents the Canberra distance value, where is the true total number of categories, is the total sample size, All unknown locations With observation location Explanatory variables The square root of the mean deviation, Observation location A collection of Unknown location A collection of; The data that meets the similarity hypothesis are screened out through formulas (5)-(6). For the data that meets the similarity hypothesis, the similarity values ​​of the observed position and the unknown position are calculated respectively through the similarity measurement method of formulas (7)-(11). The weight vector matrix based on the average similarity value is constructed through formula (12) as the similarity calculation unit.

[0010] Further, the S3 comprises the following steps: S31: Modeling the urban surface of the embodiment. In order to carry out the modeling work of the urban surface part of the embodiment area, the present invention obtains the land use type, remote sensing images, buildings and river network distribution and other geographic spatial data, all of which use the WGS-84 coordinate system. When the plane coordinate system is used, the map projection uniformly uses the UTM zone 50N projection; S32: Establishing surface grid cells based on spatial discrete grids; S33: Based on the grid data, the hydrological and hydrodynamic model is used to model the urban surface runoff generation and convergence process, and the parameter information such as rainfall data, land use type data, elevation information (DEM), building data, infiltration rate, Manning coefficient grid, etc. are input into the grid unit; Further, the S33 comprises the following steps: S331: Based on typical rainfall events in the region, using point-like time series data from the rainfall station, the spatial rainfall distribution in the embodiment region is estimated by spatial interpolation; S332: Generate the Manning coefficient grid using land use type data and building data. Taking into account the research area and the runoff characteristics of the urban underlying surface, the present invention divides the urban surface grid into five types: grassland, water body, building, road, and impervious surface; S333: Construct Manning coefficient grids according to different types as input data for hydrological and hydrodynamic models; S34: adding the time step in the rainfall grid to the hydrological and hydrodynamic model input file to adjust the simulation time and total simulation time of the area to simulate waterlogging; S35: Based on the simulation results of the study area, data exploratory analysis methods are used to find missing or incorrectly labeled data, correct and delete inappropriate data, and outlier deletion uses the MAD outlier identification method. The calculation formula is as follows: (13), In the above formula, is the mean of the data, is the standard deviation of the data. Through the above operations, the urban flooding depth dataset is constructed.

[0011] Further, the S4 comprises the following steps: S41: Rainfall dataset, waterlogging depth dataset and similarity calculation unit are matched, i.e. the observed waterlogging depth value is used as a label for model learning; Further, the S41 comprises the following steps: S411: Time matching of rainfall sequence and water depth sequence. According to the existing rainfall sequence, the start time, time step, end time and other information of rainfall are obtained, and then the start time, time step and end time defined by the configured water depth result are matched (generally, this process cannot be matched directly). The matching is mainly to lengthen two inconsistent sequences. Due to the randomness of rainfall events, there is inconsistency in the time interval of rainfall, so it is necessary to add 0 value (default is non-rainfall) to the specified rainfall simulation time interval; S412: Spatial matching of rainfall sequence and water depth sequence. For spatial matching, spatial resolution interpolation conversion is performed based on the spatial resolution data required by the experiment. The calculation formula is as follows: (14), In the above formula, For the The observations at discrete points, is the grid point The estimated value on is the number of samples involved in the calculation, is the interpolation point and The distance between sites, is the power of the distance, The value of is 1.6.

[0012] S42: High-dimensional space mapping. After completing the above similarity calculation and data set matching, all similarity values ​​are mapped to the high-dimensional space and converted into a regression problem. The mapping operation is completed based on the kernel function, and the calculation formula is as follows: (15), In the above formula, is the newly constructed kernel function, is the total sample size, is an undetermined constant, is a combination kernel function, including a combination of linear kernel function and Gaussian kernel function. and represents different vectors in the input space, and The above steps are completed as the Lagrangian operator to construct a spatiotemporal dataset of urban waterlogging hydrology.

[0013] S43: preprocess the data and process missing values, abnormal values ​​and outliers; Further, the S43 comprises the following steps: S431: In order to check the quality of the data in the urban waterlogging hydrological spatiotemporal dataset, the interquartile range of the box plot is used to check for outliers. If the data point is above the upper quartile and exceeds 1.5 times or below the lower quartile and less than 1.5 times These points are considered outliers. S432: Since there are outliers in the water depth result data, the distribution of normal data is extremely compressed. The interquartile range detection algorithm is used to remove outliers. The calculation formula is as follows: (16), represents the interquartile range value, represents the third quartile, represents the first quartile; DEM data is more consistent with the real data distribution, so no processing is done. Since the median of rainfall data is generally 0, removing some 0s can reduce the sample imbalance caused by too many 0 values; S433: Remove data features with too high correlation to complete data set screening; S434: Divide the data set. Observe the distribution of the training set and the test set, and use the KDE distribution chart to view and compare the distribution of feature variables in the training set and the test set. Basically, the training set and the test set of all variables roughly meet the same distribution, which meets the requirements for data set division.

[0014] S44: Utilizes the AutoGluon automated machine learning framework, an automated machine learning library for training highly accurate machine learning models, using multi-layer stacking ensembles and Repeat Fold bagging to optimize the model training and validation process; Further, the S44 comprises the following steps: S441: Feature construction, which is to create new features from the urban waterlogging hydrological spatiotemporal dataset, generally using feature intersection (through some mathematical operations, such as four arithmetic operations, , etc., quickly derive new variables) and decompose the original features to create new features, which is handled by AutoGluon; S442: Divide the urban waterlogging hydrological spatiotemporal dataset. Randomly select 70% of the data in the urban waterlogging hydrological spatiotemporal dataset as the training set, and the remaining 30% as the test set for model input and verification; S443: Build a model library. AutoGluon first estimates the required training time. If it exceeds the remaining time of this layer, it jumps to the next stacked layer. After each new model is trained, it is immediately saved to disk for fault tolerance. This design makes the behavior of the framework highly predictable, ensuring that predictions can be generated by training at least one model within the specified time. In addition, when AutoGluon checks the intermediate iteration process of sequential training models such as neural networks and tree models, model generation within a limited time can still be guaranteed. In addition, if the model fails during training, a jump to the next model operation will be executed for this event. Generate multiple models; S444: Model selection: Based on the candidate models, the models are scored from the perspectives of model time consumption, computing resources, and whether the simulation results comply with hydrological rules, and the top three models are selected as candidate models.

[0015] S45: Constructing a geographic similarity model. Due to the performance differences of each model under different geographic configurations, this application performs local adaptation from the top three models with the highest comprehensive scores based on the accuracy of each model at the similarity calculation unit scale to construct a similarity model that meets the specific geographic environment.

[0016] Further, the S5 comprises the following steps: S51: Setting the hyperparameter search range to build a hyperparameter pool; S52: Optimize the model by adjusting hyperparameters; Further, the S52 comprises the following steps: S521: Search for hyperparameters using the Bayesian optimization algorithm; S522: In this process, in order to limit the time consumption of model training, the epoch_valid_improve parameter is added to save model training time; S523: Use the Gaussian process regression method and performance evaluation index in the Bayesian optimization algorithm to perform iterative search to obtain the optimal parameter combination. S53: Based on the optimal parameter combination, the optimal geographic similarity model is constructed. Further, the S6 comprises the following steps: S61: Design waterlogging scenarios based on different rainfall temporal resolution and spatial resolution parameters; Further, the S61 includes the following steps: S611: Using the rainstorm intensity formula, design rainfall scenarios with different return periods. The rainstorm intensity calculation formula for the experimental area is as follows: (17), In the above formula, is the design rainstorm intensity (mm / min), is the rainfall duration (min), is the design return period (years).

[0017] According to the rainstorm intensity formula in the "Notice on the Release of Nanjing Rainstorm Intensity Formula" issued by Nanjing Water Affairs Bureau in 2024, the design duration is 180 minutes, and the comprehensive rain peak position coefficient is 0.393, for the rainstorm scenarios of 1 year, 5 years, 10 years, 20 years, 50 years and 100 years; S612: By setting different time steps, more scenarios are constructed and these multi-temporal rainfall datasets are used as inputs to the hydrological and hydrodynamic models to simulate the flooding depths under different waterlogging scenarios.

[0018] S62: Model pre-training based on the same parameters. The optimal parameters in step 5 are used as prior parameters for similarity model construction; S63: Select different water depth thresholds, different time series rainfall data, and different calculation units to perform similarity model training; Further, the S63 includes the following steps: S631: Using spatial granularity grid training, multiple rainstorm scenarios are compared, and the top three water depth thresholds are kept as the final results in all selected experimental scenarios; S632: The impact of different time series rainfall data on the final training results is one of the factors that need to be considered in modeling. The length of the selected historical time series rainfall data is 1-30, and multiple rainstorm scenarios are compared. The top three water depth thresholds are kept as the final results in all selected experimental scenarios to compare the model performance; S633: Compare two types of computing units: the grid type and the neighborhood type; S634: The tree model and ensemble model have relatively superior performance and are trained in different scenarios; S64: water depth prediction calculation. The geographical environment and rainfall of the location to be predicted are used as inputs of the similarity model to perform water depth prediction calculations and output the predicted water depth results.

[0019] The details are as follows: S631: Using spatial granularity grid training, the water depth threshold is selected as 1mm-10mm, and multiple rainstorm scenarios are compared. In all selected experimental scenarios, the top three water depth thresholds are kept as the final results. The root mean square error and mean absolute error under different water depth thresholds are compared. It is found that the smaller the water depth threshold, the smaller the root mean square error and mean absolute error. The water depth threshold with the largest correlation coefficient and goodness of fit is 5mm. Finally, the best performance water depth threshold is selected as 5mm; S632: The impact of different time series rainfall data on the final training results is one of the factors that need to be considered in modeling. The length of the selected historical time series rainfall data is 1-30. Multiple rainstorm scenarios are compared. The top three water depth thresholds are kept as the final results in all selected experimental scenarios. The root mean square error and mean absolute error in all experimental scenarios and all rainfall time series are kept between 20mm-52mm and 13mm-38mm, and the correlation coefficient and goodness of fit are kept between 0.3-0.84 and 0.03-0.72. After comprehensive consideration, it is found that the rainfall time series at the first 26-29 moments performs well in multiple indicators. S633: Comparing the two types of calculation units, the spatial granularity grid type and the neighborhood type, the root mean square error and mean absolute error indicators, correlation coefficient and goodness of fit indicators of the neighborhood type calculation unit basically change in the same way. The overall root mean square error and mean absolute error are less than 30 mm and 17 mm, and the correlation coefficient and goodness of fit are greater than 0.85 and 0.7, which are in line with the ideal results; the root mean square error and mean absolute error of the spatial granularity grid type calculation unit are greater than 30 mm and 18 mm, and the correlation coefficient and goodness of fit are less than 0.85 and 0.7. It can be seen that the neighborhood type calculation unit has greatly improved the model performance; S634: Finally, the water depth threshold of 5 mm, the rainfall data of the first 26 historical moments and the neighborhood calculation unit were selected as the optimal configuration calculation. The tree model and the integrated model had relatively superior performance, the error was basically stable within 52 mm, and the goodness of fit R2 was greater than 0.6010.

[0020] Compared with the prior art, the present invention has the following advantages and beneficial effects: 1. Combined with geographical laws, the geographical similarity theory is introduced into urban waterlogging modeling, the urban rainfall process under waterlogging scenarios and the geographical modeling mechanism implied in the urban waterlogging process are analyzed, and the application of GIS theory in urban waterlogging modeling is improved; 2. Overcome the limitation of inaccurate flood simulation results caused by considering simple factors in traditional models, incorporate geographical environmental factors into model calculations, comprehensively utilize geographical factors and multi-time series rainfall data, conduct urban waterlogging modeling from the perspective of information geography and geographical similarity, characterize the spatiotemporal evolution characteristics of real urban rainstorm waterlogging events, and improve the accuracy and timeliness of rainstorm waterlogging simulation results; 3. In view of the spatial granularity (spatial resolution) effect and parameter uncertainty of model application, different spatial granularity scenarios and different model configuration conditions are designed to optimize parameters, which is conducive to improving the spatiotemporal simulation capability of urban waterlogging models at different spatial granularities, and providing reference for the study of urban waterlogging modeling in different regions and at different granularities, thereby effectively supporting flood control and drainage decisions in urban waterlogged areas; 4. Based on the geographic similarity theory and spatial similarity calculation method, the geographic configuration factor is used to construct a geographic similarity calculation unit, which is used as a priori knowledge model for urban waterlogging depth prediction, making the urban waterlogging depth prediction method more explainable; 5. Use high-precision urban water depth data sets to divide training sets and validation sets. Through a structured automatic machine learning framework, convert water depth image data into structured tabular data and conduct model training together with geographic similarity units. Select the optimal geographic similarity model from a variety of pre-assessment model combinations to achieve water depth simulation and prediction based on rainfall and geographic configuration factors. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 It is a framework diagram of the method of the present invention; Figure 2 This is a data outlier check diagram of the present invention; Figure 3 This is a flow chart of hyperparameter training of the model of the present invention; Figure 4 The model of the present invention Schematic diagram of fold cross validation; Figure 5 This is a schematic diagram of the water depth prediction results of the example area of ​​the present invention. DETAILED DESCRIPTION

[0022] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with embodiments and drawings. The exemplary embodiments of the present invention and their description are only used to explain the present invention and are not intended to limit the present invention.

[0023] Taking a certain area as an example, the area is surrounded by the Yangtze River, Qinhuai River and the ring road. There is no water exchange with the outside. The rainwater drainage network in the area is relatively independent, and there are various artificial ditches and lakes and other water storage units. Therefore, it can be determined that the area of ​​this embodiment belongs to a typical city with relatively clear boundary conditions. The overall terrain of the embodiment area is flat, with only a few mountains with a maximum altitude of 448.9 m, and rich geographical features and hydrological types.

[0024] like Figure 1 As shown, an embodiment of the present invention discloses a method for predicting urban waterlogging depth taking into account geographical similarities, which mainly includes the following steps: S1: Collect basic geographic information data, calculate geographic environmental factors, and construct geographic configuration data sets; S2: data preprocessing, constructing geographic similarity calculation units; S3: Based on raster data, the hydrological and hydrodynamic model is used to simulate the surface runoff generation and convergence process in the study area to construct an urban waterlogging depth dataset; S4: Based on the urban waterlogging dataset and the geographic similarity calculation unit, the geographic similarity model is constructed using the automatic machine learning model library; S5: Hyperparameter optimization to optimize the geographic similarity model; S6: Calculation of predicted water depth.

[0025] Furthermore, the S1 comprises the following steps: S11: Select the Nanjing urban area and obtain the basic geographic information data of the area from the National Catalogue Service For Geographic Information. S12: Based on digital elevation data, the geographic features of the region were extracted through ArcMap built-in calculation methods, and a geographic configuration dataset based on 15 geographic factors including land use type, soil type, aspect, slope, plan curvature, profile curvature, roughness, roughness number, terrain moisture index, sediment transport index, water flow dynamics index, terrain ruggedness index, terrain position index, normalized difference vegetation index, and terrain control index was constructed; As shown in Table 1; Table 1 Geographic configuration dataset S13: Potential explanatory variables are selected through Spearman correlation analysis method. The calculation formula is as follows: (1), in, is the Spearman correlation coefficient value, is the total sample size, Representative The rank difference between the data.

[0026] The collinearity between explanatory variables was tested using the variance inflation factor (VIF). Generally speaking, a VIF greater than 10 is considered high collinearity, and variables with a VIF value exceeding 10 will be removed until all selected explanatory variables are ultimately below 10, thereby determining a geographic configuration dataset that fully expresses the characteristics of the geographic environment.

[0027] Further, the S2 comprises the following steps: S21: Standardize and normalize each factor in the geographic configuration data set, and characterize the processed data. The calculation formula is as follows: (2), (3), (4), in, is the result after processing. It is a piece of data. is the minimum value in the data, is the maximum value in the data, is the mean of the sample data, is the standard deviation of the sample data, is the explanatory variable for a certain location The value of is the set of explanatory variable values ​​for all locations; S22: Based on the observed position and unknown location Assume that there is a similarity relationship and calculate the unknown position , and observation location , The formula for the similarity is as follows: (5), (6), In the formula, is the initial similarity value, It is based on covariates and Similarity of location, The decision based on the covariance level is made by comparing the similarity of the covariance levels of all covariances. and The similarity function of S23: Evaluate the similarity between the unknown position and the observed position by calculating the distance value. The calculation formula is as follows: (7), (8), (9), (10), (11), (12), In the above formula, represents the Chebyshev distance value, represents the Manhattan distance value, represents the Euclidean distance value, represents the Minkowski distance value, represents the total number of hypothesized categories, represents the Canberra distance value, where is the true total number of categories, is the total sample size, All unknown locations With observation location Explanatory variables The square root of the mean deviation, For observation location A collection of Unknown location A collection of; The data that meets the similarity hypothesis are screened out through formulas (5)-(6). For the data that meets the similarity hypothesis, the similarity values ​​of the observed position and the unknown position are calculated respectively through the similarity measurement method of formulas (7)-(11). The weight vector matrix based on the average similarity value is constructed through formula (12) as the similarity calculation unit.

[0028] Furthermore, S3 comprises the following steps: S31: Modeling the urban surface of the embodiment. In order to carry out the modeling work of the urban surface part of the embodiment area, the present invention obtains the land use type, remote sensing images, buildings and river network distribution and other geographic spatial data, all of which use the WGS-84 coordinate system. When the plane coordinate system is used, the map projection uniformly uses the UTM zone 50N projection; S32: Establishing surface grid cells based on spatial discrete grids; S33: Based on the grid data, the hydrological and hydrodynamic model is used to model the urban surface runoff generation and convergence process, and the parameter information such as rainfall data, land use type data, elevation information (DEM), building data, infiltration rate, Manning coefficient grid, etc. are input into the grid unit; Further, the S33 comprises the following steps: S331: Based on typical rainfall events in the region, using point-like time series data from the rainfall station, the spatial rainfall distribution in the embodiment region is estimated by spatial interpolation; S332: Generate the Manning coefficient grid using land use type data and building data. Taking into account the research area and the runoff characteristics of the urban underlying surface, the present invention divides the urban surface grid into five types: grassland, water body, building, road, and impervious surface; S333: Construct the Manning coefficient grid according to different types, and the Manning roughness coefficient values ​​are shown in Table 2 below: Table 2 Manning roughness coefficient values ​​for different surface types S34: adding the time step in the rainfall grid to the hydrological and hydrodynamic model input file to adjust the simulation time and total simulation time of the area to simulate waterlogging; S35: Based on the simulation results of the study area, data exploratory analysis methods are used to find missing or incorrectly labeled data, correct and delete inappropriate data, and outlier deletion uses the MAD outlier identification method. The calculation formula is as follows: (13), In the above formula, is the mean of the data, is the standard deviation of the data. Through the above operations, the urban flooding depth dataset is constructed.

[0029] Further, the S4 comprises the following steps: S41: Rainfall dataset, waterlogging depth dataset and similarity calculation unit are matched, i.e. the observed waterlogging depth value is used as a label for model learning; Further, the S41 comprises the following steps: S411: Time matching of rainfall sequence and water depth sequence. According to the existing rainfall sequence, the start time, time step, end time and other information of rainfall are obtained, and then the start time, time step and end time defined by the configured water depth result are matched (generally, this process cannot be matched directly). The matching is mainly to lengthen two inconsistent sequences. Due to the randomness of rainfall events, there is inconsistency in the time interval of rainfall, so it is necessary to add 0 value (default is non-rainfall) to the specified rainfall simulation time interval; S412: Spatial matching of rainfall sequence and water depth sequence. For spatial matching, spatial resolution interpolation conversion is performed based on the spatial resolution data required by the experiment. The calculation formula is as follows: (14), In the above formula, For the The observations at discrete points, is the grid point The estimated value on is the number of samples involved in the calculation, is the interpolation point and The distance between sites, is the power of the distance, The value of is 1.6.

[0030] S42: High-dimensional space mapping. After completing the above similarity calculation and data set matching, all similarity values ​​are mapped to the high-dimensional space and converted into a regression problem. The mapping operation is completed based on the kernel function, and the calculation formula is as follows: (15), In the above formula, is the newly constructed kernel function, is the total sample size, is an undetermined constant, is a combination kernel function, including a combination of linear kernel function and Gaussian kernel function. and represents different vectors in the input space, and For the Lagrangian operator, complete the above steps to construct the urban waterlogging hydrological spatiotemporal dataset; S43: preprocess the data and process missing values, abnormal values ​​and outliers; Further, the S43 comprises the following steps: S431: In order to check the quality of the data in the urban waterlogging hydrological spatiotemporal dataset, the interquartile range of the box plot is used to check for outliers. If the data point is above the upper quartile and exceeds 1.5 times or below the lower quartile and less than 1.5 times These points are considered outliers. S432: The abnormal value check result is as follows Figure 2 As shown in the figure, due to the presence of outliers in the water depth result data, the distribution of normal data is extremely compressed. The interquartile range detection algorithm is used to remove outliers. The calculation formula is as follows: (16), represents the interquartile range value, represents the third quartile, represents the first quartile; DEM data is more consistent with the real data distribution, so no processing is done. Since the median of rainfall data is generally 0, removing some 0s can reduce the sample imbalance caused by too many 0 values; S433: Remove data features with too high correlation to complete data set screening; S434: Divide the data set. Observe the distribution of the training set and the test set, and use the KDE distribution chart to view and compare the distribution of feature variables in the training set and the test set. Basically, the training set and the test set of all variables roughly meet the same distribution, which meets the requirements for data set division.

[0031] S44: Automated machine learning library for training highly accurate machine learning models using AutoGluon Tabular (automated machine learning framework for structured data), using multi-layer stacking ensembles and Repeat Fold bagging to optimize the model training and validation process; Further, the S44 comprises the following steps: S441: Feature construction, which is to create new features from the urban waterlogging hydrological spatiotemporal dataset, generally using feature intersection (through some mathematical operations, such as four arithmetic operations, , etc., quickly derive new variables) and decompose the original features to create new features, which is handled by AutoGluon; S442: Divide the urban waterlogging hydrological spatiotemporal dataset. Randomly select 70% of the data in the urban waterlogging hydrological spatiotemporal dataset as the training set, and the remaining 30% as the test set for model input and verification; S443: Establish a model library. AutoGluon first estimates the required training time, and if it exceeds the remaining time of this layer, it jumps to the next stacked layer. After each new model is trained, it is immediately saved to disk for fault tolerance. This design makes the behavior of the framework highly predictable, ensuring that predictions can be generated by training at least one model within the specified time. In addition, when AutoGluon checks the intermediate iteration process of sequential training models such as neural networks and tree models, model generation within a limited time can still be guaranteed. In addition, if the model fails during training, a jump to the next model operation will be executed for this event. Generate multiple models; S444: Model selection. Based on the candidate models, the models are scored from the perspectives of model time consumption, computing resources, and whether the simulation results comply with hydrological rules, and the top three models are selected as candidate models.

[0032] S45: Constructing a geographic similarity model. Due to the performance differences of each model under different geographic configurations, this application performs local adaptation from the top three models with the highest comprehensive scores based on the accuracy of each model at the similarity calculation unit scale to construct a similarity model that meets the specific geographic environment.

[0033] Further, the S5 comprises the following steps: S51: Set the hyperparameter search range to build a hyperparameter pool. Taking Tabular_nn and CatBoost models as examples, the hyperparameter search range is shown in Table 3; Table 3 Model hyperparameters and their ranges S52: Optimize the model by adjusting hyperparameters; Further, the S52 comprises the following steps: S521: Using the Bayesian optimization algorithm, taking the CatBoost model as an example, searching for hyperparameters is performed, as shown in Table 4; Table 4 Hyperparameters of CatBoost model S522: In this process, in order to limit the time consumption of model training, the epoch_valid_improve parameter is added to save model training time; S523: Use the Gaussian process regression method and performance evaluation index in the Bayesian optimization algorithm to perform iterative search to obtain the optimal parameter combination. The specific process is as follows: Figure 3 shown.

[0034] S53: Based on the optimal parameter combination, the optimal geographic similarity model is constructed.

[0035] Further, the S6 comprises the following steps: S61: Design waterlogging scenarios based on different rainfall temporal resolution and spatial resolution parameters; Further, the S61 includes the following steps: S611: Using the rainstorm intensity formula, design rainfall scenarios with different return periods. The rainstorm intensity calculation formula for the experimental area is as follows: (17), In the above formula, is the design rainstorm intensity (mm / min), is the rainfall duration (min), is the design return period (years).

[0036] According to the rainstorm intensity formula in the "Notice on the Release of Nanjing Rainstorm Intensity Formula" issued by Nanjing Water Affairs Bureau in 2024, the design duration is 180 minutes, and the comprehensive rain peak position coefficient is 0.393, for the rainstorm scenarios of 1 year, 5 years, 10 years, 20 years, 50 years and 100 years; S612: By setting different time steps, more scenarios are constructed and these multi-temporal rainfall datasets are used as inputs to the hydrological and hydrodynamic models to simulate the flooding depths under different waterlogging scenarios.

[0037] S62: Model pre-training based on the same parameters. The optimal parameters in step 5 are used as prior parameters for similarity model construction; S63: Select different water depth thresholds, different time series rainfall data, and different calculation units to perform similarity model training; Further, the S63 includes the following steps: S631: Use spatial granularity grid training, select the water depth threshold as 1mm-10mm, and use the k-fold cross-validation method to determine the best similarity threshold, such as Figure 4As shown in the figure, multiple rainstorm scenarios are compared. The top three water depth thresholds are kept as the final results in all selected experimental scenarios. The root mean square error and mean absolute error under different water depth thresholds are compared. It is found that the smaller the water depth threshold, the smaller the root mean square error and mean absolute error. The water depth threshold with the largest correlation coefficient and goodness of fit is 5mm. Finally, the best performance water depth threshold is selected as 5mm. S632: The impact of different time series rainfall data on the final training results is one of the factors that need to be considered in modeling. The length of the selected historical time series rainfall data is 1-30. Multiple rainstorm scenarios are compared. The top three water depth thresholds are kept as the final results in all selected experimental scenarios. The root mean square error and mean absolute error in all experimental scenarios and all rainfall time series are kept between 20mm-52mm and 13mm-38mm, and the correlation coefficient and goodness of fit are kept between 0.3-0.84 and 0.03-0.72. After comprehensive consideration, it is found that the rainfall time series of the first 26-29 moments performs well in multiple indicators; S633: Compare the two types of calculation units, the spatial granularity grid type and the neighborhood type. The root mean square error and mean absolute error indicators, correlation coefficient and goodness of fit indicators of the neighborhood type calculation unit basically change in the same way. The overall root mean square error and mean absolute error are less than 30mm and 17mm, and the correlation coefficient and goodness of fit are greater than 0.85 and 0.7, which are in line with the ideal results. The root mean square error and mean absolute error of the spatial granularity grid type calculation unit are greater than 30mm and 18mm, and the correlation coefficient and goodness of fit are less than 0.85 and 0.7. It can be seen that the neighborhood type calculation units have greatly improved the model performance; S634: Finally, the water depth threshold of 5 mm, the rainfall data of the first 26 historical moments and the neighborhood calculation unit are selected as the optimal configuration calculation. The tree model and the integrated model have relatively superior performance. The error is basically stable within 52 mm, and the goodness of fit R2 is basically greater than 0.60. The training results of different scenarios are shown in Table 5: Table 5 Model training results S64: Water depth prediction calculation. The geographical environment and rainfall of the location to be predicted are used as the input of the similarity model to calculate the future water depth prediction. The accuracy of the calculation result is as follows: Figure 5 As shown, its coefficient of determination The optimal value is 0.94, the trends of the observed and predicted water depth curves are consistent, and the predicted water depth results are output.

[0038] The description of the above embodiments is only used to help understand the method of the present invention and its core concept; at the same time, for a person skilled in the art, according to the concept of the present invention, there will be changes in the specific implementation method and application scope. In summary, the content of this specification should not be understood as a limitation of the present invention.

Claims

1. A method for predicting urban flooding depth taking into account geographical similarities, characterized in that: The following steps are involved: S1: Collect basic geographic information data, calculate geographic environmental factors, and construct geographic configuration data sets; S2: data preprocessing, constructing geographic similarity calculation units; S3: Based on raster data, the hydrological and hydrodynamic model is used to simulate the surface runoff generation and convergence process in the study area and construct an urban waterlogging depth dataset; S4: Based on the urban waterlogging dataset and the geographic similarity calculation unit, the geographic similarity model is constructed using the automatic machine learning model library; S5: Hyperparameter optimization to optimize the geographic similarity model; S6: Calculation of predicted water depth.

2. The method for predicting urban flooding depth taking into account geographical similarities according to claim 1, characterized in that: The S1 comprises the following steps: S11: Obtain regional geographic information basic data from the national geographic information resource directory service system; S12: Based on digital elevation data, the regional geographic features are extracted through the built-in calculation method of ArcMap, and a geographic configuration dataset based on 15 geographic factors including land use type, soil type, aspect, slope, plan curvature, profile curvature, roughness, roughness number, terrain moisture index, sediment transport index, water flow dynamics index, terrain ruggedness index, terrain position index, normalized vegetation index, and terrain control index is constructed; S13: Potential explanatory variables are selected through Spearman correlation analysis method. The calculation formula is as follows: （1）， in, is the Spearman correlation coefficient value, is the total sample size, Representative The rank difference between the data.

3. The method for predicting urban flooding depth taking into account geographical similarities according to claim 1, characterized in that: The S2 comprises the following steps: S21: Standardize and normalize each factor in the geographic configuration data set, and characterize the processed data. The calculation formula is as follows: （2）， （3）， （4）， in, is the result after processing. It is a piece of data. is the minimum value in the data, is the maximum value in the data, is the mean of the sample data, is the standard deviation of the sample data, is the explanatory variable for a certain location The value of is the set of explanatory variable values ​​for all locations; S22: Based on the observed position and unknown location Assume that there is a similarity relationship and calculate the unknown position , and observation location , The formula for the similarity is as follows: （5）， （6）， In the formula, is the initial similarity value, It is based on covariates and Similarity of location, The decision based on the covariance level is made by comparing the similarity of the covariance levels of all covariances. and The similarity function of S23: Evaluate the similarity between the unknown position and the observed position by calculating the distance value. The calculation formula is as follows: （7）， （8）， （9）， （10）， （11）， （12）， In the above formula, represents the Chebyshev distance value, represents the Manhattan distance value, represents the Euclidean distance value, represents the Minkowski distance value, represents the total number of hypothesized categories, represents the Canberra distance value, where is the true total number of categories, is the total sample size, All unknown locations With observation location Explanatory variables The square root of the mean deviation, Observation location A collection of Unknown location A collection of; The data that meets the similarity hypothesis are screened out through formulas (5)-(6). For the data that meets the similarity hypothesis, the similarity values ​​of the observed position and the unknown position are calculated respectively through the similarity measurement method of formulas (7)-(11). The weight vector matrix based on the average similarity value is constructed through formula (12) as the similarity calculation unit.

4. The method for predicting urban flooding depth taking into account geographical similarities according to claim 1, characterized in that: The construction of the S3 urban flood depth dataset includes the following steps: S31: Model the urban surface and obtain the land use type, remote sensing images, buildings and river network distribution geospatial data, all using the WGS-84 coordinate system. When using the plane coordinate system, the map projection uniformly adopts the UTM zone 50N projection; S32: Establishing surface grid cells based on spatial discrete grids; S33: Based on the grid data, the hydrological and hydrodynamic model is used to model the urban surface runoff generation and convergence process, and the rainfall data, land use type data, elevation information, building data, infiltration rate, and Manning coefficient grid parameter information are input into the grid cells; S34: adding the time step in the rainfall grid to the hydrological and hydrodynamic model input file to adjust the regional simulation time and the total simulation time to simulate waterlogging; S35: Based on the simulation results of the study area, data exploratory analysis methods are used to find missing or incorrectly labeled data, correct and delete inappropriate data, and outlier deletion uses the MAD outlier identification method. The calculation formula is as follows: （13）， In the above formula, is the mean of the data, is the standard deviation of the data. Through the above operations, the urban flooding depth dataset is constructed.

5. The method for predicting urban flooding depth taking into account geographical similarities according to claim 1, characterized in that: S4 Construction of urban waterlogging geographical similarity model includes the following steps: S41: Rainfall dataset, waterlogging depth dataset and similarity calculation unit are matched, that is, the observed waterlogging depth value is used as a label for model learning, as follows: S411: Time matching of rainfall sequence and water depth sequence. The start time, time step and end time information of rainfall are obtained according to the existing rainfall sequence. Then, the start time, time step and end time defined by the configured water depth result are matched. S412: spatial matching of rainfall sequence and water depth sequence. For spatial matching, spatial resolution interpolation conversion is performed based on the spatial resolution data required by the experiment. The calculation formula is as follows: （14）， In the above formula, For the The observations at discrete points, is the grid point The estimated value on is the number of samples involved in the calculation, is the interpolation point and The distance between sites, is the power of the distance, The value of is 1.6, S42: High-dimensional space mapping. After completing the similarity calculation and data set matching, all similarity values ​​are mapped to the high-dimensional space and converted into a regression problem. The mapping operation is completed based on the kernel function. The calculation formula is as follows: （15）， In the above formula, is the newly constructed kernel function, is the total sample size, is an undetermined constant, is a combination kernel function, including a combination of linear kernel function and Gaussian kernel function. and represents different vectors in the input space, and For the Lagrangian operator, complete the above steps to construct the urban waterlogging hydrological spatiotemporal dataset; S43: preprocess the data and process missing values, abnormal values ​​and outliers; S44: Automated machine learning framework for structured data using AutoGluon Tabular, an automated machine learning library for training highly accurate machine learning models, using multi-layer stacking ensembles and Repeat Fold bagging to optimize the model training and validation process; S45: Construct a geographic similarity model and perform local adaptation from the top three models with the highest comprehensive scores to construct a similarity model that meets the specific geographic environment.

6. The method for predicting urban flooding depth taking into account geographical similarities according to claim 1, characterized in that: S5 Urban waterlogging geographic similarity model optimization includes the following steps: S51: Setting the hyperparameter search range to build a hyperparameter pool; S52: Select a candidate model from the model library, randomly generate initial model parameters, encode them into strings, generate new strings based on Bayesian prior probability to replace part of the old strings, and create new parameters until the performance evaluation index is met and the model optimization is completed; S53: Using the Gaussian process regression method in the Bayesian optimization algorithm, the learning rate parameter, tree depth parameter, regularization coefficient, hidden layer size, dropout rate, number of layers, absolute skewness threshold, number of iterations and maximum number of rounds hyperparameters are iteratively adjusted to search for the optimal parameter results of each candidate model, and then the optimal model is selected from multiple candidate models to complete the construction of the optimal geographic similarity model.

7. The method for predicting urban flooding depth taking into account geographical similarities according to claim 6, characterized in that: The S6 comprises the following steps: S61: Design waterlogging scenarios based on different rainfall temporal resolution and spatial resolution parameters; S62: Based on the model pre-training with the same parameters, the optimal parameters are used as prior parameters for similarity model construction; S63: Select different water depth thresholds, different time series rainfall data, and different calculation units to perform similarity model training; S64: water depth prediction calculation, taking the geographical environment and rainfall of the location to be predicted as inputs of the similarity model to perform water depth prediction calculation, and outputting the predicted water depth result.

8. The method for predicting urban flooding depth taking into account geographical similarities according to claim 5, characterized in that: S43 data preprocessing includes the following steps: S431: Use the interquartile range of the box plot to check for outliers. If the data point is above the upper quartile and exceeds 1.5 times or below the lower quartile and less than 1.5 times These points are considered outliers. S432: Use the interquartile range detection algorithm to remove outliers. The calculation formula is as follows: （16）， represents the interquartile range value, represents the third quartile, represents the first quartile; S433: Remove data features with too high correlation to complete data set screening; S434: Divide the data set, observe the distribution of the training set and the test set, and use the KDE distribution diagram to view and compare the distribution of feature variables in the training set and the test set to meet the data set division requirements.

9. The method for predicting urban flooding depth taking into account geographical similarities according to claim 7, characterized in that: The S63 model training process includes the following steps: S631: Using spatial granularity grid training, the water depth threshold is selected as 1mm-10mm, and multiple rainstorm scenarios are compared. The top three water depth thresholds are kept as the final results in all selected experimental scenarios. The root mean square error and mean absolute error under different water depth thresholds are compared. The water depth threshold with the largest correlation coefficient and goodness of fit is 5mm, and the best performance water depth threshold is finally selected as 5mm; S632, the length of the selected historical time series rainfall data is 1-30, and multiple rainstorm scenarios are compared. The top three water depth thresholds are kept as the final results in all selected experimental scenarios. The root mean square error and mean absolute error in all experimental scenarios and rainfall time series at all times are kept between 20mm-52mm and 13mm-38mm, and the correlation coefficient and goodness of fit are kept between 0.3-0.84 and 0.03-0.

72. After comprehensive consideration, it is found that the rainfall time series at the first 26-29 moments performs well under multiple indicators; S633: Comparison of the two types of calculation units, the grid-type and neighborhood-type calculation units, shows that the root mean square error and mean absolute error, correlation coefficient and goodness of fit of the neighborhood-type calculation units change consistently, with the overall root mean square error and mean absolute error less than 30 mm and 17 mm, and the correlation coefficient and goodness of fit greater than 0.85 and 0.7, which are in line with the ideal results; the root mean square error and mean absolute error of the grid-type calculation units are greater than 30 mm and 18 mm, and the correlation coefficient and goodness of fit are both less than 0.85 and 0.7, S634: Finally, the water depth threshold of 5 mm, the rainfall data of the first 26 historical moments and the neighborhood calculation unit were selected as the optimal configuration calculation. The tree model and the integrated model performed well, with the error stable within 52 mm and the goodness of fit R2 greater than 0.6010.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the method for predicting urban flooding depth taking into account geographical similarities as described in any one of claims 1 to 9 above is implemented.

Images