Decision Tree-Based Urban Flooding Meteorological Model and Prediction Method

Abstract

This invention discloses a decision tree-based meteorological model and prediction method for urban flooding based on historical precipitation percentiles. Based on urban flooding disasters, it studies the precipitation-induced disaster factors, discovering and confirming that short-duration heavy rainfall is a common characteristic of the disaster. The maximum cumulative precipitation over different durations and the duration of heavy rainfall are selected as the main disaster-inducing meteorological factors. Binary data is established, with 80% used for training and 20% for testing. The optimal decision regression tree model for the maximum inundation depth is established by adjusting parameters using the CART decision tree regression method. Based on this, the historical percentiles of precipitation factors are introduced into the decision tree regression model, resulting in a decision tree regression model based on historical precipitation percentiles that considers both regional characteristics and site-specific features. The model outperforms decision tree models that do not consider historical precipitation percentiles and can be used for early identification and forecasting of urban flooding.

Description

Technical Field

[0001] This invention belongs to the field of meteorological forecasting technology, and in particular relates to a decision tree-based urban flooding meteorological model and forecasting method based on historical precipitation percentiles. Background Technology

[0002] Urban flooding occurs when continuous or heavy rainfall exceeds the drainage capacity of the city's drainage network, preventing rainwater from draining and causing waterlogging. Urban flooding weather warnings are an effective means of urban disaster prevention and mitigation, and the research on warning models is crucial. Using quantitative precipitation forecasting to drive urban flooding models to predict water depth over the next few hours can extend the effective lead time and improve forecast accuracy, providing a scientific basis for urban flood control and prevention.

[0003] However, current research on urban flooding mainly focuses on flooding simulation and risk assessment, while research on flood forecasting and early warning is relatively limited. Furthermore, hydrological flooding models suffer from drawbacks in operational applications, including high requirements for basic data, making them unsuitable for widespread implementation. Additionally, their large computational demands affect the real-time effectiveness of online applications, thus hindering the timely implementation of urban flooding meteorological early warning systems. While machine learning methods have developed rapidly in recent years and achieved significant results, their application in urban flooding remains limited due to the scarcity of urban flooding data. Summary of the Invention

[0004] Purpose of the Invention: To address the aforementioned problems and shortcomings of existing methods, the purpose of this invention is to provide a decision tree-based urban flooding meteorological model and prediction method based on historical precipitation percentiles. This invention establishes a regional decision tree regression model for regional urban flooding. Based on the analysis of disaster-causing factors, it utilizes the main meteorological factors contributing to urban flooding to establish a decision tree regression model. Furthermore, considering the distribution characteristics of historical precipitation over different durations, a decision tree-based urban flooding model is established, thereby achieving early identification of urban flooding.

[0005] Technical Solution: To achieve the above-mentioned objectives, this invention adopts the following technical solution: a decision tree-based urban flooding meteorological model based on historical precipitation percentiles, comprising the following steps:

[0006] S1, Data Acquisition: Obtain historical flood data and historical flood-free data for cities across the country as model samples. The historical flood data includes the occurrence time of urban flooding, maximum inundation depth, 24-hour precipitation from 1959 to 2023, and 1-hour precipitation from 2013 to 2023. If there is no flooding within 3 days from the date of the flooding disaster, extract the precipitation data on the 3rd day after the flooding disaster to establish historical flood-free data. Use 80% of the model samples as training samples and the remaining 20% ​​as test samples.

[0007] S2, determine the key influencing factors of urban flooding, with the maximum cumulative precipitation in 1 hour, 2 hours, 3 hours, 4 hours, 5 hours, 6 hours, 12 hours, 24 hours and the duration of heavy rainfall as the key influencing factors of the maximum inundation depth of the city;

[0008] S3, Establishment of the decision regression tree flooding model: Construct a decision tree flooding regression model with the maximum limit depth and minimum training sample parameters, and test the model with training samples to obtain the tested decision regression tree flooding model;

[0009] S4, in each node of the decision regression tree flooding model, the decision tree threshold is updated according to the value corresponding to the historical percentile of precipitation, thus completing the model update.

[0010] Furthermore, the model update process described in step S4 includes the following steps:

[0011] (1) In the decision regression tree flooding model, in each node, search for the sample that is closest to the decision tree threshold of the node among all the sample data of the node, and obtain the precipitation value of the closest sample.

[0012] (2) Identify the station to which the sample belongs and sort the historical precipitation of that station;

[0013] (3) Based on the precipitation of the closest sample, calculate the percentile of the historical precipitation ranking of the station to which it belongs;

[0014] (4) Sort the historical precipitation of other stations in the node respectively, and search the historical precipitation of each station according to the percentile confirmed in step (3) to obtain the precipitation value of the corresponding percentile in each station.

[0015] (5) The percentile precipitation values ​​of each station obtained in step (4) are used as the threshold values ​​of the node in turn to verify the accuracy of the model obtained by the sample evaluation test, and the one with the highest accuracy is selected as the threshold value of the node.

[0016] (6) Other nodes in the decision regression tree flooding model shall complete the threshold update of each node according to steps (1)-(5).

[0017] Furthermore, the maximum depth limit value in step S3 is 2, and the minimum training sample parameter value is 7.

[0018] This invention also provides a prediction method for urban flooding meteorological models based on historical precipitation percentiles. Based on the above-mentioned decision regression tree flooding model, by inputting quantitative precipitation estimation data and / or quantitative precipitation forecast data, a future urban flooding meteorological risk forecast can be obtained.

[0019] Furthermore, the future urban flooding meteorological risk forecast is divided into blue, yellow, orange and red warning levels according to the severity of flooding. The maximum inundation depth ranges corresponding to the blue, yellow, orange and red warning levels are 0.05m~0.2m, 0.21m~0.35m, 0.36m~0.5m and greater than 0.5m, respectively.

[0020] Beneficial effects: Compared with the prior art, the present invention uses a decision regression tree and the maximum cumulative precipitation over 3 hours, 6 hours, 1 hour, 12 hours and 24 hours as thresholds to predict the maximum water depth of urban flooding. On this basis, the historical percentile of precipitation factors is introduced into the decision tree regression model to obtain a decision tree regression model based on historical percentile of precipitation that takes into account both regional characteristics and site specificity. The model performance is better than the decision tree model that does not consider historical percentile of precipitation and can be used for the early identification and forecasting of urban flooding. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating the prediction method of the urban flooding meteorological model based on historical precipitation percentiles as described in this invention.

[0022] Figure 2 The ROC curve for the Logistic regression model test of the decision tree regression model parameters of this invention;

[0023] Figure 3 This is a structural diagram of the decision tree regression model of this invention;

[0024] Figure 4 This is a density function graph of the maximum 1-12 hour precipitation in urban areas across China from 2019 to 2022, as shown in this embodiment of the invention.

[0025] Figure 5 This is a density function graph of the maximum cumulative precipitation in urban areas across China from 2019 to 2022 during the period of 12-24 hours, as shown in this embodiment of the invention. Detailed Implementation

[0026] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. After reading this invention, any modifications of the invention in various equivalent forms by those skilled in the art will fall within the scope defined by the appended claims.

[0027] The overall technical approach of this invention is mainly based on historical urban flooding events. It analyzes the maximum hourly rainfall and short-duration rainfall duration over a 1-24 hour period to identify key factors contributing to urban flooding. By comparing the effects of various regression methods, the optimal solution is identified and modified. Historical rainfall data is then incorporated into the model for percentile testing to obtain the best accuracy, thus finalizing the model. Specific steps include:

[0028] I. Data Source

[0029] This invention collects case studies of urban flooding across China from 2019 to 2023 through the China Meteorological Administration's disaster reporting system and official media information, matching them with daily precipitation. Case studies are deleted if there is no precipitation near the disaster site. Through screening, 139 case studies of urban flooding were collected, including the time of occurrence, location, and maximum inundation depth. Historical precipitation data from 24-hour observation stations (1959-2023) and 10-year (2013-2023) observation stations were obtained from the National Meteorological Center.

[0030] Currently, disaster data includes samples of urban flooding events. However, establishing a decision tree regression model requires samples from areas where flooding has not occurred, and the model's performance is highly dependent on the selection of the zero value. This invention selects the zero value based on conditions such as no flooding event occurring before rainfall and no flooding occurring before rainfall. Specifically, if there is no flooding within three days of the flooding event, rainfall data from the third day after the flooding event is extracted to create historical flood-free data; after deduplication, previous rainfall is extracted. After determining the modeling sample in this way, 80% of the sample is used for modeling, and 20% is used for validation.

[0031] II. Determination of Urban Flooding Causes

[0032] The present invention specifies the occurrence time of urban flooding to the day. To analyze the characteristics of disaster-causing precipitation, hourly precipitation on the day of urban flooding and the two days prior is extracted. The cumulative precipitation over 1 to 24 hours is calculated using a sliding method, and the maximum value is used to analyze the main characteristics of the disaster-causing precipitation. A density function graph of the maximum 1-12 hour precipitation for urban flooding across the country from 2019 to 2022 is also generated. Figure 4 (and density function graphs of the maximum cumulative precipitation and duration of short-duration heavy rainfall in urban areas across China during 12-24 hours from 2019 to 2022) Figure 5 ).

[0033] The above analysis reveals that the main cause of urban flooding is the maximum cumulative rainfall over 1-6 hours, with a phase difference of approximately 10-20 mm in the density map per hour, and a similar distribution of maximum cumulative rainfall density over 12-24 hours. Therefore, this invention selects nine factors—maximum cumulative rainfall over 1-6 hours, 12 hours, 24 hours, and the duration of heavy rainfall—to conduct a decision tree modeling experiment on the maximum inundation depth of the city.

[0034] III. Model Type Determination

[0035] To select the best-performing method, this invention experimented with four regression methods: multiple regression, KNN, random forest, and decision tree. In the KNN method, five nearest neighbors were considered, and 30 leaf nodes were used. For the latter three methods, three leaf nodes were selected. Five-fold cross-validation was used to obtain the accuracy, absolute error, and mean squared error of the training and validation data. Urban flooding meteorological regression models were established using each of the four methods, and the accuracy and error of the four regression methods were compared, as detailed in Table 1 below.

[0036] Table 1. Accuracy, Mean Absolute Error, and Mean Relative Error of the Four Regression Methods

[0037]

[0038] As shown in Table 1, for the training samples, Random Forest performed best, achieving an accuracy of 0.9. KNN and Decision Tree were similar, at 0.61 and 0.63 respectively. Multiple Regression performed the worst, with an accuracy of only 0.53. However, testing on the validation samples revealed no significant difference in performance among the four methods. KNN performed slightly better, while Random Forest performed slightly worse, but the difference was too small to be representative. The differences in Mean Absolute Error and Mean Squared Error were also minimal. The results on the validation samples are crucial for selection. Although Random Forest performed best on the experimental samples, it did not offer a significant advantage in accuracy on the validation samples.

[0039] Based on the experimental results, there is insufficient reason to choose any particular method. Machine learning methods can be categorized into white-box and black-box models from the perspective of transparency and interpretability. White-box models emphasize high transparency and interpretability, such as linear regression and decision tree methods, while black-box models emphasize predictability but neglect interpretability, such as random forests. To facilitate forecasters' better understanding of the model's internal structure and decision logic, this invention, which generates output based on input, utilizes a decision tree method to establish an urban flooding weather forecasting model, which is a better choice.

[0040] IV. Determining the Parameters of the Decision Tree Regression Model

[0041] To find better parameters for the decision tree regression model, experiments were conducted on two important parameters for growth and decision tree pruning: maximum limit depth and minimum training sample parameters. Cross-validation was also performed, and the specific experimental values ​​are shown in Table 2.

[0042] Table 2 Main Parameters of the Experiment

[0043]

[0044] The test yielded evaluation results for each parameter, such as Figure 2 As shown, it can be seen that selecting a maximum depth limit of 2 and a minimum training sample parameter of 7 yields the best results, with the AUC area of ​​the ROC curve being 0.82. Further adjustments were made to parameters such as the number of samples for each branch, selecting parameters where the scores of both test and experimental samples tended to be stable and consistent for modeling.

[0045] V. Constructing a Decision Tree Regression Model

[0046] The decision regression tree flooding model obtained after testing is as follows: Figure 3 In the diagram, `node` represents a node, and `X[i]` represents `i` variables, indicating the maximum cumulative rainfall in a given hour. From the decision regression tree, it can be seen that the model uses the maximum cumulative rainfall over 3, 24, 1, and 12 hours, as well as the duration of short-duration heavy rainfall. The first layer uses the maximum cumulative rainfall over 3 hours, with a threshold of 48.9 mm. The model's root mean square error (RMSE) for samples where urban flooding does not occur is less than 0.001, while the RMSE for the predicted maximum water depth for samples where flooding does occur is 0.1–0.3. The RMSE for predicting a maximum inundation depth exceeding 1 meter is 0.348, indicating that the performance is acceptable. Furthermore, the accuracy rate of the test data after modeling is 0.63, and the validation data score is 0.43.

[0047] VI. Introducing historical percentile precipitation into the decision tree model

[0048] To improve the accuracy of the experiment, this invention incorporates historical precipitation percentiles into the decision tree, building upon the decision tree threshold. The main method involves conducting experiments on each quantile for stations where 80% of the samples exceed the threshold's percentile. The relative percentile of the exceeding precipitation is found within the corresponding historical precipitation ranking. Then, the percentiles for each meteorological station are extracted, the threshold is updated, and modeling is performed. The effectiveness is then verified. The specific process is as follows:

[0049] (1) In the decision regression tree flooding model, in each node, search for the sample that is closest to the decision tree threshold of the node among all the sample data of the node, and obtain the precipitation value of the closest sample.

[0050] (2) Identify the station to which the sample belongs and sort the historical precipitation of that station;

[0051] (3) Using the precipitation of the closest sample as the standard, calculate the percentile of the historical precipitation ranking of the station to which it belongs;

[0052] (4) Sort the historical precipitation of other stations in the node respectively, and search the historical precipitation of each station according to the percentile confirmed in step (3) to obtain the precipitation value of each station corresponding to the percentile.

[0053] (5) The percentile precipitation values ​​of each station obtained in step (4) are used as the threshold values ​​of the node respectively, and the accuracy of the sample evaluation test model is checked. The one with the highest accuracy is selected as the threshold value of the node.

[0054] (6) Other nodes in the decision regression tree flooding model shall complete the threshold update of each node according to steps (1)-(5). As shown in Table 3, the accuracy is the result after incorporating historical precipitation percentiles.

[0055] Table 3. Accuracy before and after incorporating historical precipitation percentiles

[0056]

[0057] The results show that after introducing historical precipitation percentiles, the overall accuracy of the test samples improved to 0.51, and the mean square error decreased from 0.13 to 0.11, demonstrating improved performance. In particular, for samples with a flooding depth greater than 0.5 meters, the mean square error decreased from 0.45 meters to 0.33 meters, showing a significant improvement.

[0058] VII. Decision Tree Model Prediction

[0059] Using quantitative precipitation estimation and forecast as input data, an improved urban flooding decision tree model is used to forecast the maximum inundation depth of urban flooding.

[0060] Here, a decision tree regression method based on historical precipitation percentiles is introduced based on regression experiments. Quantitative precipitation estimation (QPE) and quantitative precipitation forecast (QPF) are used to forecast the meteorological risk of future urban flooding. The maximum inundation depth is forecasted and warned according to the blue, yellow, orange, and red warning levels corresponding to the intervals of 0.05m~0.2m, 0.21m~0.35m, 0.36m~0.5m, and greater than 0.5m. Example

[0061] The following examples demonstrate the effectiveness of the forecasting model of the present invention: From 22:00 on July 10, 2024 to 7:00 on July 11, 2024, Dianjiang County experienced widespread heavy to torrential rain. Affected by the heavy rainfall, Chengxi Town in Dianjiang County experienced flooding, with some road sections having a water depth of about 2 meters, causing significant economic losses.

[0062] It is evident that, in addition to forecasting future precipitation, the model also captures historical precipitation characteristics, and can also capture the characteristics of urban flooding in recent actual precipitation, providing a good indication of the process.

[0063] In addition, 34 cases of urban flooding collected in 2024 were examined. Among the 34 cases, 7 were red alerts, 6 were orange alerts, 5 were yellow alerts, and 4 were blue alerts, with a hit rate of 65%. 12 cases were missed, accounting for 35%, which has good indicative value.

[0064] In summary, this invention, based on the decision tree threshold, introduces historical precipitation percentiles into the decision tree, updates the threshold, and then models the data. It was found that after introducing historical precipitation percentiles, the overall accuracy of the test samples improved to 0.51, and the mean square error decreased from 0.13 to 0.11, demonstrating improved performance. In particular, for samples with a flooding depth greater than 0.5 meters, the mean square error decreased from 0.45 meters to 0.33 meters, showing a significant improvement.

[0065] This invention, based on urban flooding disasters, studies the precipitation-induced disaster factors of urban flooding, discovering and confirming that short-duration heavy rainfall is a common feature of the disaster. The maximum cumulative rainfall over different durations and the duration of heavy rainfall are selected as the main disaster-causing meteorological factors. A binary classification dataset is established, with 80% used for training and 20% for testing. The optimal decision regression tree model for maximum inundation depth is established through parameter tuning using the CART decision tree regression method. The decision regression tree uses the maximum cumulative rainfall over 3, 6, and 1 hours, and 12 and 24 hours as thresholds to predict the maximum water depth of urban flooding. Furthermore, historical percentiles of precipitation factors are introduced into the decision tree regression model, resulting in a decision tree regression model based on historical percentiles of precipitation that considers both regional characteristics and site-specific features. This model outperforms decision tree models that do not consider historical percentiles of precipitation and can be used for the early identification and forecasting of urban flooding.

[0066] The above description is only a partial embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for constructing an urban flooding meteorological model based on historical precipitation percentiles, characterized in that... Includes the following steps: S1, Data Acquisition: Obtain historical flood data and historical flood-free data for cities across the country as model samples. The historical flood data includes the occurrence time of urban flooding, maximum inundation depth, 24-hour precipitation from 1959 to 2023, and 1-hour precipitation from 2013 to 2023. If there is no flooding within 3 days from the date of the flooding disaster, extract the precipitation data on the 3rd day after the flooding disaster to establish historical flood-free data. Use 80% of the model samples as training samples and the remaining 20% ​​as test samples. S2, determine the key influencing factors of urban flooding, with the maximum cumulative precipitation in 1 hour, 2 hours, 3 hours, 4 hours, 5 hours, 6 hours, 12 hours, 24 hours and the duration of heavy rainfall as the key influencing factors of the maximum inundation depth of the city; S3, Establishment of the decision tree urban flooding meteorological model: Select the key influencing factors mentioned in step S2, and construct a decision tree flooding regression model with the maximum limit depth and minimum training sample parameters. Test the model with training samples to obtain the tested decision tree urban flooding meteorological model. S4, in each node of the decision tree urban flooding meteorological model, the decision tree threshold is updated according to the value corresponding to the historical percentile of precipitation, and the model update is completed; The model update process described in step S4 includes the following steps: (1) In each node of the decision tree urban flooding meteorological model, search for the closest sample to the decision tree threshold of the node among all sample data of the node, and obtain the precipitation value of the closest sample. (2) Identify the station to which the sample belongs and sort the historical precipitation of that station; (3) Based on the precipitation of the closest sample, calculate the percentile of the historical precipitation ranking of the station to which it belongs; (4) Sort the historical precipitation of other stations in the node respectively, and search the historical precipitation of each station according to the percentile confirmed in step (3) to obtain the precipitation value of the corresponding percentile in each station. (5) The percentile precipitation values ​​of each station obtained in step (4) are used as the threshold values ​​of the node in turn to verify the accuracy of the model obtained by the sample evaluation test, and the one with the highest accuracy is selected as the threshold value of the node. (6) Other nodes in the decision tree urban flooding meteorological model shall complete the threshold update of each node according to steps (1)-(5).

2. The method for constructing an urban flooding meteorological model based on historical precipitation percentiles according to claim 1, characterized in that: The maximum depth limit value in step S3 is 2, and the minimum training sample parameter value is 7.

3. A prediction method for urban flooding meteorological models based on historical precipitation percentiles, characterized in that: Based on the decision tree urban flooding meteorological model obtained in claim 1 or 2, a future urban flooding meteorological risk forecast can be obtained by inputting quantitative precipitation estimation data and / or quantitative precipitation forecast data.

4. The prediction method for urban flooding meteorological models based on historical precipitation percentiles according to claim 3, characterized in that: The future urban flooding meteorological risk forecast is divided into blue, yellow, orange and red warning levels according to the severity of flooding. The maximum inundation depth ranges corresponding to the blue, yellow, orange and red warning levels are 0.05m~0.2m, 0.21m~0.35m, 0.36m~0.5m and greater than 0.5m, respectively.

Images