Urban flood disaster rapid prediction method based on integrated instant learning algorithm

Abstract

This invention discloses a rapid urban flood disaster prediction method based on an integrated real-time learning algorithm, comprising: acquiring a historical dataset containing historical rainfall and flood feature vectors; selecting multiple similarity measurement methods, and based on each method, filtering relevant historical sample sets corresponding to the rainfall forecast data to be measured from the historical dataset; constructing local prediction models using each sample set and obtaining preliminary prediction results; weighted integrating multiple preliminary prediction results to obtain the final flood disaster prediction result; when historical data is insufficient, supplementing the dataset by constructing a hydrological and hydrodynamic model to simulate flood data under different rainfall scenarios. This invention effectively solves the problems of poor adaptability and overfitting caused by fixed parameters in traditional offline models. By integrating multiple similarity measures and local online modeling, it significantly improves the accuracy, stability, and real-time performance of flood prediction, providing reliable technical support for urban flood prevention and disaster reduction.

Description

Technical Field

[0001] This invention belongs to the field of urban flood early warning and forecasting technology, and in particular relates to a rapid prediction method for urban flood disasters based on an integrated real-time learning algorithm. Background Technology

[0002] Against the backdrop of intensifying global climate change, extreme heavy rainfall events are on the rise, particularly sudden floods caused by short-duration heavy rainfall, which are causing increasingly severe damage to highly urbanized areas. However, numerical models of urban flooding based on physical processes are relatively slow, making it difficult to meet the increasingly demanding timeliness requirements of forecasting and early warning. In recent years, machine learning techniques have been widely applied in fields such as computer vision, machine translation, and robotics. Due to the advantages of machine learning algorithms, such as high computational speed and high prediction accuracy, numerous researchers have developed a large number of flood prediction models based on machine learning techniques, demonstrating the significant application potential of artificial intelligence technologies, including machine learning, in rapid flood prediction.

[0003] Traditional machine learning algorithms employ offline training methods based on all historical data, resulting in fixed model structures and parameters. However, urban flooding processes encompass surface runoff confluence and evolution, pipe network drainage, and river network confluence, making them complex predictors. When traditional offline machine learning prediction models are applied online, the complexity of urban flooding processes leads to poor adaptability and overfitting, resulting in unstable prediction models and significant errors in certain scenarios. Furthermore, this method relies on modeling all historical data. As the amount of data increases, the computational load on the model also increases significantly, affecting the timeliness of predictions. Therefore, the key to maintaining excellent model performance lies in establishing a machine learning modeling framework that can dynamically adjust its computational parameters and reduce computational load, thereby improving the accuracy and rapid prediction capabilities of flood forecasting models. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention proposes a rapid urban flood disaster prediction method based on an integrated real-time learning algorithm, thereby resolving the issues present in the existing technologies.

[0005] To achieve the above objectives, this invention provides a method for rapid prediction of urban flooding based on an integrated real-time learning algorithm, comprising: Obtain a historical dataset, which includes historical rainfall feature vectors and corresponding historical flood feature vectors; Multiple similarity measurement methods are selected, and based on each similarity measurement method, relevant historical sample sets corresponding to the rainfall forecast data to be measured are selected from the historical dataset; Based on each of the selected historical sample sets, a corresponding local prediction model is constructed, and preliminary prediction results for the rainfall forecast data to be measured are obtained through each local prediction model. The multiple preliminary forecasts are weighted and integrated to obtain the final flood disaster forecast.

[0006] Optionally, the process of obtaining the historical dataset includes: Historical flood monitoring data within the forecast area are collected to form a basic training sample set; a hydrological and hydrodynamic model of the forecast area is constructed, and flood data under different rainfall scenarios are simulated to obtain a supplementary training sample set; the historical dataset is composed of the basic training sample set and the supplementary training sample set.

[0007] Optionally, the multiple similarity measurement methods include: a first similarity measurement method based on Euclidean distance; a second similarity measurement method based on a joint definition of distance and angle; and a third similarity measurement method based on partial least squares method to supervise the latent variable space.

[0008] Optionally, the process of selecting relevant historical sample sets from the historical dataset based on each similarity measurement method includes: Calculate the similarity index value of each historical sample in the historical dataset relative to the rainfall forecast data to be measured using each similarity measurement method; The similarity index values ​​calculated by each similarity measurement method are sorted in descending order; A predetermined number of historical samples with the highest ranking are selected to form a relevant historical sample set corresponding to the similarity measurement method.

[0009] Optionally, before screening, the process may also include: performing correlation analysis on the input feature parameters of the rainfall feature vector using Pearson correlation analysis; selecting input feature parameters whose absolute value of the Pearson correlation coefficient with the flood feature vector is greater than a preset threshold for similarity calculation and local prediction model construction.

[0010] Optionally, the algorithm used to construct the local prediction model is the local weighted partial least squares algorithm.

[0011] Optionally, the process of obtaining preliminary prediction results using a local prediction model includes: Based on the input variable matrix and output variable matrix of the relevant historical sample set and the rainfall forecast data to be measured, a local weighted partial least squares regression calculation is performed to obtain the preliminary prediction results.

[0012] Optionally, the process of weightedly integrating multiple preliminary predictions includes: A confidence weight is assigned to the preliminary prediction result corresponding to each of the local prediction models, and the confidence weight is determined based on the prediction error of the corresponding local prediction model; The final flood disaster prediction result is obtained by multiplying each preliminary prediction result by its corresponding confidence weight and then summing the results.

[0013] Optionally, flood data under different rainfall scenarios can be simulated, including designing rainfall processes covering different rainfall durations, rainfall intensities, and cumulative rainfall amounts, and simulating them using the hydrological and hydrodynamic model.

[0014] Compared with the prior art, the present invention has the following advantages and technical effects: This invention effectively solves the problems of poor adaptability and overfitting caused by the fixed structure and parameters and inability to update in real time in traditional offline models. Simultaneously, by integrating multiple similarity measurement methods to construct an Integrated Just-In-Time Learning (E-JITL) algorithm, it compensates for the shortcomings of single similarity measurement methods in performance selection. In the E-JITL algorithm, multiple similarity measurement methods are first used to screen samples, forming multiple sets of relevant samples; then, online local prediction models are constructed for each set of samples, and finally, a high-precision prediction result is obtained by weighted averaging. Compared with traditional global models, this method can update model parameters in real time to ensure excellent forecasting performance, significantly improving the adaptability and stability of urban flood rapid prediction models, and providing reliable technical support for urban flood control and disaster reduction. Attached Figure Description

[0015] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a comparison diagram of the modeling and calculation processes of the traditional global modeling method and the real-time learning algorithm in an embodiment of the present invention; Figure 2 This is a flowchart of the E-JITL modeling framework calculation process according to an embodiment of the present invention; Figure 3 This is a comparison diagram of the calculation principles of different similarity measurements in embodiments of the present invention; Figure 4 This is a schematic diagram of the prediction results and measured data of the E-JITL algorithm in an embodiment of the present invention. Detailed Implementation

[0016] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0017] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0018] Example 1 This embodiment provides a method for rapid prediction of urban flooding disasters based on an integrated real-time learning algorithm. The modeling and computation processes of traditional global modeling methods and real-time learning algorithms are compared as follows: Figure 1 As shown, the E-JITL modeling framework calculation flowchart is as follows: Figure 2 As shown, please follow these steps: Step 1: Collect historical flood data as training samples for machine learning algorithms. If historical monitoring data is insufficient, construct a hydrological and hydrodynamic model to simulate urban flood data under different rainfall events and supplement the algorithm training set data. Step 1 is implemented in the following steps: Step 1.1: The system collects historical flood monitoring data within the forecast area, performs preprocessing operations such as screening, cleaning, and standardization on the raw data, and organizes it into a standardized dataset, which is used as the basic training samples for machine learning algorithms. The samples are represented as {( X i , Y i )}={( x 1, y 1), ( x 2, y 2),…, ( x i , y i )}.in, i Total historical sample size; x i For the first i Rainfall feature vectors of historical samples; y i For the first i Flood feature vectors of historical samples.

[0019] Step 1.2: Collect basic geographical and hydrological parameters of the forecast area, such as topography, pipeline distribution, land use type, and soil infiltration characteristics. Based on these parameters, construct a hydrological and hydrodynamic model suitable for the area. To compensate for the lack of historical monitoring data, design diversified rainfall processes covering different rainfall durations, rainfall intensities, and cumulative rainfall. Simulate the urban flood evolution process under various rainfall scenarios using the constructed high-precision hydrological and hydrodynamic model. Extract and organize the simulated flood data as a supplementary dataset for training samples, thereby solving the problem of insufficient training sample quantity and ultimately forming a complete E-JITL algorithm training dataset.

[0020] Step 2: Define the similarity measurement method used for sample selection. (If applicable) K There are several similarity measurement methods, which can be represented as follows: S 1, ..., S k And select a weighting function to quantify the importance of each measurement method.

[0021] Step 2 is implemented in the following steps: In the E-JITL modeling framework, three similarity measurement methods are used to select relevant samples, denoted as follows: S 1. S 2 and S 3. These methods each emphasize different sample selection strategies, fully utilizing the characteristics of various similarity measurement methods to improve prediction accuracy and robustness through local modeling. Different similarity measurement methods lead to the selection of different relevant samples for local modeling. The computational principles of different similarity measurement methods are compared below. Figure 3 As shown: Similarity measurement methods S 1. By calculating the Euclidean distance, the historical data with the closest distance is selected as the associated historical sample; similarity measurement method. S 2 in S Building upon point 1, this approach expands upon the existing method by incorporating angular correlation into the consideration of spatial distance between samples; similarity measurement methods... S 3. Based on latent variables related to the output, local modeling samples are selected.

[0022] The specific calculation process for each similarity measurement method is as follows: Step 2.1, Similarity Measurement Method S 1. Based on the most commonly used Euclidean distance, its calculation formula is: in, S 1,i It is a similarity measurement methodS 1 to the first i Similarity weights for each historical sample; dis i Indicates the first i The Euclidean distance between historical samples and rainfall prediction data; Δ is a diagonal matrix whose diagonal elements are the input variables. x The variance of each dimension; par 1 is an adjustable parameter used to modify the rate at which the similarity weight decreases with distance.

[0023] Step 2.2, Similarity Measurement Method S 2. Based on the joint definition of distance and angle, its calculation formula is: in, It is a trade-off parameter between 0 and 1; yes x q and x i The angle between them.

[0024] Step 2.3, Similarity Measurement Method S 3. Similarity is calculated in the supervised latent variable space by applying partial least squares (PLS) to the historical dataset. The distance calculation formula is: in, P and Q These are the projection matrices of the input and output spaces, respectively; E and F These are the residual matrices of the input and output data, respectively. T It is the score matrix of the historical dataset in the latent variable space, where the th... i The row corresponds to the first i Latent variables of a historical sample t i .

[0025] Step 3: Use Pearson correlation analysis to perform correlation analysis on the selected E-JITL algorithm input feature parameters, and select the feature parameters with larger Pearson correlation coefficients to ensure that there is a strong correlation between the selected feature parameters and the flood feature values ​​to be predicted in the simulation.

[0026] The specific steps are as follows: The Pearson correlation coefficient ranges from -1 to 1. Specifically, when the absolute value of the Pearson correlation coefficient is less than 0.4, the variables are weakly correlated; if the absolute value is between 0.4 and 0.6, the correlation is moderate; and when the absolute value is between 0.6 and 1, the correlation is strong.

[0027] To avoid decreased prediction accuracy due to weak correlation between input parameters, parameters with an absolute Pearson correlation coefficient greater than 0.6 are selected. The formula for calculating the Pearson correlation coefficient is as follows: in, For variables x, y The Pearson correlation coefficient between them; E ( x ), E ( y ), E ( xy ) are respectively x, y, xy The mathematical expectation; , Variables x, y variance; cov( x , y ) is a variable x, y The covariance.

[0028] Step 4: When rainfall forecast data... x q When a similarity occurs, the similarity index value for all historical data is calculated using each similarity measurement method. For each similarity measurement method... S k The similarity index value of historical samples is calculated as follows: By sorting the similarity index values ​​in descending order, the most relevant samples can be identified for modeling; S k Filtered n Each relevant input and output sample is represented as follows: The similarity index values ​​corresponding to these samples can be expressed as: .

[0029] Step 4 is implemented in the following steps: When rainfall forecast data x q When the similarity occurs, it is based on the three similarity measurement methods preset by the E-JITL modeling framework. S 1. S 2. S 3. Calculate all historical rainfall-related data and... x qThe similarity index values, through multi-dimensional similarity assessment, enable accurate screening of relevant samples, providing a high-quality data foundation for subsequent local modeling. Each similarity measurement method... S k ( k =1,2,3) Independently complete the similarity calculation and screening process of historical samples to ensure that the characteristics of different sample selection strategies are fully utilized.

[0030] The similarity index calculation and ranking process is as follows: Step 4.1, when rainfall forecast data x q When a similarity occurs, the similarity index value for all historical data is calculated using each similarity measurement method. For any similarity measurement method... S k ( k =1,2,3), No. i Historical samples and rainfall forecast data x q The general formula for calculating the similarity index value is: in, ω (·) represents the similarity calculation function corresponding to each similarity measurement method.

[0031] Complete the similarity measurement method respectively S 1. Similarity Measurement Methods (Based on Euclidean Distance) S 2. Similarity Measurement Methods (Based on Joint Definition of Distance and Angle) S Similarity calculation of historical datasets for three types of measurement methods, including 3 (similarity calculation based on PLS supervised latent variable space). For similarity measurement methods... S k The similarity index value of historical samples is calculated as follows: .

[0032] Step 4.2: After calculating the similarity index values ​​of each historical sample under the corresponding measurement method, screen relevant samples for modeling according to the following steps: First, for each type of... S k Corresponding similarity index set Sort the data in descending order to obtain the sorted index set: .in, ≥ ≥ ... ≥ , For the sorted number i The similarity index value of the positions. Secondly, take the first position after sorting. n One sample is used as the relevant sample for modeling, among which nThe preset threshold for the number of local modeling samples is used, and the optimal range of values ​​is determined through model accuracy verification.

[0033] Depend on S k Filtered n Each relevant input and output sample is represented as follows: The similarity index values ​​corresponding to these samples can be expressed as: ; Step 5: Use relevant historical samples and the corresponding similarity index value For rainfall forecast data x q Build a local online model y = f k ( x ).

[0034] Step 5 is implemented in the following steps: Assuming in similarity measurement methods S k The predicted output under the given conditions is Locally Weighted Partial Least Squares (LWPLS) is an improved algorithm based on weighted least squares and partial least squares. The core of the algorithm is the weighting of data points in regression analysis, with higher weights for closer points and lower weights for farther points. This utilizes local data to improve prediction accuracy, thus it is widely used in local modeling. This algorithm can effectively capture the nonlinear characteristics of urban flooding. Addressing the common problem of collinearity among input variables (which easily leads to singular prediction matrices), it reduces the dimensionality of the input data by mapping it to a latent space, transforming it into uncorrelated latent variables before prediction.

[0035] The LWPLS algorithm first weights each sample point, then performs partial least squares regression analysis using the weighted sample data. When calculating the partial least squares regression coefficients, the weight of each sample is used to adjust its influence in the model, so that samples closer to the target data contribute more to the model's prediction results. This local weighting method makes the model more flexible and better able to adapt to local features of the data, thereby improving generalization ability and prediction accuracy. Its calculation process is as follows: Step 5.1, set and Let be the input and output variable matrix of the training samples, where the _i_th ... i The line represents the first i Input of each sample x i and output yi First, the distance between the rainfall prediction data and the training samples is calculated using a distance calculation formula. Then, the sample weights are calculated using a Gaussian function. The calculation expression is: in, d i yes x q and x i The distance between them; par It is an adjustable parameter used to adjust the rate at which the weight decreases with distance.

[0036] Based on the weights calculated using the above formula Construct a diagonal matrix.

[0037] Step 5.2: Determine the total number of latent variables. K And index the latent variables k The initial value is set to 1.

[0038] Step 5.3: Calculate the weighted mean of the input and output matrices of the training sample set, expressed as follows: and .set up Obtain the input matrix. and output matrix And set the weighted output mean. .

[0039] Step 5.4, from the input matrix Extracting the first and second rainfall samples respectively from the input rainfall sample k One latent variable: in, It is the eigenvector corresponding to the largest eigenvalue. .

[0040] Step 5.5, calculate the first... k Loading vectors of 1 latent variable and the k A vector of regression coefficients : Step 5.6, use the first kThe predicted values ​​of the rainfall forecast data output updated by the four latent variables are as follows: Step 5.7: Check if the requirements are met. k = K If this condition is met, the final output of the rainfall forecast data is determined using the above formula; otherwise, proceed to step 5.8.

[0041] Step 5.8: Determine the input and output matrices for the training data, and the output vector for the rainfall prediction data: Step 5.9, Settings k = k +1, return to step 5.4 and repeat the subsequent calculation process until the condition in step 5.7 is met. k = K Then, output the result and stop the calculation.

[0042] Step 6: Integrate the predicted outputs obtained from each similarity measurement method using a weighted method; assuming It is S k Given the weighted dataset, the final ensemble prediction result is: .

[0043] Step 6 is implemented in the following steps: After completing the relevant sample screening in step 4 and the local model construction and prediction corresponding to each similarity measurement method in step 5, the process enters the integration stage of multi-local model prediction outputs. Through reasonable weight allocation and integration calculation, the prediction results under different similarity strategies are merged to improve the accuracy, robustness, and generalization ability of the final prediction output. The integration process uses each similarity measurement method... S k ( k Based on the local model prediction outputs corresponding to (=1,2,3), a weighted ensemble is performed to obtain the final prediction result.

[0044] Step 6.1: To quantify the reliability of the prediction results of each local model, a weight dataset is introduced. This weight dataset is directly related to the relevant samples and corresponding similarity indicators obtained by the methods in Step 5. The weight coefficients meet the normalization conditions to ensure that the weight allocation is reasonable and to avoid the excessive influence of a single sample on the prediction results.

[0045] Step 6.2: Calculate the credibility weight of each local model to measure the overall predictive reliability of the local models corresponding to different similarity methods. The calculation of the credibility weight comprehensively considers the goodness of fit and predictive stability of the local models, and is determined based on the cross-validation error: the smaller the average prediction error of a local model, the greater its credibility weight, indicating that its prediction results are more reliable, and it occupies a higher weight proportion in the integration process. Moreover, the sum of the credibility weights of all local models is 1, ensuring the additivity and interpretability of the weights.

[0046] Step 6.3: The final integrated prediction result is obtained by weighted summation, which involves weighting and fusing the prediction outputs of each local model according to their confidence weights. Assumptions It is S k Given the weighted dataset, the final ensemble prediction result is: .

[0047] Step 7 is implemented in the following steps: Based on the rolling updates of short-term rainfall forecast data, steps 4 to 6 are repeated sequentially to achieve rolling prediction of flood events in the forecast area and output rolling updated flood evolution forecast data.

[0048] The effects of the present invention are illustrated below with specific examples: A city has implemented a smart flood control platform system, centered on a rapid flood disaster prediction technology based on the E-JITL algorithm. By inputting short-term rainfall forecast data from the meteorological department, the system enables rapid prediction of flood disasters within its jurisdiction. The smart control platform system, relying on the rapid flood disaster prediction results provided by this invention, offers risk warnings and emergency response plans to relevant management departments.

[0049] Feature parameters with absolute Pearson correlation coefficients greater than 0.6 were selected to ensure a strong correlation between the chosen parameters and the flood characteristics to be predicted in the simulation, thus improving the performance and reliability of the prediction model. The correlation analysis results of the flood prediction input feature parameters are shown in Table 1. The next step is to construct an E-JITL prediction algorithm suitable for the study area based on the selected feature parameters and to predict urban flooding processes.

[0050] Table 1 To verify the performance of the E-JITL algorithm, it was compared with that based on a single similarity measurement. S 1. S 2. S JITL 3 S 1. JITL S 2. JITL S3. Algorithm Prediction Results. Algorithm parameters affect prediction performance. A random grid search method was used for optimization, and the final number of similar samples was determined to be 35 (to ensure accuracy and avoid overfitting). Similarity measurement method... S 1 and S 2 parameters par 1 and par Setting 2 to 1 indicates the similarity measurement method. S 2 parameters The similarity measurement method is 0.6. S 3 parameters par 3 is 2.

[0051] RMSLE , NSE and MAPE Indicators are used to evaluate the predictive performance of each method. RMSLE It is sensitive to wide-ranging, unevenly distributed data, making the assessment more robust. NSE When measuring data similarity, the closer the value is to 1, the higher the accuracy. MAPE The error is measured using absolute values ​​to avoid the cancellation of positive and negative values ​​and to accurately reflect the error range. Table 2 compares the prediction results of the four algorithms, and the comparison between the predicted waterlogging data and the measured data is shown in [Table 3]. Figure 4 (a) is JITL_ S 1, (b) is JITL_ S 2 and (c) are JITL_ S 3.

[0052] Table 2 From Table 2 and Figure 4 It is evident that the prediction performance of JITL algorithms based on a single similarity measurement is generally lower than that of E-JITL algorithms, because they struggle to fully capture the complex relationships between samples. Among these, JITL_ S 1 (Euclidean distance) performed the worst, predicting the area of ​​waterlogging. RMLSE =0.11、 MAPE =0.42 (high error) NSE =0.82 (the fit still has room for improvement); JITL_ S 2. By introducing angular information to enrich the similarity metric, the indicators were optimized to RMLSE=0.09, MAPE=0.31, and NSE=0.89, resulting in reduced error and improved fit; JITL_ S 3. By extracting latent space and removing irrelevant information through PLS, the performance index is further improved to... RMLSE =0.07、 MAPE =0.22、 NSE =0.92, the prediction accuracy is optimal.

[0053] The E-JITL algorithm, based on a weighted ensemble of the first three algorithms, performs best in predicting water accumulation area. RMLSE =0.06、 MAPE =0.14、 NSE =0.94; Prediction of water volume and depth RMLSE They are 0.02 and 0.01 respectively. MAPE The values ​​are 0.11 and 0.08 respectively. NSE All values ​​were above 0.98, demonstrating their accuracy and robustness in urban flood forecasting. Figure 4 The scatter plot shows that the E-JITL data points are more closely clustered near the diagonal, with smaller deviations from actual values ​​and high-precision hydrodynamic model simulation results.

[0054] The advantage of the E-JITL algorithm stems from its multi-similarity ensemble strategy, which integrates multi-dimensional features such as Euclidean distance and angle information to comprehensively capture the complex relationships between samples, thereby improving the algorithm's generalization ability and predictive performance in different scenarios.

[0055] This invention provides a rapid urban flood disaster prediction method based on the Integrated Just-In-Time Learning (E-JITL) algorithm. It employs multiple similarity measurement methods to select relevant historical samples for each rainfall forecast data point, addressing the inaccuracy of single similarity measurement methods. Through different similarity measurement methods, multiple sets of relevant samples are identified as training data, constructing multiple online local models. In the E-JITL algorithm, Local Weighted Partial Least Squares (LWPLS) is used as the local modeling algorithm to train the local models. Subsequently, the prediction outputs of each local model are fused using a weighted average to obtain the final prediction. Furthermore, a hydrodynamic model specific to the study area is constructed to address the problem of insufficient training samples. This hydrodynamic model is used to simulate flood data under various types of rainfall events, supplementing the training dataset for the E-JITL algorithm. Compared with traditional global models, this method enables real-time updates of model parameters to ensure good forecast performance, improving the adaptability and stability of the rapid urban flood forecasting model.

[0056] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for rapid prediction of urban flooding disasters based on an integrated real-time learning algorithm, characterized in that, Includes the following steps: Obtain a historical dataset, which includes historical rainfall feature vectors and corresponding historical flood feature vectors; Multiple similarity measurement methods are selected, and based on each similarity measurement method, relevant historical sample sets corresponding to the rainfall forecast data to be measured are selected from the historical dataset; Based on each of the selected historical sample sets, a corresponding local prediction model is constructed, and preliminary prediction results for the rainfall forecast data to be measured are obtained through each local prediction model. The multiple preliminary forecasts are weighted and integrated to obtain the final flood disaster forecast.

2. The method for rapid prediction of urban flooding disasters based on an integrated real-time learning algorithm according to claim 1, characterized in that, The process of obtaining the historical dataset includes: Historical flood monitoring data within the forecast area are collected to form a basic training sample set; a hydrological and hydrodynamic model of the forecast area is constructed, and flood data under different rainfall scenarios are simulated to obtain a supplementary training sample set; the historical dataset is composed of the basic training sample set and the supplementary training sample set.

3. The method for rapid prediction of urban flooding disasters based on an integrated real-time learning algorithm according to claim 1, characterized in that, The various similarity measurement methods include: a first similarity measurement method based on Euclidean distance; a second similarity measurement method based on a joint definition of distance and angle; and a third similarity measurement method based on partial least squares method to supervise the latent variable space.

4. The method for rapid prediction of urban flooding disasters based on an integrated real-time learning algorithm according to claim 1, characterized in that, The process of selecting relevant historical sample sets from the historical dataset based on each similarity measurement method includes: Calculate the similarity index value of each historical sample in the historical dataset relative to the rainfall forecast data to be measured using each similarity measurement method; The similarity index values ​​calculated by each similarity measurement method are sorted in descending order; A predetermined number of historical samples with the highest ranking are selected to form a relevant historical sample set corresponding to the similarity measurement method.

5. The method for rapid prediction of urban flooding disasters based on an integrated real-time learning algorithm according to claim 1, characterized in that, Before the screening, the process also includes: using Pearson correlation analysis to perform correlation analysis on the input feature parameters of the rainfall feature vector; selecting input feature parameters whose absolute value of the Pearson correlation coefficient with the flood feature vector is greater than a preset threshold for similarity calculation and local prediction model construction.

6. The method for rapid prediction of urban flooding based on an integrated real-time learning algorithm according to claim 1, characterized in that, The algorithm used to construct the local prediction model is the local weighted partial least squares algorithm.

7. The method for rapid prediction of urban flooding based on an integrated real-time learning algorithm according to claim 1, characterized in that, The process of obtaining preliminary prediction results using a local prediction model includes: Based on the input variable matrix and output variable matrix of the relevant historical sample set and the rainfall forecast data to be measured, a local weighted partial least squares regression calculation is performed to obtain the preliminary prediction results.

8. The method for rapid prediction of urban flooding disasters based on an integrated real-time learning algorithm according to claim 1, characterized in that, The process of weightedly integrating multiple preliminary predictions includes: A confidence weight is assigned to the preliminary prediction result corresponding to each of the local prediction models, and the confidence weight is determined based on the prediction error of the corresponding local prediction model; The final flood disaster prediction result is obtained by multiplying each preliminary prediction result by its corresponding confidence weight and then summing the results.

9. The method for rapid prediction of urban flooding disasters based on an integrated real-time learning algorithm according to claim 2, characterized in that, Simulate flood data under different rainfall scenarios, including designing rainfall processes covering different rainfall durations, rainfall intensities, and cumulative rainfall amounts, and using the aforementioned hydrological and hydrodynamic model for simulation.

Images