Urban flood simulation method fusing particle filter assimilation algorithm and machine learning model

Abstract

The invention belongs to the technical field of urban flood simulation, and particularly relates to an urban flood simulation method fusing a particle filter assimilation algorithm and a machine learning model, and the method comprises the following steps: (1) obtaining historical rainfall and flood data, and extracting rainfall characteristic indexes and ponding depth; (2) by taking the rainfall characteristic index as an input condition and the ponding depth as an output condition, constructing a GBDT model to simulate the rainfall process; (3) optimizing the GBDT model by adopting a PF algorithm, and dynamically correcting the ponding depth and GBDT model parameters to obtain a GBDT-PF model; and (4) simulating the rainfall process by adopting a GBDT-PF model. According to the method, the GBDT model is optimized by using the PF algorithm, dynamic correction of the urban flood model parameters and the ponding depth is realized, and the simulation performance of the urban flood ponding process is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of urban flood simulation, and in particular relates to an urban flood simulation method integrating a particle filter assimilation algorithm and a machine learning model. Background Art

[0002] Hydrological and hydrodynamic models are effective tools for simulating water flow processes and have been widely used in real-time flood forecasting, urban flood simulation and other fields. However, urban flood simulation models based on hydrological and hydrodynamic models require a large amount of basic data, and the software and time costs of calculation are high. In recent years, many scholars have proposed machine learning models. Common machine learning models include support vector machines (SVM), random forests (RF), gradient boosting decision trees (GBDT), etc. Machine learning models can mine effective information from a large amount of data and establish any nonlinear relationship between input and output information. The model has a fast calculation speed, avoiding the high time cost of hydrodynamic models, and has been successfully applied to many fields such as urban flood simulation and dam break flood simulation.

[0003] However, the existing urban flood simulation methods based on machine learning models, on the one hand, do not take into account the uncertainty of the output results caused by slight changes in model parameters. On the other hand, the rainfall-water accumulation process is a dynamic process, and constant model parameters make it difficult to accurately simulate the water accumulation depth at each moment.

[0004] Therefore, how to reduce the uncertainty of the output results caused by changes in model parameters while reducing the simulation time cost, and improve the simulation accuracy of the water depth at each moment, thereby improving the simulation performance of urban flooding and waterlogging processes, is a difficult problem that needs to be urgently solved in the field of urban flood simulation. Summary of the invention

[0005] In order to solve the above technical problems, the present invention proposes an urban flood simulation method integrating a particle filter assimilation algorithm and a machine learning model, which mainly includes the following steps:

[0006] Step (1), obtaining historical rainfall and flood data, extracting rainfall characteristic indicators and water depth;

[0007] Step (2), taking rainfall characteristic index as input condition and water accumulation depth as output condition, constructing GBDT model to simulate rainfall process;

[0008] Step (3), using the PF algorithm to optimize the GBDT model, dynamically correct the water depth and GBDT model parameters, and obtain the GBDT-PF model;

[0009] Step (4), using the GBDT-PF model to simulate the rainfall process.

[0010] Furthermore, in step (1), based on the measured historical rainfall and waterlogging data, the corresponding indicators of rainfall and waterlogging data are calculated by an equidistant splitting and recombination method to construct a rainfall characteristic indicator-waterlogging depth data set.

[0011] Furthermore, the rainfall characteristic indicators include rainfall duration, cumulative rainfall, rainfall peak, peak multiple ratio, rainfall intensity variance, concentration skewness and peak position coefficient.

[0012] Furthermore, the corresponding rainfall events were selected for verification simulation, and the simulation performance of the GBDT model was evaluated based on the RMSE and KGE indicators.

[0013] Furthermore, in the step (2), the Morris screening method is selected to screen the data assimilation analysis parameters of the GBDT model based on the sensitivity criteria.

[0014] Furthermore, in step (3), the control variable method is used to determine the reasonable values ​​of the number of particles and the standard deviation of the observation error in the particle filter algorithm.

[0015] Compared with the prior art, the beneficial effects of the present invention are at least as follows:

[0016] 1. The particle filter algorithm is used to optimize the gradient boosting model, which realizes the dynamic correction of urban flood model parameters and waterlogging depth, and improves the simulation performance of urban flood and waterlogging processes.

[0017] 2. Use the Morris screening method to identify the influence of each parameter in the model on the simulation results, and screen out the data assimilation analysis parameters of the GBDT model, thereby reducing the uncertainty of the output results caused by changes in model parameters. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 is a flow chart of the urban flood simulation method integrating the particle filter assimilation algorithm and the machine learning model of the present invention;

[0019] Figure 2 It is a schematic diagram of the gradient boosted decision tree (GBDT) of the present invention to construct a regression model;

[0020] Figure 3 It is a schematic diagram of the particle filtering algorithm of the present invention;

[0021] Figure 4 It is the particle filter algorithm and machine learning model fusion process of the embodiment of the present invention;

[0022] Figure 5 It is a comparison between the measured depth and the simulated depth of some observation points of the #15 rainfall event in the urban flood and waterlogging process simulation model constructed based on GBDT in an embodiment of the present invention;

[0023] Figure 6 is the sensitivity analysis result of the GBDT model parameters in the embodiment of the present invention;

[0024] Figure 7 is the assimilation performance of different numbers of particles and standard deviations of observation errors in the particle filter algorithm of the embodiment of the present invention;

[0025] Figure 8 It is a comparison of the simulation results of the GBDT-PF and GBDT models under the #15 rainfall event of an embodiment of the present invention;

[0026] Fig. 9 is a scatter comparison result of the simulated water depth and the measured water depth in three characteristic time periods of an embodiment of the present invention;

[0027] Fig.10 is the time-varying characteristic of the highly sensitive parameter in the GBDT model of the embodiment of the present invention;

[0028] Fig.11 This is the dynamic correction process of the water depth in the embodiment of the present invention. DETAILED DESCRIPTION

[0029] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0030] It is understood that the terms "first", "second", etc. used in this application may be used herein to describe various elements, but unless otherwise specified, these elements are not limited by these terms. These terms are only used to distinguish a first element from another element. For example, without departing from the scope of this application, a first xx script may be referred to as a second xx script, and similarly, a second xx script may be referred to as a first xx script.

[0031] The present invention provides Figure 1 The urban flood simulation method shown in FIG. 1 is a method integrating a particle filter assimilation algorithm and a machine learning model, and mainly includes the following steps:

[0032] Step (1), obtaining historical rainfall and flood data, extracting rainfall characteristic indicators and water depth;

[0033] Step (2), taking rainfall characteristic index as input condition and water accumulation depth as output condition, constructing GBDT model to simulate rainfall process;

[0034] Specifically, the GBDT model is a nonlinear model that requires building a new decision tree at each iteration and optimizing the decision tree in the direction of smaller residuals. The residuals are updated each time to effectively avoid overfitting. The GBDT regression model is constructed by combining the decision tree algorithm with the gradient boosting algorithm to establish the relationship between various rainfall characteristic indicators and water depth. Figure 2 To construct a schematic diagram of the GBDT regression model, the specific steps are as follows:

[0035] (1) Input the training data set, where X and Y represent the input and output variables, representing rainfall characteristic indicators and water depth respectively. Initialize the model with a constant value and minimize the following loss function:

[0036]

[0037] Where: F0(X) is the initialization learner; L(·) is the loss function; δ is the value of the output constant corresponding to the leaf node in the decision tree;

[0038] (2) Use gradient boosting to reduce model residuals: For m = 1:M, perform the following iterations to calculate the gradient of the loss function and set it as the residual, train a new decision tree to fit the residual, that is, use the training data set to train the decision tree:

[0039]

[0040] Where: is the negative gradient of the i-th sample in the m-th iteration, i.e., the residual; Y i is the actual value; F(X i ) is the predicted value;

[0041] (3) According to the principle of minimizing empirical risk, the decision tree is divided into J regions (i.e., the number of leaf nodes in the regression tree), the best fitting value of each region is calculated, and then the model is updated:

[0042]

[0043] Where: jm is the best fitting value for each leaf area; R jm is the region of leaf nodes corresponding to the mth classifier; j = 1, 2, 3, ..., J; J represents the number of leaf nodes; F m (X) is the learner updated m times iteratively; F m-1 (X) is the learner updated for m-1 iterations; X i is the input index of the i-th sample; I is the indicator function, which is related to the area where the sample is located; δ jm The product of the vector and I is the residual fitting value of the sample under this round of tree;

[0044] (4) The weak learners of each iteration are superimposed to obtain a strong learner, namely the GBDT regression model.

[0045]

[0046] Where: F(X) is the final strong learner; F0(X) is the initial learner; M is the total number of iterations;

[0047] Step (3), using the PF algorithm to optimize the GBDT model, dynamically correct the water depth and GBDT model parameters, and obtain the GBDT-PF model;

[0048] Specifically, the particle filter (PF) algorithm is developed based on Bayesian estimation, which is a non-parametric Monte Carlo method for approximate filtering tasks. It can gradually adjust the state variables according to the observed data to obtain the optimal value, thereby preventing the accumulation or transmission of errors and reducing uncertainty. The particle filter algorithm uses the distribution of weighted particles to fit the posterior probability distribution of the target variable. Figure 3 As shown in Figure 1, the framework mainly consists of five steps, namely initializing the particle set, prediction, observation, update, and resampling. The details are as follows:

[0049] (1) Initialize the particle set: randomly generate a set of particles according to the prior distribution.

[0050] (2) Prediction: The particle propagates backward based on the state transfer equation of the system to obtain the predicted value of the particle at each moment.

[0051]

[0052] Where: f(·) is the state transfer equation; u t is the control variable; is the state of the i-th particle at time t-1;

[0053] (3) Observation: Obtain the observation value z of the corresponding particle according to the observation equation t .

[0054] (4) Update: Calculate the weight of each particle based on its predicted value and observed value. The weight of a particle can be expressed as:

[0055]

[0056] Where: is the weight of the i-th particle at time t; is the weight of the i-th particle at time t-1; is the likelihood probability of the i-th particle at time t;

[0057] Normalized weights:

[0058]

[0059] Where: is the normalized weight; is the particle weight after sampling of the importance density function; is the sum of the weights of N particles;

[0060] The value of the particle weighted state variable at time t

[0061]

[0062] Where: is the optimal value after weighting; is the normalized weight; is the simulation value of the i-th particle at time t;

[0063] (5) Resampling: During the continuous iteration process, when the weights of some particles are almost 0 and the weights of some particles are large, particle degradation will occur. This not only wastes the computational workload, but also leads to the degradation of the particle filter assimilation performance, requiring particle resampling. The essence of resampling is to copy particles with large weights and remove particles with small weights.

[0064] Step (4), using the GBDT-PF model to simulate the rainfall process.

[0065] Specifically, for the GBDT urban flood simulation model, the idea of ​​coupling the particle filter algorithm with GBDT is proposed. For the trained GBDT model, a certain moment of rainfall is simulated to obtain Np particles containing parameters and state variables, and then corrected by observation data to obtain the best simulation of GBDT model parameters and water depth. These parameters are used for subsequent simulations of the waterlogging process and further correction of the simulation results. Figure 4 As shown, the specific process is as follows:

[0066] (1) Np particles are set at each water accumulation point. Each particle includes the water depth and GBDT high sensitivity parameters generated by random sampling. The weight of each particle is set to 1 / Np.

[0067] (2) Input the rainfall characteristic index and water depth of each waterlogging point training set into the GBDT model for training;

[0068] (3) Simulate the water depth at time t, and then update the Np particles at time t based on the observed water depth at time t to obtain the corresponding weight of each particle. The normal distribution function is used to represent the likelihood function in formula (7), and the specific weight calculation formula is as follows:

[0069]

[0070] Where: j is the number of waterlogging points; is the weight of the i-th particle at time t; is the observed water depth at time t; is the simulated water depth of the ith particle at time t; σ is the standard deviation of the observation error;

[0071] (4) Select the system resampling algorithm to resample the particles, remove the particles with smaller weights, copy the particles with larger weights, always keep the number of Np particles unchanged, and then normalize the particle weights;

[0072] (5) Calculate the optimal simulated water depth at time t:

[0073]

[0074] Where: is the optimal simulated water depth of the ith particle at time t; is the weight of the i-th particle at time t; is the simulated water depth of the ith particle at time t;

[0075] (6) Let t = t + 1, and return to step (3) to simulate and correct the water depth at the next moment, until the last moment T of the waterlogging process, complete the correction and obtain the best waterlogging process simulation result;

[0076] Furthermore, in step (1), based on the measured historical rainfall and waterlogging data, the corresponding indicators of rainfall and waterlogging data are calculated by an equidistant splitting and recombination method to construct a rainfall characteristic indicator-waterlogging depth data set.

[0077] Furthermore, the rainfall characteristic indicators include rainfall duration, cumulative rainfall, rainfall peak, peak multiple ratio, rainfall intensity variance, concentration skewness and peak position coefficient.

[0078] Furthermore, the simulation performance of the GBDT model was evaluated according to the RMSE and KGE indicators.

[0079] Specifically, the water depth simulation performance of the GBDT model and the fusion model (GBDT-PF) is evaluated by the root mean square error (RMSE) and Kling-Gupta efficiency (KGE). RMSE refers to the absolute error between the observed water depth and the simulated water depth, so the smaller the value, the better the simulation effect. KGE is used to verify the goodness of the model simulation results. KGE = 1 indicates that the simulated water depth is completely consistent with the observed water depth. RMSE and KGE are expressed as:

[0080]

[0081] Where: is the simulated water depth of GBDT and GBDT-PF models at time i; is the observed water depth at time i; T is the total waterlogging period; r is the linear relationship between the simulated water depth and the observed water depth sequence; (μ s ,σ s ) and (μ0,σ0) are the mean and standard deviation of the simulated and observed water depth series, respectively;

[0082] Furthermore, in the step (2), the Morris screening method is used to screen the data assimilation analysis parameters of the GBDT model based on the sensitivity criteria.

[0083] Specifically, the Morris screening method has high computational efficiency and strong operability, which helps to quickly identify parameters that have a greater impact on model results. When applying the Morris screening method for local sensitivity analysis, the parameter value is perturbed using a given change step size to obtain the value of the objective function under different parameter values, and finally the average of multiple Morris values ​​is taken as the parameter sensitivity value S. The specific calculation formula is as follows:

[0084]

[0085] Where: S is the parameter sensitivity value; D0 is the initial result of the model; D i is the model simulation result value after the parameter passes through the i-th time; P i is the change range of the parameter value from the initial value; n is the sample size of the parameter. Table 1 gives the evaluation criteria of parameter sensitivity corresponding to different sensitivity values.

[0086] Table 1 Morris screening method parameter sensitivity evaluation criteria

[0087]

[0088] Furthermore, in step (3), the control variable method is used to determine the reasonable values ​​of the number of particles and the standard deviation of the observation error in the particle filter algorithm.

[0089] Reference below Figure 5-Figure 11 The urban flood simulation research method of the present invention which integrates the particle filter assimilation algorithm and machine learning is used to analyze specific embodiments.

[0090] The data involved in this embodiment includes two parts, including rainfall data and water accumulation data.

[0091] Automatic rain gauges at 12 rain gauge stations provided by the Zhengzhou Meteorological Department recorded rainfall data. Sixteen rainfall events from 2014 to 2018 were selected as sample data for this study, with a temporal resolution of 10 minutes. The recording order of these rainfall events is as follows: 20140807, 20140830, 20140914, 20140917, 20150803, 20150826, 20150904, 20160714, 20160719, 20160825, 20170522, 20170718, 20170812, 20170825, 20180515, and 20180626. The above rainfall events are simplified as follows: #1, #2, … #16. On this basis, the data of the 12 rain gauges were divided into waterlogging points, and the Kriging spatial interpolation method was used to interpolate the location of the rainfall process corresponding to each waterlogging point. Based on the same method, the above steps were repeated to obtain the rainfall process of each waterlogging point in the 16 rainfall events.

[0092] Waterlogging data was obtained through road monitoring equipment, and flooding location and process data were provided by the Zhengzhou City Urban Management Department. Data from 15 waterlogging points were selected as sample data. The distribution of waterlogging points is mainly located in the central area of ​​Zhengzhou City. This may be mainly because the central and western parts of Zhengzhou City are Erqi District and Zhongyuan District, which are early construction areas. The design of the drainage network is difficult to cope with the current climate change, which makes these areas prone to waterlogging. The eastern area is a new development zone under construction, and some facilities may not keep up, resulting in waterlogging.

[0093] The 180-minute rainfall process is split into 18 groups of data with a time resolution of 10 minutes by using the method of equidistant splitting and reorganization, and then reorganized into 18 sub-rainfall processes (0-10 minutes, 0-20 minutes, ..., 0-180 minutes). Finally, the indicators of each group of rainfall processes are calculated. Table 2 lists the sample set of 16 rainfall events, with a total of 10,080 groups of data and 80,640 indicators, without any missing data. Among them, seven rainfall characteristic indicators are used as input variables, and water depth is used as output variable.

[0094] Table 2 Rainfall characteristic index and waterlogging depth sample set

[0095]

[0096] In the machine learning model, the value of the model parameters has a great influence on the model structure and simulation performance. For the GBDT model, there are framework parameters and weak learner parameters in the model parameters. Setting too large or too small is not conducive to model training, and relatively reasonable parameters need to be set during the model construction process. This embodiment refers to the corresponding literature, and taking into account the number of samples and features, as well as the specific requirements of different model parameters, seven model parameters are finally selected for setting, namely: the number of regression trees, learning rate, subsampling rate, regression tree depth, minimum number of sample partitions, minimum number of sample leaves, and maximum sample features. Table 3 gives the value of each model parameter.

[0097] Table 3 GBDT model parameter value settings

[0098]

[0099] For the 16 rainfall events with a total of 10080 sets of data, the first 13 with a total of 8190 sets of data were selected as training data, and the remaining 1890 sets of rainfall data (#14, #15 and #16) were selected as test data. Table 4 shows the model performance of the GBDT model tested using three rainfall events. Overall, the average values ​​of RMSE and KGE were 0.039 and 0.918, respectively, indicating that it is reasonable and feasible to choose the GBDT model to simulate the rainfall-water accumulation process. However, there are differences in the simulation performance of different rainfall events, for example, the KGE of #15 is 0.876 (<0.9), and the simulation accuracy needs to be further improved.

[0100] Table 4 Simulation performance of waterlogging process under different rainfall events

[0101]

[0102] In order to intuitively reflect the simulation effect of the GBDT model on the waterlogging process, the embodiment selected the #15 rainfall event and extracted some waterlogging points with poor simulation (such as Figure 5 As shown in the figure, the shortcomings of GBDT in simulating the waterlogging process are consistent at different waterlogging points, that is, the simulation accuracy is good during the water level rise period (100-200 minutes), but poor during the water level peak (200-300 minutes) and water level drop period (300-400 minutes). The above results show that when constructing the GBDT model, the simulation performance of different rainfall events and different waterlogging periods is different when the parameter values ​​are fixed. Therefore, it is necessary to dynamically adjust the parameters to improve the accuracy of the model simulation.

[0103] The identification and selection of highly sensitive parameters in the model helps to improve the assimilation efficiency and reduce the uncertainty of the simulation results. In this embodiment, since the maximum sample feature is set to None, the remaining six parameters are considered when building the model. Table 5 shows the perturbation range and change step size of the GBDT parameter sensitivity analysis. When a parameter is perturbed, the initial values ​​of the remaining parameters remain unchanged.

[0104] Table 5 GBDT model parameter sensitivity analysis disturbance settings

[0105]

[0106]

[0107] Figure 6 The sensitivity values ​​of different parameters under three rainfall events are shown, which are obtained by calculating the average sensitivity value of the parameters at each waterlogging point. The sensitivity of the parameters under different rainfall events is different, but the minimum number of sample divisions is significantly weaker than other parameters. In addition, according to the sensitivity evaluation criteria (Table 1), the average sensitivity values ​​of the number of regression trees, learning rate, subsampling rate, and regression tree depth exceed 1, which are high-sensitivity parameters, indicating that changes in values ​​have a significant impact on the simulation results. Therefore, in the data assimilation step, consider selecting these four high-sensitivity parameters as assimilation parameters for particle filtering.

[0108] The number of particles and the standard deviation of the observation error will significantly affect the assimilation performance of the particle filter algorithm. In order to determine the appropriate values ​​of the two parameters, the number of particles and the standard deviation of the observation error are changed within a certain range. Table 6 provides the preset ranges of the number of particles and the standard deviation of the observation error. When the value of one parameter is analyzed, the value of the other parameter remains unchanged.

[0109] Table 6 Preset ranges of particle number and observation error standard deviation

[0110]

[0111] Figure 7The changes in average KGE under different observation error standard deviations and particle numbers are shown respectively. For the observation error standard deviation, the average KGE increases as the observation error standard deviation decreases, but slowly increases and stabilizes when the value is 0.01, and the average KGE is 0.913 at this time. On the contrary, the average KGE increases with the increase in the number of particles. After the number of particles is greater than 100, the average KGE increases slowly and remains stable. Similar to the application of the hydrodynamic model, the running time is also given. It can be seen from the running time that the time cost of the machine learning model is low. It is worth noting that the running time shows a linear increase trend, but when the number of particles is greater than 100, the increase in the number of particles leads to an uneven increase in KGE and time cost. Therefore, after comprehensively considering the impact of the number of particles and the observation error standard deviation on the assimilation performance and running time, the embodiment sets the number of particles and the observation error standard deviation to 100 and 0.01, respectively.

[0112] The water depths and parameters of waterlogging points with poor simulation accuracy in the GBDT model were assimilated. Figure 8 The assimilation results of the waterlogging process under the #15 rainfall event are given. After fusing the particle filter algorithm, the changing trend of the simulated waterlogging process is generally consistent with the observed waterlogging process, and is closer to the observed waterlogging process. The simulation performance is significantly better than the waterlogging process simulated by a single GBDT model. Table 7 shows the evaluation results of the GBDT and GBDT-PF simulations of the three rainfall events. By simulating each waterlogging point with the GBDT-PF model, the RMSE is significantly reduced, and the KGE is increased to 0.9. In addition, the simulation accuracy of the waterlogging points that originally performed well in the simulation has been further improved, and for the waterlogging points that performed poorly in the simulation, the improvement effect is significant. For example, the KGE of J6 has increased by an average of 15.06%.

[0113] Table 7 Comparison of simulation performance between GBDT and GBDT-PF models

[0114]

[0115]

[0116] Fig. 9 Comparison diagrams of simulated water depth and observed water depth of GBDT and GBDT-PF models in different time periods are given. For the three characteristic time periods, the scattered points of GBDT-PF model are closer to the 1:1 line and more concentrated, and the simulation performance of GBDT-PF model is significantly better than that of GBDT model, which shows that the fusion particle filter algorithm avoids using the same model parameters for each waterlogging point simulation in the entire waterlogging process, dynamically adjusts the model parameters, generates the optimal simulation value at each moment, and improves the simulation accuracy of the entire time period.

[0117] Take the #14 rainfall event as an example. Fig.10The dynamic correction process of different parameters is given. The optimal estimated values ​​of the parameters at different waterlogging points have time-varying characteristics, and the change characteristics are inconsistent. In general, the change amplitude of the learning rate is higher than that of the other three parameters, which is consistent with the results of parameter sensitivity analysis. On the time scale, before 100 minutes, the changes of the four parameters are small and basically remain stable, indicating that the simulation results of the original fixed parameters are better. The optimal estimates of the parameters fluctuate on the subsequent time scales. J6 has a poor effect in the GBDT simulation (RMSE = 0.035, KGE = 0.817), while the accuracy is significantly improved after GBDT-PF simulation (RMSE = 0.010, KGE = 0.953), so the change amplitude of the four parameters is higher than that of the remaining waterlogging points. The above parameter changes show that the fusion particle filter algorithm can effectively improve the simulation accuracy of the waterlogging process by dynamically adjusting the model parameters over time.

[0118] Fig.11 The dynamic correction process of the simulated water depth at the J5 waterlogging point under the #14 rainfall event and the predicted water depth process based on time-varying parameters are shown. As can be seen from the figure, there is a certain gap between the predicted water depth process of the parameters in the early stage of the waterlogging process and the observed process. However, as the correction process progresses, the prediction period gradually shortens, the parameters are continuously adjusted dynamically, and the subsequent prediction results are close to the actual prediction results. The accuracy of the simulation results after correction by the particle filter algorithm is significantly improved.

[0119] This application proposes a comprehensive framework that integrates the GBDT model and the particle filter algorithm to improve the simulation accuracy of urban flood and waterlogging processes. Taking the main urban area of ​​Zhengzhou City, Henan Province, China as a model application case, the choice of the GBDT model is reasonable and feasible for the simulation of the waterlogging process, but the use of fixed parameter simulation leads to a large overall simulation error of the waterlogging process. In contrast, the GBDT-PF model dynamically adjusts highly sensitive parameters and corrects the water depth simulation process, effectively improving the simulation accuracy. Under different rainfall events, the average reduction rate of RMSE is 57.9%, and KGE is increased to 0.9. In addition, the simulation accuracy of different waterlogging periods is significantly improved. The time-varying characteristics of the parameters of the GBDT-PF model and the dynamic prediction and correction process show that the fusion particle filter algorithm effectively avoids the errors caused by the fixed parameters used in the subsequent rainfall event simulation, and meets the accuracy requirements of urban flood simulation. The proposed model has higher potential in improving the performance of urban flood simulation, prediction and defense.

[0120] It should be understood that, although each step in the flow chart of each embodiment of the present invention is shown in sequence according to the indication of the arrow, these steps are not necessarily performed in sequence according to the order indicated by the arrow. Unless there is a clear explanation in this article, the execution of these steps does not have a strict order restriction, and these steps can be performed in other orders. Moreover, at least a portion of the steps in each embodiment may include a plurality of sub-steps or a plurality of stages, and these sub-steps or stages are not necessarily performed at the same time, but can be performed at different times, and the execution order of these sub-steps or stages is not necessarily performed in sequence, but can be performed in turn or alternately with at least a portion of other steps or sub-steps or stages of other steps.

[0121] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program. The above-mentioned program can be stored in a non-volatile computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0122] The technical features of the above-mentioned embodiments can be combined arbitrarily. In order to make the description concise, not all possible combinations of the technical features in the above-mentioned embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0123] The above-mentioned embodiments only express several implementation methods of the present invention, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the patent of the present invention. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.

[0124] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for simulating urban floods by integrating a particle filter assimilation algorithm and a machine learning model, characterized in that: The following steps are involved: Step (1), obtaining historical rainfall and flood data, extracting rainfall characteristic indicators and water depth; Step (2), taking rainfall characteristic index as input condition and water accumulation depth as output condition, constructing GBDT model to simulate rainfall process; Step (3), using the PF algorithm to optimize the GBDT model, dynamically correct the water depth and GBDT model parameters, and obtain the GBDT-PF model; Step (4), using the GBDT-PF model to simulate the rainfall process.

2. The urban flood simulation method integrating the particle filter assimilation algorithm and the machine learning model according to claim 1 is characterized in that: In the step (1), based on the measured historical rainfall and waterlogging data, the corresponding indicators of rainfall and waterlogging data are calculated by an equidistant splitting and reorganization method to construct a rainfall characteristic indicator-waterlogging depth data set.

3. The urban flood simulation method integrating the particle filter assimilation algorithm and the machine learning model according to claim 1 is characterized in that: The rainfall characteristic indicators include rainfall duration, cumulative rainfall, rainfall peak, peak ratio, rainfall intensity variance, concentration skewness and peak position coefficient.

4. The urban flood simulation method integrating the particle filter assimilation algorithm and the machine learning model according to claim 1 is characterized in that: The simulation performance of the GBDT model was evaluated based on the RMSE and KGE indicators.

5. The urban flood simulation method integrating the particle filter assimilation algorithm and the machine learning model according to claim 1 is characterized in that: In the step (2), the Morris screening method is used to screen the data assimilation analysis parameters of the GBDT model based on the sensitivity criteria.

6. The urban flood simulation method integrating the particle filter assimilation algorithm and the machine learning model according to claim 1 is characterized in that: In the step (3), the control variable method is used to determine the reasonable values ​​of the number of particles and the standard deviation of the observation error in the particle filter algorithm.

7. The urban flood simulation method integrating the particle filter assimilation algorithm and the machine learning model according to claim 1 is characterized in that: The specific process of simulating the rainfall process using the GBDT-PF model is as follows: Step (4-1), Np particles are set at each water accumulation point, each particle includes the water depth and GBDT high sensitivity parameters generated by random sampling; the weight of each particle is set to 1 / Np; Step (4-2), input the rainfall characteristic index and water accumulation process of each water accumulation point training set into the GBDT model for training; Step (4-3), simulate the water depth at time t, and then update the Np particles at time t based on the observed water depth at time t to obtain the corresponding weight of each particle; the specific weight calculation formula is as follows: Where: j is the number of waterlogging points; is the weight of the i-th particle at time t; is the observed water depth at time t; is the simulated water depth of the ith particle at time t; σ is the standard deviation of the observation error; Step (4-4), select the system resampling algorithm to resample the particles, remove the particles with smaller weights, copy the particles with larger weights, always keep the number of Np particles unchanged, and then normalize the particle weights; Step (4-5), calculate the optimal simulated water depth at time t: Where: is the optimal simulated water depth of the ith particle at time t; is the weight of the i-th particle at time t; is the simulated water depth of the ith particle at time t; Step (4-6), let t=t+1, and return to step (4-3) to simulate and correct the water depth at the next moment, until the last moment T of the water accumulation process, complete the correction and obtain the best water accumulation process simulation result.

Images