A classification-based stochastic flood forecasting method based on machine learning and cloud model

Abstract

This invention discloses a random flood forecasting method based on machine learning and cloud models. The steps are as follows: using typical historical floods as basic data for model calibration and testing, selecting classification indicators that meet the conditions, and classifying historical floods based on a self-organizing map neural network (SOM); using the maximum information coefficient method (MIC) to screen the impact factors of classified floods, and establishing a random flood forecasting model based on different machine learning methods; for different types of floods, solving the fusion weights of different forecasting models based on the cloud model, and weighting the simulation results of each model to obtain the model integrated forecast result; analyzing and calculating the relative forecast error, and establishing the joint distribution function of the relative forecast errors at adjacent moments based on the Copula method; finally, obtaining real-time online flood information to achieve random flood forecasting. This invention can improve the accuracy of flood forecasting and provide a new approach to hydrological prediction and forecasting.

Description

Technical Field

[0001] The present invention belongs to the technical field of hydrological forecasting, and in particular relates to a classified flood stochastic forecasting method based on machine learning and cloud models. Background Art

[0002] Accurate and timely flood forecasting facilitates the development of scientifically sound water conservancy project scheduling plans, ensuring water security in river basins and possessing significant economic and social benefits. In recent years, machine learning has garnered significant attention in the field of hydrological forecasting due to its strong nonlinear fitting capabilities and simple model construction. Using machine learning models, such as artificial neural networks, as hydrological forecasting models has become a stable and effective forecasting tool. However, given the significant impact of structural uncertainty in individual models on forecast results, scholars both domestically and internationally have proposed the concept of multi-model ensemble forecasting. These methods primarily utilize methods such as simple averaging, neural network models, and Bayesian model averaging to integrate forecast results from multiple models and improve overall forecast accuracy. However, these methods primarily calculate fusion weights based on model simulation accuracy and fail to consider the similarity between model distributions. In principle, forecast errors are inevitable, regardless of how hydrological forecasting technology improves. However, most current stochastic simulations with uncertainty fail to consider the correlation between errors at preceding and following forecast times. Summary of the Invention

[0003] In view of the uncertainty of forecast results from different models and the defects that the flood forecasting process that generally considers forecast errors does not consider the correlation before and after the forecast time, the present invention proposes a classified flood stochastic forecasting method based on machine learning and cloud model, which can improve the accuracy of flood forecasting.

[0004] The present invention adopts the following technical solutions:

[0005] A classification flood stochastic forecasting method based on machine learning and cloud model includes the following steps:

[0006] Step 1: Using typical historical flood events as the basic data for model calibration and testing, select classification indicators that meet the conditions and classify historical flood events based on the self-organizing map neural network (SOM);

[0007] Step 2: Use the maximum information coefficient (MIC) method to screen the influencing factors of classified floods, and establish classified flood forecasting models based on different machine learning methods to obtain the optimal parameters of different models corresponding to various flood types;

[0008] Step 3: For different types of floods, the fusion weights of different forecast models are solved based on the cloud model, and the simulation results of each model are weighted to obtain the integrated forecast results of the classified flood model;

[0009] Step 4: Analyze and calculate the relative forecast error, establish the joint distribution function of the relative forecast errors at adjacent moments based on the Copula method, and determine the cumulative probability distribution function of the relative forecast error;

[0010] Step 5: Obtain real-time flood information online and implement random flood forecasting.

[0011] In the above technical solution, further, the classification indicators in Step 1 include one or more of total rainfall, maximum three-hour rainfall, rainfall intensity, rainfall center, previous impact rainfall and flood peak.

[0012] Furthermore, in Step 2, the maximum information coefficient method MIC is used to screen the influencing factors, including the following methods: for different types of floods, the rainfall P predicted at time t is used as the t And the previous rainfall t-N ,P t-N+1 ,P t-N+2 ,…,P t-1} is used as a set of candidate prediction factors, and the current flow rate Q is calculated based on the maximum information coefficient method MIC t Correlations (MIC values) with different candidate predictors were used, and those with MIC values ​​greater than 0.30 were selected as the inputs of the prediction model;

[0013] Furthermore, the Step 2 establishes a classification flood forecasting model based on different machine learning methods, including the following steps:

[0014] Step 2-1: Divide the flood series into a training period and a validation period according to the number of floods. Use the candidate predictors with MIC values ​​greater than 0.30 obtained in Step 2 as model inputs, and the number of floods as model outputs. All data sets are normalized to the interval (0, 1).

[0015] Step 2-2: Establish a forecast model based on different machine learning methods {M1, M2, ..., M K}(K≥2), output the model results, and obtain the simulated runoff value after denormalization in, represents the simulated value of the i-th forecast model at time t, i = 1, 2, …, K, K is the number of forecast models, and T is the duration of different flood events (h);

[0016] Step 2-3: Compare the secondary flood observations in the training set in, represents the observation value at time t, and selects the correlation coefficient (R), Nash coefficient (NSE), mean absolute error (MAE) and root mean square error (RMSE) as evaluation indicators. The grid method is used to optimize the parameters of different machine learning models;

[0017] Step 2-4: Input the validation set data into the trained model to obtain the simulation results, and then obtain the runoff simulation value after denormalization. The correlation coefficient, Nash coefficient, mean absolute error and root mean square error are used as evaluation indicators to judge whether the simulation results of the model are qualified.

[0018] Furthermore, the Step 2-1 machine learning method includes two or more of a neural network model, a support vector machine model, a long short-term memory network model, a gated recurrent unit model, and an extreme gradient boosting tree.

[0019] Furthermore, in Step 3, the fusion weights of different forecast models are solved based on the cloud model, and the simulation results of each model are weighted to obtain the model integration forecast result, which includes the following steps:

[0020] Step 3-1: Select the observed and simulated values ​​during the training period, and establish the corresponding cloud model pair (Cq sim,i ,Cq obs ), converting the distribution characteristics of the sequence into corresponding digital features, where Cq sim,i represents the cloud model expectation curve constructed by the simulated value of the i-th forecast model, Cq obs The expected curve of the cloud model constructed from the observations;

[0021] Step 3-2: Calculate the simulation values ​​of different forecast models based on the overlap of cloud model expectation curves and observations The global similarity S i (Q sim,i ,Q obs ):

[0022] S i (Q sim,i ,Q obs )=sim(Cq sim,i ,Cq obs ) (1)

[0023] Step 3-3: Based on the global similarity S i (Q sim,i ,Q obs ) Solve the fusion weight w of each forecast model i :

[0024]

[0025] Step 3-4: Perform weighted combination of the simulation results of each model to obtain the model ensemble prediction result:

[0026]

[0027] Where, It represents the simulated value of the i-th forecast model at the t-th moment, i = 1, 2, …, K, K is the number of forecast models, K ≥ 2.

[0028] Furthermore, the correlation coefficient (R), Nash coefficient (NSE), mean absolute error (MAE) and root mean square error (RMSE) of the model ensemble prediction results of Step 3-4 are selected as evaluation indicators to determine whether the ensemble prediction results are qualified.

[0029] Furthermore, the Step 4 analyzes and calculates the relative forecast error, and establishes a joint distribution function of the relative forecast errors at adjacent moments based on the Copula method, including the following steps:

[0030] Step 4-1: Assume that the relative error of runoff forecast is The relative error distribution of runoff forecast is constructed as the marginal distribution, and the joint distribution of the previous and next moments is established based on different Copula functions. The maximum likelihood estimation method is used to solve the unknown parameters in the joint distribution.

[0031] Step 4-2: Use the Akaike Information Criterion (AIC) to evaluate the obtained joint distribution function, and select the copula function with the smallest AIC value as the final joint distribution function;

[0032] Furthermore, the Copula function used in Step 4-1 includes but is not limited to Gumbel, Clayton and Frank Copula functions;

[0033] Furthermore, in Step 5, online real-time flood information is obtained to implement random flood forecasting, which includes the following steps:

[0034] Step 5-1: Obtain real-time online flood information, extract classification factor data from the real-time rainfall information, and perform online flood classification based on the self-organizing map neural network (SOM) established in Step 1 to determine the category to which the flood belongs;

[0035] Step 5-2: Select the optimal model parameters corresponding to this type of flood obtained in Step 2 and the fusion weights obtained by the method in Step 3 to perform deterministic integrated forecasting;

[0036] Step 5-3: Use the Gibbs sampling method to perform random simulation of the forecast error, couple the deterministic forecast results, and realize the random flood forecast interval.

[0037] Furthermore, Step 5-3 uses the Gibbs sampling method to perform random simulation of the forecast error at the forecast time, including the following steps:

[0038] Step 5-3-1 Generate two random numbers α0, α1∈(0,1), where α0 and α1 are probability values;

[0039] Step 5-3-2 Based on conditional probability distribution Where X1 represents the relative forecast error at the current moment, X0 represents the relative forecast error 1 hour before the hypothetical flood forecast moment, and C is the joint distribution function of the copula relative forecast errors at adjacent moments, obtained based on the method in Step 4. Substituting α0 and α1 into the conditional probability distribution, we obtain x1, which is the cumulative probability distribution value of the relative forecast error at the current forecast moment.

[0040] Step 5-3-3 Randomly generate α2, α3, …, α T ∈(0,1), solve the following system of equations as follows:

[0041]

[0042] The relative error cumulative probability distribution values ​​x2, x3...x at time 2 to T can be obtained. T ;

[0043] Step 5-3-4 Repeat Step 5-3-3 N times to obtain N groups of cumulative probability distribution values ​​of relative forecast errors.

[0044] Step 5-3-5 Based on the obtained cumulative probability distribution function of relative forecast error, the inverse function of the joint distribution function of relative forecast errors at adjacent moments established by the Step 4 method is reversed to obtain N groups of relative forecast errors. The relative forecast error is coupled with the determined integrated forecast result to realize random flood forecasting.

[0045] Compared with the prior art, the present invention has the following beneficial effects:

[0046] (1) Using multiple hydrological forecast models based on machine learning methods for ensemble forecasting can avoid the risk of single model forecasting and improve forecast accuracy;

[0047] (2) Using cloud models to perform weighted averaging of different forecast models can compensate for the uncertainty defects of model forecasts;

[0048] (3) The random simulation of forecast errors based on Copula multivariate functions takes into account the mutuality of different time periods and can improve the accuracy of random forecasts. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 is a flow chart of the method of the present invention;

[0050] Figure 2 It is a specific flow chart of Step 5 in the present invention;

[0051] Figure 3 This is the C1 flood forecast result map in the implementation case;

[0052] Figure 4 This is the C2 flood forecast result map in the implementation case;

[0053] Figure 5 This is a graph of random flood forecast results in the implementation case. DETAILED DESCRIPTION

[0054] The present invention is further illustrated below with reference to specific examples. It should be understood that these examples are only used to illustrate the present invention and are not used to limit the scope of the present invention. After reading the present invention, modifications of various equivalent forms of the present invention made by those skilled in the art all fall within the scope defined by the claims attached to this application.

[0055] like Figure 1 As shown, a classification flood stochastic forecasting method based on machine learning and cloud model includes the following steps:

[0056] Step 1: Using typical historical flood events as the basic data for model calibration and testing, select classification indicators that meet the conditions and classify historical flood events based on the self-organizing map neural network (SOM);

[0057] Step 2: Use the maximum information coefficient method (MIC) to screen the influencing factors of classified floods, and establish classified flood forecasting models based on different machine learning methods to obtain the optimal parameters of different models corresponding to various flood types;

[0058] Step 3: For different types of floods, the fusion weights of different forecast models are solved based on the cloud model, and the simulation results of each model are weighted to obtain the integrated forecast results of the classified flood model;

[0059] Step 4: Analyze and calculate the relative forecast error, establish the joint distribution function of the relative forecast errors at adjacent moments based on the Copula method, and determine the cumulative probability distribution function of the relative forecast error;

[0060] Step 5: Obtain real-time flood information online and implement random flood forecasting.

[0061] The classification indicators in Step 1 include but are not limited to total rainfall, maximum three-hour rainfall, rainfall intensity, rainfall center, previous impact rainfall and flood peak.

[0062] In Step 2, the maximum information coefficient method MIC is used to screen the influencing factors: for different types of floods, the rainfall P at time t is predicted. t And the previous rainfall t-N ,P t-N+1 ,P t-N+2 ,…,P t-1} is used as a set of candidate prediction factors, and the current flow rate Q is calculated based on the maximum information coefficient method MIC t Correlations (MIC values) with different candidate predictors were used, and those with MIC values ​​greater than 0.30 were selected as the inputs of the prediction model;

[0063] Step 2 establishes a classification flood forecasting model based on different machine learning methods, including the following steps:

[0064] Step 2-1: Divide the flood series into a training period and a validation period according to the number of floods. Use the candidate predictors with MIC values ​​greater than 0.30 obtained in Step 2 as model inputs, and the number of floods as model outputs. All data sets are normalized to the interval (0, 1).

[0065] Step 2-2: Establish a forecast model based on different machine learning methods {M1, M2, ...M i ,…M K}(K≥2), output the model results, and obtain the simulated runoff value after denormalization in, represents the simulated value of the i-th forecast model at time t, i = 1, 2, …, K, K is the number of forecast models, and T is the duration of different flood events (h);

[0066] Step 2-3: Compare secondary flood observations in, represents the observation value at time t, and the grid method is used to optimize the parameters of different machine learning models;

[0067] Step 2-4 Input the validation set data into the trained model to obtain the simulation results, and then obtain the runoff value after denormalization.

[0068] The Step 2-1 machine learning method includes two or more of a neural network model, a support vector machine model, a long short-term memory network model, a gated recurrent unit model, and an extreme gradient boosting tree.

[0069] The Step 2-3 model parameters are optimized, and the correlation coefficient (R), Nash coefficient (NSE), mean absolute error (MAE) and root mean square error (RMSE) are selected as evaluation indicators.

[0070] In Step 3, the fusion weights of different forecast models are solved based on the cloud model, and the simulation results of each model are weighted to obtain the model integration forecast result, which includes the following steps:

[0071] Step 3-1: Select the observed and simulated values ​​during the training period, and establish the corresponding cloud model pair (Cq sim,i ,Cq obs ), converting the distribution characteristics of the sequence into corresponding digital features, where Cq sim,i represents the cloud model expectation curve constructed by the simulated value of the i-th forecast model, Cq obs The expected curve of the cloud model constructed from the observations;

[0072] Step 3-2: Calculate the simulated values ​​Q of different forecast models based on the overlap of the expected curve of the cloud model sim,i and observation value Q obs The global similarity S i (Q sim,i ,Q obs ):

[0073] S i (Q sim,i ,Q obs )=sim(Cq sim,i ,Cq obs ) (1)

[0074] Step 3-3: Based on the global similarity S i (Q sim,i ,Q obs ) Solve the fusion weight w of each forecast model i :

[0075]

[0076] Step 3-4: Perform weighted combination of the simulation results of each model to obtain the model ensemble prediction result:

[0077]

[0078] Where, It represents the simulated value of the i-th forecast model at the t-th moment, i = 1, 2, …, K, K is the number of forecast models, K ≥ 2.

[0079] The integrated prediction results of Step 3-4 select correlation coefficient (R), Nash coefficient (NSE), mean absolute error (MAE) and root mean square error (RMSE) as evaluation indicators to judge whether the model integrated prediction results are qualified.

[0080] The step 4 analyzes and calculates the forecast error, and establishes a joint distribution function of the relative forecast errors at adjacent moments based on the Copula method, including the following steps:

[0081] Step 4-1: Assume that the relative runoff forecast error is The relative runoff forecast error distribution is constructed as the marginal distribution, and the joint distribution of the previous and next moments is established based on different Copula functions. The maximum likelihood estimation method is used to solve the unknown parameters in the joint distribution.

[0082] Step 4-2: Use the Akaike Information Criterion (AIC) to evaluate the obtained joint distribution function, and select the copula function with the smallest AIC value as the final joint distribution function;

[0083] The Copula function used in Step 4-1 includes but is not limited to Gumbel, Clayton and Frank Copula functions;

[0084] like Figure 2 As shown, Step 5 obtains online real-time flood information and implements random flood forecasting, which includes the following steps:

[0085] Step 5-1: Obtain real-time online flood information, extract classification factor data from real-time rainfall information, and determine the category of the flood based on the self-organizing map neural network (SOM) online flood classification established in Step 1;

[0086] Step 5-2: Select the optimal model parameters corresponding to this type of flood obtained in Step 2 and the fusion weights obtained in Step 3 to perform deterministic integrated forecasting;

[0087] Step 5-3: Use the Gibbs sampling method to randomly simulate the relative forecast error, couple the simulated relative forecast error with the deterministic integrated forecast result to realize the random flood forecast interval.

[0088] Furthermore, Step 5-3 uses the Gibbs sampling method to perform random simulation of the forecast error at the forecast time, including the following steps:

[0089] Step 5-3-1 Generate two random numbers α0, α1∈(0,1), where α0 and α1 are probability values;

[0090] Step 5-3-2 Based on conditional probability distribution Where X1 represents the relative forecast error at the current moment, X0 represents the relative forecast error 1 hour before the hypothetical flood forecast moment, and C is the joint distribution function of the copula relative forecast errors at adjacent moments, obtained based on the method in Step 4. Substituting α0 and α1 into the conditional probability distribution, we obtain x1, which is the cumulative probability distribution value of the relative forecast error at the current forecast moment.

[0091] Step 5-3-3 Randomly generate α2, α3, …, α T ∈(0,1), solve the following system of equations as follows:

[0092]

[0093] The relative error cumulative probability distribution values ​​x2, x3...x at time 2 to T can be obtained. T ;

[0094] Step 5-3-4 Repeat Step 5-3-3 N times to obtain N groups of cumulative probability distribution values ​​of relative forecast errors.

[0095] Step 5-3-5 Based on the obtained cumulative probability distribution function of relative forecast error, the inverse function of the joint distribution function of relative forecast errors at adjacent moments established by the Step 4 method is reversed to obtain N groups of relative forecast errors. The relative forecast error is coupled with the determined integrated forecast result to realize random flood forecasting.

[0096] The effectiveness and rationality of the method of the present invention are now illustrated by taking the classified flood forecast of a certain reservoir as an example. 22 typical floods in the history of the reservoir were selected and classified based on the self-organizing map neural network SOM. The classification results are divided into two categories: C1 and C2. For the two types of floods, the MIC method is used to screen the influencing factors; the influence with a MIC value greater than 0.3 is selected as the model input, and the momentary runoff is used as the output. The forecast models based on LSTM, GRU and Xgboost are established respectively, and the models are calibrated and verified; then the cloud model is used to perform weighted fusion on the results of different forecast models to obtain the model deterministic forecast results. The flood simulation results of different models are shown in the figure. Figure 3 and Figure 4 The four indicators of correlation coefficient (R), Nash coefficient (NSE), mean absolute error (MAE), and root mean square error (RMSE) were used to evaluate the forecast results, as shown in Tables 1 and 2. It can be seen that while weighted averaging using the cloud model cannot guarantee optimal results, it can effectively compensate for the uncertainty of the model.

[0097] Based on the deterministic integrated forecast results, the forecast error is calculated and analyzed, the marginal distribution of the forecast error is established, and the joint distribution function of the time period before and after the forecast error is constructed. The error is further randomly simulated based on Gibbs sampling, and the deterministic forecast results are superimposed to obtain the final random forecast interval. The overall coverage is good, and the results are as follows: Figure 5 As shown, the confidence interval is 85%.

[0098] Table 1 Evaluation results of indicators during the training and validation periods of different forecast models under Class C1 flood

[0099]

[0100] Table 2 Evaluation results of indicators during the training and validation periods of different forecast models under type C2 flood

[0101]

Claims

1. A classification flood stochastic forecasting method based on machine learning and cloud model, characterized in that: The following steps are involved: Step 1: Using typical historical flood events as the basic data for model calibration and testing, we select classification indicators and classify historical flood events based on a self-organizing map (SOM) neural network. The classification indicators include one or more of the following: total rainfall, maximum three-hour rainfall, rainfall intensity, rainfall center, previous impact rainfall, and flood peak. Step 2: Use the maximum information coefficient method (MIC) to screen flood influencing factors for different types of floods, and establish classification flood forecasting models based on different machine learning methods to obtain the optimal parameters of different models corresponding to different types of floods; Step 3: For different types of floods, the fusion weights of different forecast models are solved based on the cloud model. The simulation results of each model are weighted based on the fusion weights to obtain the model integrated forecast results; Step 4: Analyze and calculate the relative forecast error, establish the joint distribution function of the relative forecast errors at adjacent moments based on the Copula method, and determine the cumulative probability distribution function of the relative forecast error; Step 5: Obtain real-time online flood information and implement random flood forecasting; In Step 2, a classification flood forecasting model based on different machine learning methods is established, which includes the following steps: Step 2-1: Divide the flood series into training and validation sets according to the number of floods. Use the candidate predictors with MIC values ​​greater than 0.30 as model input and the number of floods as model output. Normalize all data sets to the interval (0, 1). Step 2-2: Establish a forecast model based on different machine learning methods {M1, M2, …M i , … M K } , i=1, 2, …, K, K≥2, where the machine learning method is two or more of the following: ANN, SVM, LSTM, GRU, and XGBoost; the model output is calculated and denormalized to obtain the runoff simulation value ,in, represents the simulated value of the i-th forecast model at time t, , K is the number of forecast models, T is the duration of different floods (h); Step 2-3: Compare the secondary flood observations in the training set ,in, represents the observation value at time t, and uses the correlation coefficient, Nash coefficient, root mean square error, and mean absolute error as evaluation indicators, and adopts the grid method to optimize the parameters of different machine learning models; Step 2-4: Input the validation set data into the trained model to obtain the simulation results, and then obtain the runoff simulation value after denormalization. , using correlation coefficient, Nash coefficient, mean absolute error and root mean square error as evaluation indicators to judge whether the simulation results of the model are qualified; In Step 5, online real-time flood information is obtained to implement random flood forecasting, which specifically includes the following steps: Step 5-1: Obtain real-time online flood information, extract classification indicator data from the real-time rainfall information, and perform online flood classification based on the self-organizing map neural network (SOM) established in Step 1 to determine the category to which the flood belongs; Step 5-2: Select the optimal model parameters of the flood category obtained in Step 2 and the fusion weights obtained based on the method in Step 3 to perform deterministic integrated forecasting; Step 5-3: Use the Gibbs sampling method to perform random simulation of the relative forecast error, and couple the simulated relative forecast error with the deterministic integrated forecast result to achieve random flood forecasting; Step 5-3 uses the Gibbs sampling method to perform random simulation of the relative forecast error at the forecast time, including the following steps: Step 5-3-1 Generate two random numbers α0, α1∈(0, 1), where α0 and α1 are probability values; Step 5-3-2 Based on conditional probability distribution , where X1 represents the relative forecast error at the current moment, X0 represents the relative forecast error 1 hour before the hypothetical flood forecast moment, and C is the joint distribution function of the copula relative forecast errors at adjacent moments, obtained based on the method in Step 4; Substitute α0 and α1 into the conditional probability distribution to obtain the cumulative probability distribution value x1 of the relative forecast error at the current forecast moment; Step 5-3-3 Randomly generate probability values ​​α2, α3, …, α T ∈(0, 1), solve the following system of equations as follows: ， The relative error cumulative probability distribution values ​​x2, x3, ... x at time 2 to T can be obtained. T ; Step5-3-4 Repeat Step5-3-3 N times to obtain N groups of cumulative probability distribution values ​​of relative forecast errors ( ), j=1, 2, …, N; Step 5-3-5 Based on the obtained cumulative probability distribution function of relative forecast error, the inverse function of the joint distribution function of relative forecast errors at adjacent moments established by the Step 4 method is reversed to obtain N groups of relative forecast errors ( ), j = 1, 2, …, N, the relative forecast error is coupled with the determined integrated forecast result to realize the random flood forecast.

2. The classification flood stochastic forecasting method based on machine learning and cloud model according to claim 1, characterized in that: In Step 2, the maximum information coefficient method MIC is used to screen the influencing factors, including the following methods: for different types of floods, analyze the selected time t forecast rainfall P t And the previous rainfall t-N , P t-N+1 , P t-N+2 ,…, P t-1 } as a set of candidate predictors; Calculate the current flow rate Q based on the MIC method t The correlation with different candidate predictors, i.e., MIC values, was used, and the candidate predictors with MIC values ​​greater than 0.30 were selected as the input of the prediction model.

3. The classification flood stochastic forecasting method based on machine learning and cloud model according to claim 2, characterized in that: In Step 3, the fusion weights of different forecast models are solved based on the cloud model, and the simulation results of each model are weighted based on the fusion weights to obtain the model integrated forecast result, including the following steps: Step 3-1: Select the observed values ​​and simulated values ​​in the training set, and establish the corresponding cloud model pair based on the inverse normal cloud generator. , thereby converting the distribution characteristics of the sequence into corresponding digital features, where represents the cloud model expectation curve constructed by the simulated value of the i-th forecast model, The expected curve of the cloud model constructed from the observations; Step 3-2: Calculate the simulation values ​​of different forecast models based on the overlap of cloud model expectation curves and observations The global similarity of : ， Step 3-3: Global similarity Solve the fusion weights of each forecast model : ， Step 3-4: Perform weighted combination of the simulation results of each model to obtain the integrated forecast result: ， Where, represents the simulated value of the i-th forecast model at time t, i=1, 2, …, K, K is the number of forecast models, K≥2; The integrated results are obtained by weighted averaging the simulated data of different models in the validation set, and the correlation coefficient, Nash coefficient, mean absolute error and root mean square error are used as evaluation indicators to judge whether the model integration prediction results are qualified.

4. The classification flood stochastic forecasting method based on machine learning and cloud model according to claim 3, characterized in that: In Step 4, a joint distribution function of relative forecast errors at adjacent moments is established based on the Copula method, which specifically includes the following steps: Step 4-1: Assume that the relative runoff forecast error is , the relative runoff forecast error distribution is constructed as the marginal distribution, the joint distribution of the previous and next moments is established based on different Copula functions, and the maximum likelihood estimation method is used to solve the unknown parameters in the joint distribution; Step 4-2: Use the AIC index to evaluate the obtained joint distribution function, and select the Copula function with the smallest AIC index value as the final joint distribution function.

Images