Intelligent flood forecasting method for coupling error correction and joint modeling

By constructing a multi-model structure and an error correction model, and combining Bayesian posterior results, the problems of insufficient uncertainty modeling and limited error correction capabilities in existing flood forecasting methods are solved. This achieves an accurate expression of the posterior probability distribution of measured flow, thereby improving the stability and reliability of flood forecasting.

CN121390431APending Publication Date: 2026-01-23HOHAI UNIV +2
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202511510029.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-10-22
Publication Date
2026-01-23

AI Technical Summary

Technical Problem

Existing flood forecasting methods struggle to effectively express rainfall uncertainty, non-Gaussian and heteroscedastic errors when faced with uncertainty issues, and traditional models are unable to accurately derive the posterior probability distribution of measured flow rates.

Method used

A real-time flood rolling probability forecasting method using coupled error correction and joint modeling is adopted. By constructing a multi-model structure, introducing disturbance input scenarios, building an error correction model, and fusing Bayesian posterior results, the method achieves an accurate expression of the posterior probability distribution function of measured flow.

Benefits of technology

It improves the stability and reliability of flood forecasting systems under complex conditions, and can more accurately express the uncertainty of rainfall input, the non-Gaussianity and heteroscedasticity of correction errors, and derive the posterior distribution of measured flow.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN121390431A_ABST
    Figure CN121390431A_ABST
Patent Text Reader

Abstract

The invention discloses an intelligent flood forecasting method for coupling error correction and joint modeling, and relates to a deep learning and uncertainty modeling technology. At the input end, constructing a future random rainfall scene through hourly dynamic normal disturbance; the method comprises the following steps: at a model end, introducing a multi-structure and multi-objective function combination based on Kolmogorov-Arnold Networks and Transform, and forming a multi-member ensemble forecast; at an error end, a probabilistic modeling method based on a numerable asymmetric Laplacian mixed density network is provided, and fine error correction is realized; a Vine copula function is adopted to construct high-dimensional joint distribution, and a Bayesian model averaging and expectation maximization algorithm is combined to realize weighted fusion of multi-member posterior results; according to the method, uncertainty in flood forecasting can be comprehensively described, the stability and adaptability of a forecasting system are improved, and the method is suitable for a basin-level real-time flood ensemble forecasting scene.
Need to check novelty before this filing date? Find Prior Art

Description

TECHNICAL FIELD

[0001] The present application relates to the field of hydrological forecasting, and in particular to a flood intelligent forecasting method coupling error correction and joint modeling. BACKGROUND

[0002] Flood forecasting has important application value in disaster prevention and mitigation, water resources scheduling and other fields. The traditional flood forecasting method mainly uses deterministic model for single flow forecasting, which is difficult to cope with the uncertainty problem widely existing in practice, especially in the future rainfall input, model structure and parameter estimation, which is easy to lead to large deviation of the forecasting result, affecting the scientificity of flood control decision. In order to cope with uncertainty, ensemble forecasting method is gradually introduced into the field of flood forecasting. This method improves the coverage ability of possible future scenarios by introducing multiple input scenarios and model members. However, the existing ensemble forecasting method is mostly based on simple resampling or static disturbance method in input disturbance processing, which fails to effectively express the rainfall uncertainty; in error processing, linear correction or empirical distribution adjustment is generally used, which is difficult to express the non-Gaussian and heteroscedasticity of error; at the same time, the modeling of joint distribution between predicted flow and measured flow is mostly limited to binary correlation structure, which is difficult to accurately derive the posterior probability distribution function of measured flow. SUMMARY

[0003] The present application aims to overcome the defects in the prior art, and to solve the problems of insufficient modeling of uncertainty, limited error correction ability and lack of joint distribution expression in the existing flood ensemble forecasting method. The present application proposes a real-time flood rolling probability forecasting method coupling error correction and joint modeling, which realizes accurate expression and rolling output of the posterior probability distribution function of measured flow by constructing multi-model structure, introducing disturbance input scenario, constructing error correction model and fusing Bayesian posterior results, and improves the stability and reliability of the forecasting system under complex conditions.

[0004] To achieve the above-mentioned purpose, the present application provides a flood intelligent forecasting method coupling error correction and joint modeling, comprising the following steps: S1, data staging division: the historical hydro-meteorological data is divided into three stages of rating period, calibration period and verification period, the rating period is used for training deterministic forecasting model, the calibration period is used for future random rainfall generation, error modeling and joint distribution parameter rating, and the verification period is used for testing the generalization ability of the framework; S2, future random rainfall scenario construction: the input uncertainty modeling method based on Gaussian disturbance is used to process the measured rainfall data in the forecast period, and a future random rainfall scenario set reflecting the uncertainty of rainfall input is constructed; S3, Constructing a multi-member set prediction set: constructing a Kolmogorov-Arnold Networks model and a classic Transformer model, and using MSE and MAE as objective functions for model construction and parameter training, and combining the future random rainfall scenarios constructed in S2 to form a multi-member set prediction set; S4, Constructing a CMAL-MDN error correction model based on the historical error sequence of the set members; S5, Vine copula joint modeling and posterior probability distribution derivation: based on the set member results after error correction, a Vine copula function is used to construct the joint distribution structure between the benchmark flow, the predicted flow and the measured flow, and the posterior probability distribution function of the measured flow corresponding to each set member is derived based on the Sklar theorem; S6, BMA fusion and rolling probability prediction output: based on the BMA method, the posterior probability distribution of each prediction member is weighted and fused, the fusion weight is iteratively optimized through the EM algorithm, and the final rolling probability prediction result of the measured flow is output.

[0005] Preferably, step S2 specifically comprises: S201, during the calibration period, the future measured rainfall is taken as the future predicted rainfall; during the calibration period and the verification period, the future random rainfall scenario is constructed; the starting prediction time is t, the maximum prediction period is M time periods, and the measured rainfall data of the prediction period 1 to M independent perturbation is performed to randomly generate a perturbation factor at t+m time points following a normal distribution; S202, for each time point of the measured rainfall value hourly perturbation is performed, and the expression is as follows: ; In the formula, is the perturbed future rainfall value; is the measured rainfall value; is a perturbation factor following a normal distribution; S203, during the perturbation process, a plurality of future rainfall scenario sets are generated by repeating the perturbation multiple times, and a group of typical rainfall scenarios are selected.

[0006] Preferably, step S3 specifically comprises: S301, based on the selected a group of typical rainfall scenarios, each is combined with the prediction model KAN, Transformer and objective function MSE, MAE, each combination constitutes a prediction member, specifically including:

[0007] To express the input, structure and parameter uncertainties, a multi-lead rolling prediction mechanism is adopted, and the prediction output of the i th prediction member in the m th lead is expressed as As the model input in the m + 1 th lead, the rolling continuous prediction process is formed, and the expression is as follows: ; In the formula, is the historical influence period; and is the model structure and parameter set of the i th prediction member in the m + 1 th lead; is the prediction value of the i th prediction member in the m + 1 th lead; is the historical rainfall input of the model; is the future rainfall sequence after disturbance; is the historical runoff input of the model; is the predicted runoff of the i th prediction member in the 1 st to m th lead; S302, in the calibration period, the measured rainfall sequence is used as the future prediction input without introducing rainfall disturbance, and the corresponding parameters under each model structure are trained; in the calibration period and the verification period, the future random rainfall scenarios generated in S2 are introduced to construct the input uncertainty, and the disturbance scenarios are combined with the model structure and the objective function to generate the set prediction results of N prediction members.

[0008] Preferably, step S4 specifically comprises: S401, constructing the error sequence between the prediction model output value and the measured runoff in the calibration period and the verification period under different leads and determining the input and output of the model; S402, constructing the ALD probability density function; S403, constructing and training the CMAL-MDN model; S404, error correction implementation.

[0009] Preferably, the contents of constructing the error sequence and preparing the model input in S401 are as follows: The error formula is constructed , and the expression is as follows: ; In the formula, is the prediction value of the i th set member at time t; is the measured value of the i th set member at time t; Determine the input of the model is the combination variable of the historical error term and the current prediction value, and the expression is as follows: ; In the formula, error term of each time interval from time t to time L2; determining the output of the model as the error prediction value at time t .

[0010] Preferably, the content of constructing the ALD probability density function in S402 is as follows: constructing the ALD probability density function , the expression is as follows: ; wherein, is a location parameter; is a scale parameter; is a skewness parameter; and y is a target variable to be predicted; constructing the ALD expectation function , the expression is as follows: .

[0011] Preferably, the content of constructing and training the CMAL-MDN model in S403 is as follows: constructing the overall error conditional probability density function , the expression is as follows: ; wherein, is a weighting coefficient of the kth ALD component in the mixed distribution; is a location parameter of ; is a scale parameter of ; is a skewness parameter of ; using a negative log-likelihood loss as a training loss function , the expression is as follows: ; wherein, is a weighting coefficient of the kth ALD component in the mixed distribution; is a location parameter of ; is a scale parameter of ; is a skewness parameter of .

[0012] Preferably, the content of implementing error correction in S404 is as follows: calculating the error prediction value , the expression is as follows: ; wherein, Let i be the position parameter of the i-th member at time t; Let be the scale parameter of the i-th member at time t; Let be the skew parameter of the i-th member at time t; Error prediction value The predicted value added to the i-th set member at time t Obtain the output of the i-th forecast member after error correction. The expression is as follows: .

[0013] Preferably, step S5 specifically includes: S501. Perform joint modeling on the error-corrected set member results; The Vine copula function is used to construct a multivariate joint distribution among baseline, predicted, and measured flow rates; the marginal distribution functions of each variable are optimized using AIC; and the joint distribution function of the three variables is determined according to Sklar's theorem. Represented as a combination of the marginal distribution function and the copula function The expression is as follows: ; In the formula, The baseline flow rate; For predicting traffic flow; The measured flow rate is represented by U, V, and W, which are variables. The marginal distribution function, where , , ; Represents the copula distribution function; , and They are respectively The threshold value; The Vine copula structure was constructed using the C-vine form, and the parameters were estimated using the maximum likelihood method. Build C-vine (root= The joint density of the three variables is expressed as follows: ; ; ; in, Represents the density function; Represents the copula density function; Let U be the binary copula density function between U and W in the first layer. is the conditional distribution function of V given U, is the conditional copula density function of V given U and W after given W, is the conditional distribution function of V given U, V is the conditional distribution function of V given U, is the conditional distribution function of V given U, U is the conditional distribution function of V given U, W is the conditional distribution function of V given U, is the conditional distribution function of V given U, constructing C-vine (root = W) ; ; ; where, is the conditional distribution function of V given U, is the conditional distribution function of V given U, is the conditional copula density function of V given U and W after given W, is the conditional distribution function of V given U, W is the conditional distribution function of V given U, U is the conditional distribution function of V given U, is the conditional distribution function of V given U, W is the conditional distribution function of V given U, V is the conditional distribution function of V given U, S502, according to the multivariate joint distribution constructed by Vine copula, and based on the generalization form of conditional distribution of Sklar theorem, the posterior probability distribution function of each prediction member corresponding to the measured flow is obtained, the specific process is: selecting the reference flow as the center node, first, the bivariate copula between the reference flow and the predicted flow, the reference flow and the measured flow is constructed, and the conditional copula function of the predicted flow and the measured flow under the condition of the reference flow is obtained by conditional h-function calculation; the posterior probability distribution function of the measured flow is obtained, the specific content is as follows: the posterior probability density function of the measured flow is obtained, the expression is as follows: ; the posterior probability distribution function of the measured flow is obtained, the expression is as follows: ; selecting the predicted flow For the central node, the bivariate copula function of the predicted flow and the benchmark flow and the predicted flow and the measured flow is constructed, and the conditional copula function of the benchmark flow and the measured flow under the condition of the given predicted flow is obtained, the corresponding posterior probability distribution function of the measured flow is obtained, and the specific content is as follows: The posterior probability density function of the measured flow is obtained, and the expression is as follows: ; The posterior probability distribution function of the measured flow is obtained, and the expression is as follows: .

[0014] Preferably, step S6 specifically comprises: S601, the posterior probability distribution functions of a plurality of prediction members are weighted and fused by using the BMA method, and the weighting coefficients of the members are determined by using the EM algorithm ; the posterior probability distribution functions of each prediction member obtained based on the Vine copula function are weighted and fused by BMA, and finally the set posterior probability density function is obtained, and the expression is as follows: ; S602, based on the set posterior probability density function, the typical statistical quantity results corresponding to the measured flow are calculated, including the expected value, the median and the upper and lower limits under a given confidence level, and the specific operation steps include: S6021, expected value calculation; According to the set posterior probability density function, the mathematical expectation value is obtained by integration, and the expression is as follows: ; S6022, median calculation, and the expression is as follows: ; S6023, given confidence level , the lower limit and the upper limit corresponding to 90% or 95% confidence level are obtained, the confidence interval is estimated, and the expression is as follows: ; .

[0015] The application also provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to realize the steps of the method in any one of the application.

[0016] ​Therefore, the application adopts the above-mentioned one, and has the beneficial effects of: 1、The application proposes a future rainfall scenario generation method of hourly dynamic normal disturbance at the input end, which is different from the traditional static disturbance and resampling method, can more truly and carefully express the uncertainty of rainfall input, and combines a multi-forecast model structure and a multi-objective function combination to comprehensively cover three types of uncertainties of input, structure and parameters, so as to significantly improve the stability and adaptability of the ensemble prediction system.

[0017] 2、The application innovatively introduces a countable asymmetric Laplace mixed density network (CMAL-MDN) in the error correction link, can simultaneously depict the non-Gaussianity, heteroscedasticity and autocorrelation of the error, breaks through the limitations of the existing linear correction and Gaussian distribution assumption, realizes fine correction of the ensemble prediction result, and effectively improves the precision and robustness of the prediction.

[0018] 3、In the joint modeling and result fusion aspect, the application first combines the high-dimensional dependence structure of the Vine copula with the Bayesian model average (BMA), and adaptively optimizes the weighting coefficient by using the expectation maximization algorithm, not only breaks through the limitation of the traditional binary copula modeling, but also can more accurately derive the posterior distribution of the measured flow, and finally realizes the high-credibility and rollable update of the flood probability prediction output.

[0019] The technical solutions of the application will be further described in detail below with the help of the drawings and examples. DESCRIPTION OF DRAWINGS

[0020] Figure 1 It is a whole flow chart of a flood intelligent prediction method of coupling error correction and joint modeling of the embodiment of the application. Figure 2 It is a comparison chart of Nash efficiency coefficients before and after error correction of the application method in the verification set of the forecast period 1 to the forecast period 7. Figure 3 It is a comparison chart of probability prediction results of the application method and the traditional HUP-BMA model in the forecast period 1. Figure 4 It is a comparison chart of probability prediction results of the application method and the traditional HUP-BMA model in the forecast period 7. DETAILED DESCRIPTION

[0021] For the purpose of the embodiments of the present application, the technical solutions and advantages are more clearly and clearly understood, the embodiments of the present application are further described in detail below in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the embodiments of the present application, and are not used to limit the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor belong to the scope of protection of the present application. The examples of the embodiments are shown in the drawings, wherein the same or similar reference numerals represent the same or similar elements or elements with the same or similar functions throughout.

[0022] It should be noted that the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or server including a series of steps or units, not necessarily limited to those clearly listed, but can include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0023] The embodiments of the present application will be described in detail below in combination with Figures 1-4 The embodiments of the present application will be described in detail below in combination with

[0024] Embodiments The present application proposes a flood intelligent prediction method combining error correction and joint modeling, comprising the following steps: S1, the collected historical hydrological and meteorological data is divided into three stages: calibration period, calibration period and verification period. Among them, the calibration period is used to train the deterministic prediction model; the calibration period is used for future random rainfall generation, error modeling and joint distribution parameter calibration; the verification period is used to test the generalization ability of the prediction framework, and the data used in the modeling process does not participate in any parameter estimation; S2, the measured rainfall data in the prediction period is dynamically constructed, and a set of future random rainfall scenarios is constructed. The input uncertainty modeling method based on Gaussian disturbance is adopted, and the future measured rainfall is dynamically disturbed by disturbance factor Distribution to reflect the uncertainty of rainfall input; Step S2 specifically includes: S201, in the calibration period, do not construct future random rainfall, take the future measured rainfall as the future prediction rainfall; in the calibration period and the verification period, construct future random rainfall. For the starting prediction time t, the maximum prediction period is M time periods, and the measured rainfall data of the prediction period 1 to M is independently disturbed, and the disturbance factor of each time is randomly generated from the normal distribution with mean 0 and standard deviation 0.2, that is, .

[0025] S202, obtaining the measured rainfall value at each time point The hourly disturbance is performed, and the formula is: ; wherein, is the future rainfall value after disturbance, is the measured rainfall value, is a disturbance factor subject to normal distribution, reflecting the input uncertainty of future rainfall.

[0026] S203, during the disturbance process, a plurality of future rainfall scenario sets are generated by repeating the disturbance multiple times. Among all the disturbance-generated scenarios, a group of representative rainfall processes are selected, wherein a is 2, which is used as a typical input scenario for subsequent model driving and posterior analysis.

[0027] S3, under the premise of considering rainfall uncertainty, two deep learning models with obvious structural differences are adopted, which are a neural network model (Kolmogorov-Arnold Networks, KAN) constructed on the basis of Kolmogorov-Arnold representation theorem and a classic Transformer model, and mean squared error (Mean Squared Error, MSE) and mean absolute error (Mean Absolute Error, MAE) are used as objective functions for model construction and parameter training, reflecting the uncertainty of model structure and parameters. The future random rainfall scenario is combined with the model to form a multi-member ensemble prediction set; Step S3 specifically includes: S301, based on the selected two groups of typical rainfall scenarios and , respectively combined with the prediction models KAN, Transformer and the objective functions MSE, MAE, each combination constitutes a prediction member, including , , , , , , , A total of 8 ensemble prediction members are used to express the input, structure and parameter uncertainties. The above prediction strategy adopts a multi-forecast period rolling prediction mechanism, and the prediction output of the i-th prediction member at the forecast period m will be used as the model input at the forecast period m+1, constituting a rolling continuous prediction process, described as follows: ; In the formula, is the historical influence period; and The model structure and parameter set of the i th prediction member at the forecast period m+1; The prediction value of the i th prediction member at the forecast period m+1; The historical rainfall input of the model; The future rainfall sequence after disturbance; The historical runoff input of the model; The predicted runoff of the i th prediction member at the forecast period 1 to m; S302, the model calibration and verification are hierarchically processed according to the data stage. In the calibration period, in order to ensure that the model training is stable under the condition of true rainfall and does not introduce rainfall disturbance, all measured rainfall sequences are used as future prediction inputs, and the corresponding parameters under each model structure are trained; in the calibration period and the verification period, the future random rainfall scenario set generated in step S2 is introduced to construct the input uncertainty, and the disturbance scenario is combined with the model structure and the objective function to generate The set prediction results of the i th prediction member are shown in Table 1.

[0028] Table 1 Statistical index table of deterministic prediction results of prediction members at forecast period 1 and forecast period 7

[0029] Among them, the models M1-M8 represent P1-KAN-MSE, P1-KAN-MAE, P1-Transformer-MSE, P1-Transformer-MAE, P2-KAN-MSE, P2-KAN-MAE, P2-Transformer-MSE, and P2-Transformer-MAE.

[0030] S4, for the historical error sequence of each set member, an error correction model based on Countable Mixtures of Asymmetric Laplacians Mixture Density Network (CMAL-MDN) is constructed to capture the non-Gaussian, autocorrelation and heteroscedasticity characteristics of the error, and the expected value of the prediction error distribution is used to correct the original prediction, thereby improving the prediction accuracy and stability; Step S4 specifically includes: S401, first, the error sequence between the prediction model output value and the measured runoff at different forecast periods in the calibration period and the verification period is constructed, and the prediction value of the i th set member at time t is , the measured value is , and the error is: ; The input of the model is the past​ Sliding window data for a given period, including a combination of historical error terms and current forecast values. The output is the predicted error value at time t. .

[0031] S402. A countable asymmetric Laplace mixture density network consists of K asymmetric Laplace distributions (ALDs), and the probability density function of an ALD is: ; in, It is a position parameter. It is a scale parameter. For skew parameters, and Characterizing the asymmetry of the distribution; y The target variable to be predicted is denoted as .

[0032] The expected value of ALD is: .

[0033] S403. This invention introduces a countable asymmetric Laplace mixed density network model, which can effectively model the asymmetry and heavy-tailed characteristics of the error. For the i-th prediction member, the network input is from S401. The output includes: (1) the weights of the K components. ,satisfy (2) Corresponding position parameters, scale parameters and skew parameters .

[0034] The model output target is the overall error conditional probability density function: ; In the formula, represents the weighting coefficient of the k-th ALD component in the mixed distribution; for Position parameters; for Scale parameters; for The skew parameter; The model is trained by minimizing the negative log-likelihood loss: ; In the formula, represents the weighting coefficient of the k-th ALD component in the mixed distribution; for Position parameters; for Scale parameters; for skewness parameter of the i th ensemble member at time t.

[0035] S404, after obtaining the error probability density function of the model output, the error prediction value is calculated: ; wherein, is the location parameter of the i th ensemble member at time t; is the scale parameter of the i th ensemble member at time t; is the skewness parameter of the i th ensemble member at time t; Finally, the error prediction value is added to the prediction value of the i th ensemble member at time t to obtain the error-corrected i th ensemble member output: .

[0036] The single-point prediction results of the ensemble members are post-processed through the error correction strategy, the prediction accuracy is improved, and high-quality inputs are provided for subsequent joint distribution construction and posterior probability fusion.

[0037] S5, based on the error-corrected ensemble member results, a Vine copula function is used to construct the joint distribution structure among the benchmark flow, the predicted flow and the measured flow, and the structure and parameters of the Vine copula are optimized, and the posterior probability distribution function of the measured flow corresponding to each ensemble member is derived according to the Sklar theorem; Step S5 specifically includes: S501, the error-corrected ensemble member results are jointly modeled, and a Vine copula function is used to construct a multivariate joint distribution among the benchmark flow , the predicted flow and the measured flow . Compared with the traditional binary copula model, the Vine copula can effectively capture the asymmetric and nonlinear dependence relationship between the flow data, provide more flexible and accurate modeling capability, and is suitable for joint modeling of high-dimensional data.

[0038] Suppose U, V, W represent the marginal distribution functions (corresponding to the benchmark flow, the measured flow and the predicted flow, respectively) of variables , and , the commonly used marginal distribution functions include Gamma distribution, normal distribution, lognormal distribution, Pearson III distribution, generalized extreme value distribution, etc., and the marginal distribution functions of each variable are optimized through Akaike Information Criterion (AIC). According to the Sklar theorem, the three-variable joint distribution function can be expressed as a combination of marginal distribution functions and copula functions: ; where, , , , denotes the copula distribution function, which is used to model the joint dependence structure among variables.

[0039] The Vine copula structure adopts C-vine form, and the pair-copula function is selected from five commonly used copula functions: Gaussian, Student-t, Clayton, Gumbel, and Frank. The parameters are estimated by maximum likelihood method.

[0040] 1) C-vine (root = U) Three-variable joint density: ; ; ; where, denotes the density function; denotes the copula density function; is the bivariate copula density function between U and W in the first layer (Tree1); is the bivariate copula density function between U and V in the first layer (Tree1); is the conditional copula density function of V and W given U in the second layer (Tree2); is the conditional distribution function of V given U; is the conditional distribution function of W given U.

[0041] 2) C-vine (root = W) Three-variable joint density: ; ; ; where, is the bivariate copula density function between W and U in the first layer (Tree1); is the bivariate copula density function between W and V in the first layer (Tree1); is the conditional copula density function of U and V given W in the second layer (Tree2); the conditional distribution function of U given W; the conditional distribution function of V given W.

[0042] S502, the posterior probability distribution function of the measured flow corresponding to each prediction member is derived according to the multivariate joint distribution constructed by the Vine copula, and further combined with the generalization form of the Sklar theorem under the conditional distribution.

[0043] When the reference flow is taken as the center node, first, the bivariate copula between the reference flow and the predicted flow, and the reference flow and the measured flow is constructed, and further, the conditional copula function of the predicted flow and the measured flow under the condition of the given reference flow is calculated through the conditional ℎ-function. Thus, the posterior probability distribution function of the measured flow is obtained , which is in the form of the combination of the marginal distribution function and the conditional copula function 1) the posterior probability density function (root= ) of the measured flow : ; 2) the posterior probability distribution function (root= ) of the measured flow : ; When the predicted flow is taken as the center node, the bivariate copula function of the predicted flow and the reference flow, and the predicted flow and the measured flow is constructed, and further, the conditional copula function of the reference flow and the measured flow under the condition of the given predicted flow is derived, and then the corresponding posterior probability distribution function of the measured flow is obtained.

[0044] 1) the posterior probability density function (root= ) of the measured flow : ; 2) the posterior probability distribution function (root= ) of the measured flow : ; ​​In the three-dimensional case, the benchmark flow and the predicted flow are respectively taken as the root node to construct the Vine copula structure. Due to the different selection of the root node, the decomposition form and the conditional copula are different, so they can be regarded as two different Vine structures. Then under each structure, different pair-copula functions and their parameter combinations are traversed, the posterior distribution of the observed flow is derived based on the Sklar theorem, and the Continuous Ranked Probability Score (CRPS) is calculated on the calibration set. The minimum CRPS is used as the criterion to optimize the optimal structure and parameters.

[0045] S6, the posterior distribution of each prediction member is weighted and fused by the Bayesian model averaging method, and the fusion weight is iteratively optimized by the expectation maximization algorithm. The final rolling probability prediction result of the observed flow is output.

[0046] Step S6 specifically includes: S601, the Bayesian model averaging (BMA) method is used to weight and fuse the posterior probability distribution of multiple prediction members , and the expectation maximization (EM) algorithm is used to determine the weighting coefficient of each member , which reflects the credibility of each prediction member in actual prediction. BMA weights and fuses the posterior probability distribution of each prediction member derived based on the Vine copula function, and finally obtains the set posterior probability density function: ; S602, based on the set posterior probability density function, the typical statistical quantity results corresponding to the observed flow are calculated, including the expected value, the median and the upper and lower limits under a given confidence level, specifically including the following operation steps: (1) expected value calculation. According to the set posterior probability density function , the mathematical expectation value is obtained by the following integral: ; (2) median calculation. The median is the observed flow value that satisfies the following condition: ; (3) for a given confidence level , usually set =0.1 or =0.05, the lower limit and the upper limit are obtained corresponding to the confidence level of 90% or 95%, the confidence interval is estimated, and it satisfies: ; 。

[0047] In another aspect, the application provides an electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, the processor implementing the steps of the method of any of the application when executing the computer program.

[0048] The application selects the hydrological station measured data of a typical basin to carry out comparative experiments. In the aspect of deterministic prediction evaluation, five commonly used indexes such as root mean square error (RMSE), mean absolute error (MAE), correlation coefficient (R), Nash efficiency coefficient (NSE) and Kling-Gupta efficiency coefficient (KGE) are adopted for analysis; in the aspect of probability prediction evaluation, four indexes such as coverage (PICP), average interval width (PINAW), Coverage Width-based Criterion (CWC) and continuous rank probability score (CRPS) are selected for comparative evaluation. Figure 2 The comparative results of the method of the application and other models in the rolling flood prediction task are shown, wherein (a) is a statistical index table of deterministic prediction results of prediction period 1 and prediction period 7 prediction members; (b) is a comparison chart of Nash efficiency coefficients before and after error correction of the verification set of prediction period 1 to prediction period 7; (c) is a comparison chart of probability prediction results of the method of the application and the traditional HUP-BMA model. The results in the chart show that the prediction effects of the models are different, and with the increase of the prediction period, the overall fitting precision tends to decrease. After error correction, the Nash efficiency coefficient is obviously improved at different prediction periods, which shows that the error correction method of the application can significantly improve the deterministic prediction precision of the model. In the aspect of probability prediction, compared with the HUP-BMA model, the method of the application significantly reduces the prediction interval width while maintaining reasonable coverage, and the CRPS and CWC are lower, which embodies stronger uncertainty description and risk quantification ability.

[0049] The application further provides a computer readable storage medium having a computer program stored thereon, the computer program implementing the steps of the method of any of the application when executed by a processor.

[0050] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the application but not to limit it, although the application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the application can still be modified or replaced by equivalents, and these modifications or equivalent replacements cannot make the modified technical solutions deviate from the spirit and scope of the technical solutions of the application.

Claims

1. A smart flood forecasting method based on coupled error correction and joint modeling, characterized in that, Includes the following steps: S1. Data phase division: Historical hydrological and meteorological data are divided into three phases: calibration period, calibration period, and validation period. The calibration period is used to train deterministic forecast models, the calibration period is used for future random rainfall generation, error modeling, and joint distribution parameter calibration, and the validation period is used to test the framework's generalization ability. S2. Construction of future stochastic rainfall scenarios: The measured rainfall data within the forecast period are processed using a Gaussian perturbation-based input uncertainty modeling method to construct a set of future stochastic rainfall scenarios that reflect the uncertainty of rainfall input. S3. Constructing a multi-member ensemble forecast set: Construct the Kolmogorov–Arnold Networks model and the classic Transformer model, and use MSE and MAE as objective functions for model construction and parameter training, and combine them with the future stochastic rainfall scenarios constructed in S2 to form a multi-member ensemble forecast set; S4. Construct the CMAL-MDN error correction model based on the historical error sequence of the set members; S5. Joint Modeling with Vine Copula and Derivation of Posterior Probability Distribution: Based on the error-corrected set member results, the Vine Copula function is used to construct the joint distribution structure between the baseline flow, the predicted flow, and the measured flow. Based on Sklar's theorem, the posterior probability distribution function of the measured flow corresponding to each set member is derived. S6, BMA Fusion and Rolling Probability Forecast Output: Based on the BMA method, the posterior probability distributions of each forecast member are weighted and fused. The fusion weights are iteratively optimized through the EM algorithm, and the final measured traffic rolling probability forecast result is output.

2. The intelligent flood forecasting method based on coupling error correction and joint modeling according to claim 1, characterized in that, Step S2 specifically includes: S201. During the calibration period, future measured rainfall is used as future forecast rainfall; during the calibration and validation periods, future stochastic rainfall scenarios are constructed; the initial forecast time is t, and the maximum lead time is M time intervals, using measured rainfall data from lead time 1 to M. Independent perturbations are applied to randomly generate perturbation factors that follow a normal distribution at time t+m; S202, Measured rainfall values ​​at each moment. The hourly perturbation is expressed as follows: ; In the formula, This represents the future rainfall value after the disturbance. This is the measured rainfall value; The perturbation factor is a normally distributed disturbance factor. S203. During the perturbation process, multiple sets of future rainfall scenarios are generated by repeating the perturbation multiple times, and group a of typical rainfall scenarios is selected.

3. The intelligent flood forecasting method based on coupling error correction and joint modeling according to claim 1, characterized in that, Step S3 specifically includes: S301. Based on the selected typical rainfall scenarios in group a, each scenario is combined with the forecast models KAN and Transformer, and the objective functions MSE and MAE, respectively. Each combination constitutes a forecast member, specifically including: ; To express three types of uncertainties—input, structure, and parameters—a multi-forecast rolling forecast mechanism is adopted, which outputs the forecasts of the i-th forecast member for a forecast period m. As the model input for the forecast period m+1, it constitutes the rolling continuous forecast process, as shown in the following expression: ; In the formula: This refers to a period of historical influence. and The model structure and parameter set for the i-th forecast member during the forecast period m+1; Let be the forecast value of the i-th forecast member in the forecast period m+1; Input historical rainfall data for the model; This represents the future rainfall sequence after the disturbance. Input historical runoff for the model; The predicted runoff of the i-th forecast member during the forecast period from 1 to m; S302. During the calibration period, without introducing rainfall disturbances, measured rainfall sequences are used as future forecast inputs to train the parameters corresponding to each model structure. During the calibration and validation periods, the set of future random rainfall scenarios generated in S2 is introduced to construct input uncertainties. The disturbance scenarios are then combined with the model structure and objective function to generate... The ensemble forecast results of each forecast member.

4. The intelligent flood forecasting method based on coupling error correction and joint modeling according to claim 1, characterized in that, Step S4 specifically includes: S401. Construct the error sequence between the forecast model output value and the measured runoff under different forecast periods during the calibration and validation periods, and determine the model input and output; S402, Construct the ALD probability density function; Construction and training of S403 and CMAL-MDN models; S404, Error correction implementation.

5. The intelligent flood forecasting method based on coupling error correction and joint modeling according to claim 4, characterized in that, The steps for constructing the error sequence and preparing the model input in S401 are as follows: Constructing the error formula The expression is as follows: ; In the formula, Let be the predicted value of the i-th set member at time t; Let be the measured value of the i-th set member at time t; Determine the input of the model The variable is a combination of historical error terms and current predicted values, expressed as follows: ; In the formula, This represents the error term for each time period from historical time t to time L2; The model output is determined to be the error prediction value at time t. .

6. The intelligent flood forecasting method based on coupling error correction and joint modeling according to claim 4, characterized in that, The content for constructing the ALD probability density function in S402 is as follows: Constructing the ALD probability density function The expression is as follows: ; In the formula, For position parameters; For scale parameters; y is the skewness parameter; y is the target variable to be predicted; Constructing the ALD Expected Value Function The expression is as follows: 。 7. The intelligent flood forecasting method based on coupling error correction and joint modeling according to claim 4, characterized in that, The construction and training of the CMAL-MDN model in S403 are as follows: Construct the overall error conditional probability density function The expression is as follows: ; In the formula, represents the weighting coefficient of the k-th ALD component in the mixed distribution; for Position parameters; for Scale parameters; for The skewness parameter; Negative log-likelihood loss is used as the training loss function. The expression is as follows: ; In the formula, represents the weighting coefficient of the k-th ALD component in the mixed distribution; for Position parameters; for Scale parameters; for The skew parameter.

8. The intelligent flood forecasting method based on coupling error correction and joint modeling according to claim 4, characterized in that, The error correction implementation in S404 is as follows: Calculate the predicted error value The expression is as follows: ; In the formula, Let i be the position parameter of the i-th member at time t; Let be the scale parameter of the i-th member at time t; Let be the skew parameter of the i-th member at time t; Error prediction value The predicted value added to the i-th set member at time t Obtain the output of the i-th forecast member after error correction. The expression is as follows: 。 9. The intelligent flood forecasting method based on coupling error correction and joint modeling according to claim 1, characterized in that, Step S5 specifically includes: S501. Perform joint modeling on the error-corrected set member results; The Vine copula function is used to construct a multivariate joint distribution among baseline, predicted, and measured flow rates; the marginal distribution functions of each variable are optimized using AIC; and the joint distribution function of the three variables is then determined according to Sklar's theorem. Represented as a combination of the marginal distribution function and the copula function The expression is as follows: ; In the formula, The baseline flow rate; For predicting traffic flow; The measured flow rate is represented by U, V, and W, which are variables. The marginal distribution function, where , , ; Represents the copula distribution function; , and They are respectively , and The threshold value; The Vine copula structure was constructed using the C-vine form, and the parameters were estimated using the maximum likelihood method. Build C-vine (root= The joint density of the three variables is expressed as follows: ; ; ; in, Represents the density function; Represents the copula density function; Let U be the binary copula density function between U and W in the first layer. Let U be the binary copula density function between U and V in the first layer. Let V and W be the conditional copula density functions of the second layer given U; Given the condition of , V The conditional distribution function; For a given U Under the conditions, W The conditional distribution function; Build C-vine (root= Joint density of three variables: ; ; ; in, Let W be the binary copula density function between the first layer W and U. Let W be the binary copula density function between W and V in the first layer. Let U be the conditional copula density function of the second layer given W and V. For a given W Under the conditions, U The conditional distribution function; For a given W Under the conditions, V The conditional distribution function; S502. Based on the multivariate joint distribution constructed using Vine copula, and using the generalized form of Sklar's theorem under conditional distributions, the posterior probability distribution function of the measured flow corresponding to each forecast member is obtained. The specific process is as follows: Selecting a baseline flow rate Using the baseline flow rate as the central node, we first construct bivariate copulas between the baseline flow rate and the predicted flow rate, and between the baseline flow rate and the measured flow rate. Then, we calculate the conditional copula functions of the predicted and measured flow rates under the given baseline flow rate using conditional ℎ-functions. Finally, we obtain the posterior probability distribution function of the measured flow rate. The specific details are as follows: Obtain actual traffic The posterior probability density function is expressed as follows: ; Obtain actual traffic The posterior probability distribution function is expressed as follows: ; Select predicted flow Using the central node, we construct bivariate copula functions for predicted flow and baseline flow, and predicted flow and measured flow, and obtain the conditional copula functions for baseline flow and measured flow under given predicted flow conditions. This yields the corresponding posterior probability distribution function for the measured flow. The specific details are as follows: Obtain the measured flow rate The posterior probability density function is expressed as follows: ; Obtain the measured flow rate The posterior probability distribution function is expressed as follows: 。 10. The intelligent flood forecasting method based on coupling error correction and joint modeling according to claim 1, characterized in that, Step S6 specifically includes: S601. Using the BMA method to analyze the posterior probability distribution function of multiple forecast members. Weighted fusion is performed, and the EM algorithm is used to determine the weighting coefficients of each member. BMA performs weighted fusion of the posterior probability distributions of each forecast member obtained based on the Vine copula function derivation, ultimately yielding the ensemble posterior probability density function. The expression is as follows: ; S602. Based on the ensemble posterior probability density function, calculate the typical statistics of the corresponding measured flow rate, including the expected value, median, and upper and lower limits at a given confidence level. The specific steps include the following: S6021, Expected value calculation; Based on the posterior probability density function of the set, the expected value is obtained by integration, as shown in the following expression: ; S6022, Median The calculation is expressed as follows: ; S6023, Given a confidence level The lower limit is obtained at a confidence level of 90% or 95%. and upper limit Estimate the confidence interval using the following expression: ; 。

Citation Information

Cited By

  • Salt tide probability forecasting method and device based on physical scene adaptive driving

    CN121919699A