CN-122021729-A - Flood peak enhanced physical base flow and residual error correction cooperative runoff prediction method

CN122021729ACN 122021729 ACN122021729 ACN 122021729ACN-122021729-A

Abstract

The invention discloses a physical base flow and residual error correction collaborative runoff prediction method for flood peak enhancement, which belongs to the field of hydrologic prediction and comprises the steps of dividing a training set and a verification set according to a proportion, and carrying out oversampling treatment on flood peak samples to construct a time sequence window. And (3) discretizing and expressing a Xinanjiang model by adopting a normal differential equation, and inputting rainfall and potential evaporation data to obtain an intermediate variable. And constructing a physical base stream and residual error correction double-channel module, and calculating the physical base stream and residual error correction through two fully-connected networks. And calculating a final runoff predicted value by adopting a residual error connection structure, overlapping flood peak sample error weighted terms by taking basic NSE loss as a core, strengthening the fitting precision of the flood peak, and updating physical parameters and neural network weights by a back propagation algorithm. And verifying the model, and respectively calculating the prediction indexes of the training set and the verification set. The method realizes the fusion of the traditional hydrologic model and the deep learning method, enhances the physical interpretation of the model and improves the accuracy of the drainage basin runoff prediction.

Inventors

WANG JINGYI
JIANG ZHIQIANG
JIN LINA
ZHANG CHI
LI ZHIJIN
LU QIANG
Jia Xufan
LUO ZHIMIN
LIANG YONGZHEN

Assignees

华中科技大学

Dates

Publication Date: 20260512
Application Date: 20260202

Claims (9)

1. The flood peak enhanced physical base flow and residual error correction cooperative runoff prediction method is characterized by comprising the following steps of: S1, collecting drainage basin historical hydrologic data, dividing a training set and a verification set according to a proportion, and carrying out oversampling treatment on a flood peak sample so as to construct a time sequence window; s2, a normal differential equation discretization expression Xinanjiang model is adopted, each parameter in a differential equation is explicitly defined, a parameter constraint range is defined, rainfall and potential evaporation data are input, and intermediate variables with clear physical significance are obtained through physical model calculation; S3, selecting a plurality of intermediate quantity splicing sliding windows, the lowest temperature, the highest temperature and the water vapor pressure, entering an LSTM after normalization, extracting time sequence dynamic mode and nonlinear correction information, constructing a physical base stream and residual error correction dual-channel module, and calculating the physical base stream and residual error correction through a first full-connection network and a second full-connection network; s4, calculating a predicted runoff sequence by adopting a residual error connection structure, and calculating the loss of the predicted runoff sequence and a real runoff sequence, wherein the loss function comprises an evaluation item of the fitting precision of the whole sequence, and updating physical parameters and neural network weights of flood peak sample errors by a back propagation algorithm; s5, verifying the model, and respectively calculating prediction indexes of a training set and a verification set, wherein the prediction indexes comprise Nash coefficients NSE, root mean square errors RMSE and relative errors RE.
2. The flood peak enhanced physical base flow and residual correction collaborative runoff prediction method according to claim 1, wherein S1 specifically comprises the following sub-steps: S11, collecting basic hydrological data in a river basin and other meteorological data, wherein the basic hydrological data comprise date, actual measurement runoff, rainfall data and evaporation data, and the other meteorological data comprise the time-interval-by-time maximum air temperature and water vapor pressure; S12, dividing the sequence into training data set verification sets according to a preset proportion; S13, adopting a repeated splicing strategy, determining a flood peak threshold value based on the measured flow, screening flood peak samples, enhancing through oversampling, retaining the data time sequence relevance, and avoiding the prediction deviation caused by the excessively low ratio of the flood peak samples during model training; s14, in order to consider the influence of the history window, at the time t, constructing a characteristic sliding window of rainfall P and evaporation E with a time sequence window of N and a sliding step length of M 。
3. The flood peak enhanced physical base flow and residual correction collaborative runoff prediction method according to claim 1, wherein S2 specifically comprises the following sub-steps: s21, expressing a drainage basin hydrologic process by adopting a normal differential equation based on a new Anjiang model confluence mechanism, wherein the expression is as follows: ; ; In the formula, Is the state variable of the Xinanjiang model at the time t, As an output variable of the model, In order to be able to input the input, 、 For the differentiable function of the confluence of new anjiang products, As a parameter to be learned, Is the time step; S22, physical parameter set Comprises a drainage basin water storage capacity SM, a curve index B, a surface runoff coefficient KI and an underground runoff coefficient KG, and the constraint ranges of parameters are defined to be that SM is more than or equal to 10mm and less than or equal to 50mm, B is more than or equal to 0.1 and less than or equal to 0.8, KI is more than or equal to 0.01 and less than or equal to 0.7, KG is more than or equal to 0.05 and less than or equal to 0.3, KG is less than KI, and the state variable at time t-1 is , And (3) with The water content and the free water storage of the soil at the time t-1 are respectively; S23, inputting rainfall at the moment t Latent evaporation Sequence of constitutions The physical model calculation is used for obtaining 7 new Anjiang intermediate variables with definite physical meaning, namely actual evaporation Moisture content of soil Free water accumulation Surface runoff yield Flow rate of underground runoff Amount of surface runoff confluence Flow collection of underground runoff 7 Xinanjiang intermediate quantity composition sequences at time t Specifically 。
4. A method for peak enhanced physical base flow and residual correction collaborative runoff prediction according to claim 3, wherein the calculation formula in S23 is: the calculation formula of the actual evaporation: ; The calculation formula of the soil water content is as follows: ; the calculation formula of the free water storage quantity comprises the following steps: ; the calculation formula of the surface diameter abortion flow comprises the following steps: ; The calculation formula of the underground radial flow is as follows: ; the calculation formula of the surface runoff converging flow comprises the following steps: ; The calculation formula of the underground runoff converging flow comprises the following steps: 。
5. The flood peak enhanced physical base flow and residual correction collaborative runoff prediction method according to claim 4, wherein the step S3 specifically comprises: s31, selecting a plurality of intermediate quantities to splice other factors, entering an LSTM after normalization, extracting time sequence dynamic mode and nonlinear correction information, and outputting a high-dimensional hidden vector ; S32, the sequence composed of 7 Xinanjiang intermediate quantities Inputting the first full connection network to generate a physical base stream predicted value : ; In the formula, 、 Weights of the hidden layer and the output layer of the first fully connected network respectively, 、 Bias items of a hidden layer and an output layer of the first fully connected network respectively, The 7-dimensional intermediate state quantity is output for S3; s33, outputting the LSTM network with high-dimensional hidden vector at the time t And the physical base stream predicted value Splicing along the feature dimension to form a fusion correction feature vector : ; S34, the fusion correction feature vector Inputting the second full connection network and outputting residual error correction term : ; In the formula, 、 Weights of the hidden layer and the output layer of the second fully connected network respectively, 、 And the bias items of the hidden layer and the output layer of the second fully connected network respectively.
6. The flood peak enhanced physical base flow and residual correction collaborative runoff prediction method according to claim 5, wherein the step S31 specifically comprises: S3101, selecting 5 variables from 7 new An Jiang intermediate quantities, wherein the variables comprise actual evaporation, surface runoff yield, underground runoff yield, surface runoff converging flow and underground runoff converging flow, and performing multidimensional feature vector splicing on the variables, rainfall, potential evaporation history window, highest air temperature, lowest air temperature and water vapor pressure: s3102, constructing a normalization layer for the fusion feature vector, wherein the normalization layer adopts Layer Normalization for processing: ; In the formula, In order to normalize the processed vector values, And Respectively are The mean and variance of (c) are determined, And (3) with In order to be able to learn the parameters of scaling and offset, Is a numerical stable term, equal to 1e-6; S3103, the normalized fusion features Inputting LSTM network, extracting time sequence dynamic mode and nonlinear correction information, and outputting high-dimensional time sequence feature vector 。
7. The method for collaborative runoff prediction with peak enhancement physical base flow and residual correction according to claim 5, wherein the first fully-connected network and the second fully-connected network specifically comprise: The two fully-connected networks are respectively composed of a hidden layer and a scalar output layer, the hidden layer adopts a ReLU activation function to model a nonlinear hydrologic process, and the output layer adopts linear mapping to ensure the stability of gradient propagation, so that the whole model is ensured to have continuous microminiaturization in the process of counter propagation, and the joint optimization with a physical module is supported.
8. The flood peak enhanced physical base flow and residual correction collaborative runoff prediction method according to claim 5, wherein the S4 specifically comprises: s41, calculating predicted runoff by adopting residual error connection structure ; ; Wherein the method comprises the steps of Is the predicted flow; S42, calculating the loss between the predicted runoff and the real runoff by adopting an Adam optimizer, wherein the loss function comprises an evaluation item of the fitting precision of the whole sequence and a flood peak sample error: ; In the formula, For a true radial flow, For the nash efficiency coefficient, And (3) with Respectively an original runoff sequence and a flood peak threshold value of simulated runoffs; S43, performing iterative updating on all the learnable parameters in the model by using the calculated loss gradient through a back propagation algorithm, and realizing end-to-end joint optimization of the physical parameters and the deep learning network weights.
9. The flood peak enhanced physical base flow and residual correction collaborative runoff prediction method according to claim 8, wherein the step S5 specifically comprises: s51, verifying the model, and respectively calculating prediction indexes of a training set and a verification set, wherein the calculation indexes comprise Nash coefficients NSE, root mean square errors RMSE and relative errors RE; S52, an evaluation index calculation formula is as follows: ; ; ; In the middle of In order to observe the average value of the runoff, The number of the historical runoffs; S53, the closer NSE is to 1, the better the simulation effect, the closer RMSE is to 0, the smaller the degree of dispersion of the prediction error, the stronger the stability, Reflecting that the deviation of the total amount of runoff is small.

Description

Flood peak enhanced physical base flow and residual error correction cooperative runoff prediction method Technical Field The invention belongs to the technical field of hydrologic forecasting, and particularly relates to a flood peak enhanced physical base flow and residual error correction cooperative runoff forecasting method. Background Basin runoff prediction is a core topic in the hydrologic field and is important for flood control, disaster reduction and water resource management. The current mainstream prediction method comprises traditional hydrologic models such as a Xinanjiang model and a water tank model based on a physical equation and a data driving deep learning model represented by LSTM. The traditional model has definite parameters and clear mechanism, but is difficult to process complex nonlinear relation, the deep learning model has strong fitting capability, but lacks a physical basis, and the predicted result has the difficult problem of violating hydrologic rules. In recent years, how to combine the two has become a research hotspot. The fusion of the physical model and the neural network model mainly comprises the steps of taking output of the physical model as input characteristics of the neural network so as to combine mechanism constraint with nonlinear fitting capability, combining deep learning with traditional hydrologic model parameter calibration, solving the limitations that a traditional calibration method depends on experience and has poor regional adaptability, and trying to replace an empirical module in the physical model through the neural network so as to realize data-driven optimization of a local process. In practical application, the existing fusion method has obvious defects that on one hand, the output of a physical model is only one of input features of deep learning, the input features and the input features are relatively independent, and on the other hand, the gating mechanism of LSTM and other networks completely depends on data training, and key physical state information such as soil humidity, free water and the like cannot be effectively utilized, so that the model is caused to understand the hydrologic process and flow on the surface. Most of the methods stay at a loose coupling level, the physical process and the deep learning are still in a splitting state, the prediction result cannot clearly distinguish the physical basic flow and the data driving correction amount, and trust is difficult to establish in actual business. Therefore, an innovative method is needed to directly embed the hydrologic physical state into the deep learning network architecture, so that the physical mechanism and the data driver are organically unified at the algorithm level. The method not only can improve the prediction precision, but also can enhance the interpretability of the model, and provides more reliable decision support for the river basin management. Disclosure of Invention Aiming at the improvement demand of the prior art, the invention provides a flood peak enhanced physical base flow and residual error correction cooperative runoff prediction method, which aims at providing scientific basis for analyzing runoff prediction. In order to achieve the above purpose, the technical scheme adopted by the invention for solving the technical problems is that a physical base flow and residual error correction cooperative runoff prediction method for flood peak enhancement comprises the following steps: s1, collecting historical hydrological data of a river basin, including measured flow, rainfall data, potential evaporation data, minimum temperature, maximum temperature and steam pressure, dividing a training set and a verification set according to a proportion, and performing oversampling treatment on a flood peak sample to construct a time sequence window; s2, a normal differential equation discretization expression Xinanjiang model is adopted, each parameter in a differential equation is explicitly defined, a parameter constraint range is defined, rainfall and potential evaporation data are input, and intermediate variables with clear physical significance are obtained through physical model calculation; S3, selecting a plurality of intermediate quantity splicing sliding windows, the lowest temperature, the highest temperature and the water vapor pressure, entering an LSTM after normalization, extracting time sequence dynamic mode and nonlinear correction information, constructing a physical base stream and residual error correction dual-channel module, and calculating the physical base stream and residual error correction through a first full-connection network and a second full-connection network; S4, calculating predicted runoffs by adopting a residual error connection structure, calculating losses of the predicted runoffs and real runoffs, wherein a loss function comprises an evaluation item of fitting precision of an integral sequence, and updating physical paramet