CN-122020134-A - Intelligent prediction method for medium-long-term runoff

CN122020134ACN 122020134 ACN122020134 ACN 122020134ACN-122020134-A

Abstract

The invention relates to the technical field of hydrological forecasting and discloses a medium-long-term runoff intelligent forecasting method which comprises the steps of collecting river basin data, preprocessing, extracting features, generating a feature vector sequence, dividing a training set, constructing a physical constraint cyclic neural network module, training based on a total loss function of physical constraint loss, constructing a sliding window online self-adaptive correction module, initializing the sliding window to store actual measurement runoff values and preliminary runoff forecasting values at the latest O moments, inputting the feature vector at the current moment into the trained physical constraint cyclic neural network module, generating the preliminary runoff forecasting values, calculating historical average deviation according to window states, correcting to obtain a final forecasting result, finally, forming a new sample pair by the actual measurement runoff value and the preliminary runoff forecasting values, adding the new sample pair into the window, removing the oldest sample, and realizing window dynamic updating.

Inventors

YAN YIQI
XIE LI
LAN YUNLONG
LU JIAQI
WANG YAJIE
ZHANG XINGDONG
WANG YUHENG
WANG CHUNQING
XU KEYAN
FANG XIUQIN
Zhu Qiuan
JIANG SHANHU
ZHANG NAN
ZHANG LI
REN XIAOFENG

Assignees

河海大学
黄河水利委员会水文局

Dates

Publication Date: 20260512
Application Date: 20260414

Claims (6)

1. The intelligent prediction method for the medium-long-term runoff is characterized by comprising the following steps of: The method comprises the steps of constructing a medium-long-term runoff intelligent forecasting model, wherein the intelligent forecasting model comprises a physical constraint cyclic neural network module and a sliding window online self-adaptive correction module; collecting meteorological, hydrological and underlying data of a target river basin, carrying out data preprocessing and feature extraction to generate a feature vector sequence, and dividing a training set, a verification set and a test set according to a time sequence; inputting the feature vector sequence of the training set into a physical constraint cyclic neural network module, constructing a total loss function based on physical constraint loss, training the physical constraint cyclic neural network module, and generating a trained physical constraint cyclic neural network module; Initializing a sliding window with a fixed size in the sliding window on-line self-adaptive correction module, wherein the sliding window is used for storing the measured runoff values at the latest O moments and the corresponding preliminary runoff forecast values; Inputting the feature vector at the current moment into a trained physical constraint cyclic neural network module, generating a preliminary runoff forecast value at the current moment, and calculating the average deviation of the history forecast in the sliding window according to the state of the sliding window; Correcting the preliminary runoff forecast value at the current moment based on the average deviation of the history forecast in the sliding window, generating a corrected runoff forecast value, and taking the corrected runoff forecast value as a medium-long-term runoff intelligent forecast result at the current moment; And forming a new sample pair by the actual measurement runoff value at the current moment and the preliminary runoff forecast value at the current moment, adding the new sample pair into a sliding window to update the sliding window, and finally realizing intelligent updating and forecasting of the medium-long-term runoff.
2. The intelligent prediction method for medium-long runoff according to claim 1, wherein the process of collecting meteorological, hydrological and underlying data of a target river basin, performing data preprocessing and feature extraction to generate a feature vector sequence, and dividing a training set, a verification set and a test set according to time sequence is as follows: Collecting meteorological, hydrological and underlying surface data of a target river basin, and generating preprocessed meteorological, hydrological and underlying surface data after outlier rejection, missing value complementation and normalization; extracting precipitation, air temperature, evaporation and actual measurement runoff of the preprocessed meteorological, hydrological and underlying surface data to generate a basic meteorological hydrological feature vector, namely: Wherein, the Is the first The basis meteorological hydrological feature vector of the moment, Is the first The precipitation amount at the moment of time, Is the first The air temperature at the moment of time, Is the first The amount of evaporation at the moment in time, Is the first The actual measured runoff quantity at the moment, Is a four-dimensional real vector space; Extracting actual measurement runoff of the preprocessed meteorological, hydrological and underlying surface data at a plurality of moments to generate a hysteresis feature vector, namely: Wherein, the Is the first The lag feature vector of the moment in time, 、、 Respectively the first Time of day, the first Time of day, the first The actual measured runoff quantity at the moment, Is a six-dimensional real vector space; Based on the preprocessed meteorological, hydrological and underlying data, extracting physical feature vectors including early-stage influence rainfall features and soil humidity index features, namely: Wherein, the Is the first The physical feature vector of the moment in time, Is the first The early stages of the moment of time influence the rainfall characteristics, Is the first The characteristic of the soil humidity index at the moment, Is a two-dimensional real number vector space, To reflect the rate of soil moisture decay for the decay factor, Is the first The early stages of the moment of time influence the rainfall characteristics, In order to accumulate the time steps, In order for the attenuation coefficient to be a factor, Is the attenuation coefficient of To the power of the two, Is the first Precipitation at the moment; Splicing the basic meteorological hydrological feature vector, the hysteresis feature vector and the physical feature vector to generate a feature vector sequence, namely: Wherein, the As a sequence of feature vectors, 、、 Respectively the 1 st time, the 2 nd time and the 2 nd time The feature vector of the moment in time, For the transpose operation, Is that A dimension real number vector space; and dividing the characteristic vector sequence into a training set, a verification set and a test set according to the time sequence.
3. The medium-long runoff intelligent forecasting method of claim 1, wherein the feature vector sequence of the training set is input into a physical constraint cyclic neural network module, a total loss function based on physical constraint loss is constructed, the physical constraint cyclic neural network module is trained, and the process of generating the trained physical constraint cyclic neural network module is as follows: Feature vector sequence of training set A cyclic neural network of physical constraints is input, wherein, 、、 Respectively the 1 st time, the 2 nd time and the 2 nd time in the training set Feature vector of moment; introducing physical constraint loss in a physical constraint cyclic neural network, including prediction loss, water balance loss and monotonicity constraint loss, and constructing a total loss function based on the physical constraint loss, namely: Wherein, the As the value of the total loss function, For predicting loss, the method is used for measuring the difference between the actual measured radial flow value and the preliminary radial flow forecast value output by the physical constraint cyclic neural network, In order to balance the loss of the water quantity, In order to be a monotonic constraint loss, Is a weight coefficient of the water balance constraint, Is a weight coefficient of the monotonicity constraint, As the total number of time instants, Is the first The measured runoff value at the moment, Is the first The preliminary runoff forecast value output by the physical constraint cyclic neural network at the moment, Is the change amount of the water storage capacity of the soil, Is the first The precipitation amount at the moment of time, Is the first The amount of the transpiration at the moment, Is the first The radial flow is deep at the moment, For the length of the time period, In order to be a flow field area, In order to operate at the maximum value, The partial derivative of the preliminary runoff forecast value output by the physical constraint cyclic neural network on the precipitation; The physical constraint cyclic neural network performs forward propagation through a gating cyclic mechanism to generate a preliminary runoff forecast value, namely: Wherein, the Is the first The update gate of the moment of time, For the sigmoid activation function, 、、 Respectively updating the weight matrix of the gate, resetting the gate and candidate hidden states, 、、 Bias items of the update gate, the reset gate and the candidate hidden states respectively, Is the first A reset gate at the moment of time, 、 Respectively the first Time of day, the first The hidden state of the moment of time, For the hyperbolic tangent activation function, Is the first The feature vector that is input at the moment, Is the first The candidate hidden state of the moment in time, For the element-by-element multiplication operation, In order to output the layer weight matrix, Bias items for the output layer; And updating the weight parameters and the bias items of the physical constraint cyclic neural network through back propagation, minimizing the total loss function, and finally generating the trained physical constraint cyclic neural network.
4. The intelligent prediction method for medium-long runoff according to claim 1, wherein the process of calculating the average deviation of the history prediction in the sliding window according to the state of the sliding window is as follows: Judging whether the sliding window is empty, if so, judging the average deviation of the history forecast in the sliding window If the historical forecast is 0, judging whether the sliding window is full, if so, calculating the average deviation of the historical forecast in the sliding window by adopting an equal weight average method, namely: Wherein, the Is the first The average deviation of the history forecast within the time sliding window, Is the size of the sliding window; Otherwise, calculating average deviation of history forecast in the sliding window by adopting an exponential weighted average method, namely: Wherein, the For the total number of time instants within the current window, Is an attenuation factor.
5. The intelligent prediction method of medium-long term runoff according to claim 4, wherein the calculation formula of the corrected runoff prediction value is: Wherein, the Is the first And (5) a runoff forecast value corrected at the moment.
6. The intelligent medium-long term runoff forecasting method of claim 5, wherein a new sample pair is formed by an actual measurement runoff value at the current moment and a preliminary runoff forecasting value at the current moment, a sliding window is added to execute updating of the sliding window, and finally the intelligent medium-long term runoff updating and forecasting process is realized as follows: Forming a new sample pair by the actual measurement runoff value at the current moment and the preliminary runoff forecast value at the current moment And adding the sliding window to generate an updated sliding window, namely: Wherein, the The updated sliding window is obtained; Judging whether the updated sliding window size exceeds a preset upper limit value O, if so, removing the oldest sample pair in the updated sliding window to generate a re-updated sliding window, namely: Wherein, the In order to update the sliding window again, Is the first The measured radial flow value at the moment, Is the first Preliminary runoff forecast values at time.

Description

Intelligent prediction method for medium-long-term runoff Technical Field The invention relates to the technical field of hydrologic forecasting, in particular to an intelligent forecasting method for medium-long-term runoff. Background The medium-long term runoff forecasting is a core technical support for water resource scientific regulation and control and flood control and disaster reduction, and the forecasting accuracy directly influences the running efficiency of hydraulic engineering and regional water safety. In the prior art, runoff forecasting methods are mainly divided into three types: The first type is a physical hydrologic model, describes a hydrologic physical process based on partial differential equations such as a san-Vinan equation set, has definite physical significance, but has the problems of difficult parameter calibration, high calculation resource consumption, weak real-time forecasting capability and the like. The second type of data driving model mainly comprises machine learning models such as LSTM, SVM, RF and the like, and can effectively mine nonlinear relations in data, but has the following limitations that an LSTM model is good at capturing time sequence dependency relations, but is easy to fit and lacks of physical interpretability, an SVM model has strong generalization capability in small sample data, but is difficult to adapt to nonlinear characteristics of a complex hydrologic process, an RF model can process high-dimensional data redundancy, and a prediction result is easy to be influenced by parameters such as the number of decision trees and the like. The third type is a hybrid model, and combines a physical model with a data driving method, but the existing hybrid model is fixed after training is completed, and lacks self-adaptive capability for climate change and non-stationarity of hydrologic process. In conclusion, the prior art has the problems of limited forecasting precision, insufficient physical consistency, lack of on-line self-adaption capability and the like. Disclosure of Invention Aiming at the defects in the prior art, the invention provides a medium-long term runoff intelligent forecasting method, which improves the accuracy, physical consistency and dynamic adaptability of medium-long term runoff forecasting by organically integrating a physical constraint cyclic neural network and online self-adaptive correction of a sliding window and solves the problems of limited accuracy, insufficient physical consistency and lack of online self-adaptive capacity of medium-long term runoff forecasting in the prior art. In order to achieve the aim of the invention, the invention adopts the following technical scheme: an intelligent forecasting method for medium-long term runoff comprises the following steps: The method comprises the steps of constructing a medium-long-term runoff intelligent forecasting model, wherein the intelligent forecasting model comprises a physical constraint cyclic neural network module and a sliding window online self-adaptive correction module; collecting meteorological, hydrological and underlying data of a target river basin, carrying out data preprocessing and feature extraction to generate a feature vector sequence, and dividing a training set, a verification set and a test set according to a time sequence; inputting the feature vector sequence of the training set into a physical constraint cyclic neural network module, constructing a total loss function based on physical constraint loss, training the physical constraint cyclic neural network module, and generating a trained physical constraint cyclic neural network module; Initializing a sliding window with a fixed size in the sliding window on-line self-adaptive correction module, wherein the sliding window is used for storing the measured runoff values at the latest O moments and the corresponding preliminary runoff forecast values; Inputting the feature vector at the current moment into a trained physical constraint cyclic neural network module, generating a preliminary runoff forecast value at the current moment, and calculating the average deviation of the history forecast in the sliding window according to the state of the sliding window; Correcting the preliminary runoff forecast value at the current moment based on the average deviation of the history forecast in the sliding window, generating a corrected runoff forecast value, and taking the corrected runoff forecast value as a medium-long-term runoff intelligent forecast result at the current moment; And forming a new sample pair by the actual measurement runoff value at the current moment and the preliminary runoff forecast value at the current moment, adding the new sample pair into a sliding window to update the sliding window, and finally realizing intelligent updating and forecasting of the medium-long-term runoff. Further, collecting meteorological, hydrological and underlying surface data of a target river basin,