CN-122022080-A - Reservoir flood forecasting method based on physical constraint and space-time double-flow coupling

CN122022080ACN 122022080 ACN122022080 ACN 122022080ACN-122022080-A

Abstract

The invention discloses a reservoir flood forecasting method based on physical constraint and space-time double-flow coupling, which comprises the steps of S1, obtaining multi-source hydrological data, S2, decomposing the multi-source hydrological data into a historical state sequence and a future driving sequence, S3, extracting features of the historical state sequence data through a physical enhancement long-short time memory network of a historical inertia feature extraction branch to obtain historical inertia features, extracting features of the future driving sequence through a time domain convolution network of a future forced feature extraction branch to obtain future forced features, S4, generating fusion features by weighting and fusing the historical inertia features and the future forced features, S5, inputting the fusion features into a decoder to obtain a predicted water level increment, and superposing the predicted water level increment to a current water level to obtain a predicted value. The prediction method improves the timeliness of prediction, improves the prediction precision of the water withdrawal stage, and enhances the physical consistency of the prediction result.

Inventors

WANG MINGMING
LIU BINBIN
ZHU XIAOLEI
HONG YIQUAN
HUANG JIN
FANG DONG
ZHANG WANBO

Assignees

安徽省（水利部淮河水利委员会）水利科学研究院（安徽省水利工程质量检测中心站）

Dates

Publication Date: 20260512
Application Date: 20260413

Claims (10)

1. The reservoir flood forecasting method based on physical constraint and space-time double-flow coupling is characterized by comprising the following steps: s1, acquiring multi-source hydrological meteorological data of a reservoir river basin; S2, decomposing the multi-source hydrological meteorological data into a historical state sequence and a future driving sequence; S3, performing feature extraction on the historical state sequence data through a historical inertia feature extraction branch to obtain historical inertia features By future forced feature extraction branching, for future drive sequences Extracting features to obtain future forced features ; S4, historical inertia characteristics And future forcing features Generating fusion features by weighted fusion ; S5, fusing the characteristics Input decoder to obtain predicted water level increment And is superimposed to the current water level to obtain a predicted value ; In step S3, the history inertia characteristic extraction branch adopts a physical enhancement long-short time memory network, and the future forced characteristic extraction branch adopts a time domain convolution network.
2. The reservoir flood forecasting method based on physical constraint and space-time double flow coupling as claimed in claim 1, wherein the input gate of the physical enhanced long-short time memory network is as follows: In the formula, To physically enhance the state of the input gate activation, The function is activated for Sigmoid, And (3) with As a weight matrix of the input gates, Is that The historical state of the moment of time is a multi-source input vector, For the last moment Is a hidden layer state vector of (c), As a result of the bias term, Representing the hadamard product; for the time index within the historical time sliding window, For the moment of the current prediction initiation, For the length of the history observation window, Represent the first The average rainfall of the surface of the observation basin at the moment, Indicating time of day Is used for the flow-producing gain factor of (1), And (3) with Respectively a weight matrix and a bias term which can be learned in the neural network, Is a proxy variable for river basin soil saturation.
3. The reservoir flood forecasting method based on physical constraint and space-time double flow coupling as claimed in claim 1, wherein in the model initialization stage, forgetting gate bias parameters of the physical enhanced long-short time memory network are used Set as intervals Any constant within.
4. The reservoir flood prediction method based on physical constraint and space-time double flow coupling as claimed in claim 1, wherein the model is subjected to constraint training through a mixed loss function, wherein the mixed loss function comprises a water balance physical regularization loss value The following formula: Wherein, the For a batch size delineated at each network training, For the index of the samples in the current batch, For the moment of the current prediction initiation, For the future foreseeable window length, For future time index of the foreseeable window, Representation of The numerical weather forecast rainfall at the moment, Representation of The reservoir schedule at the moment plans the delivery flow, For a net flux of the liquid, Is that A predicted value of the water level at the moment, Is that The dam front water level observed at the moment, Is the predicted total variation of the water level.
5. The reservoir flood prediction method based on physical constraint and space-time double flow coupling as claimed in claim 4, wherein the mixing loss function The following formula is given: In the formula, Is of batch size; Indexing for the samples; And Respectively for model number Predicted water level and actual observed water level of each sample output; is a weight balance coefficient.
6. The reservoir flood forecasting system based on the coupling of physical constraint and space-time double flow is characterized by comprising a data acquisition module, a data decomposition module, a double-flow heterogeneous feature extraction module, a feature fusion module, a decoding module and a prediction calculation module, The data acquisition module is used for acquiring multi-source hydrometeorological data of a reservoir river basin, The data decomposition module is used for decomposing the multi-source hydrological meteorological data into a historical state sequence and a future driving sequence, The dual-flow heterogeneous feature extraction module comprises a historical inertia feature extraction branch and a future forced feature extraction branch which are arranged in parallel, wherein the historical inertia feature extraction branch is used for carrying out feature extraction on historical state sequence data to obtain historical inertia features The future forced feature extraction branch is used for extracting features of a future driving sequence to obtain future forced features The history inertia characteristic extraction branch adopts a physical enhanced long-short time memory network, the future forced characteristic extraction branch adopts a time domain convolution network, The feature fusion module is used for dynamically calculating the contribution weights of the two branches and generating fusion features through weighted fusion , The decoding module is used for fusing the characteristics Decoding and outputting the water level increment of the current prediction step , The prediction calculation module calculates the water level increment according to the current prediction step Calculating a predicted water level 。
7. The reservoir flood forecast system based on the coupling of physical constraints and space-time double flow according to claim 6, wherein the physical enhanced long-short time memory network corrects the calculation formula of the standard input gate to: 。
8. The reservoir flood forecast system based on physical constraints and spatiotemporal double flow coupling of claim 6, further comprising a hybrid loss function for constraint training of models, the hybrid loss function comprising a water balance physical regularization loss value The following formula: 。
9. Reservoir flood forecasting equipment based on coupling of physical constraint and space-time double flow is characterized by comprising a memory, a processor and a reservoir flood forecasting program which is stored on the memory and can run on the processor and is based on coupling of physical constraint and space-time double flow, wherein the reservoir flood forecasting program based on coupling of physical constraint and space-time double flow is configured with a reservoir flood forecasting method based on coupling of physical constraint and space-time double flow as claimed in any one of claims 1-5.
10. A storage medium, wherein a reservoir flood forecasting program based on coupling of physical constraint and space-time double flow is stored on the storage medium, and the reservoir flood forecasting program based on coupling of physical constraint and space-time double flow realizes the reservoir flood forecasting method based on coupling of physical constraint and space-time double flow according to any one of claims 1-5 when being executed.

Description

Reservoir flood forecasting method based on physical constraint and space-time double-flow coupling Technical Field The invention relates to the technical field of hydraulic engineering safety monitoring and hydrologic and water resource management, in particular to a method for carrying out high-precision real-time rolling forecasting on a reservoir water level evolution process by utilizing historical hydrologic monitoring data and forecast period weather scheduling data in combination with a physical mechanism constraint and deep learning space-time modeling technology. Background The reservoir is used as a key hub for flood control, water supply and power generation in a river basin, and the accurate forecast of the running water level change is a precondition for making scientific scheduling decisions. With the aggravation of global climate change, the sudden and intensity of the extreme rainfall event is obviously increased, and extremely high requirements are put on the response speed and the forecasting precision of the reservoir flood control dispatching system. Particularly in the flood season, the evolution of the reservoir water level is affected by complex nonlinearity of multiple factors such as upstream rainfall converging, interval inflow, manual flood discharge scheduling and the like, and how to realize real-time water level prediction with long prediction period, high precision and compliance with the hydraulic law is always a difficulty to be solved in the field of hydrologic prediction. The existing reservoir flood forecasting method is mainly divided into two main types, namely a physical cause model and a data driving model. Traditional physical cause models, such as a Xinanjiang model, a semi-distributed hydrological model (TOPMODEL) based on a topography index and the like, generalize a yield convergence process based on a hydrologic mechanism, have definite physical meaning, but the models usually contain a large number of experience parameters which need to be calibrated, are sensitive to changes of conditions under a flow field, have high computational complexity, and are difficult to meet the requirement of modern flood control scheduling on second-level real-time response. In recent years, data-driven models represented by long and short term memory networks (LSTM) have been widely used in hydrographic time series prediction due to their strong nonlinear fitting ability. The model can automatically learn the law of water level change from massive historical monitoring data, and has the advantages of high calculation speed and strong generalization capability. However, existing data-driven forecasting models still have significant technical limitations in practical engineering applications. Firstly, the traditional model mostly adopts a 'pure history driving' mode, namely, only the rainwater condition data at the past moment is used for deducing the future, the future rainfall information provided by numerical weather forecast (NWP) and the established scheduling plan of the reservoir are ignored, so that obvious phase lag phenomenon often occurs in the rising stage of flood of the model, and sufficient operation time cannot be reserved for flood control and rescue. Secondly, the general deep learning model essentially belongs to a 'black box' model, lacks the constraint of a hydrologic physical mechanism, aims at minimizing fitting errors only in model training, easily causes a predicted result to violate a basic physical conservation law, predicts a false phenomenon that the water level greatly rises under the condition of no effective rainfall and inflow, and seriously influences the trust degree of a dispatcher on the model result. In addition, when the gating mechanism of the standard LSTM unit is used for processing a long-duration flood, the memory of the early-stage high water storage potential energy is easy to decay rapidly along with time, so that the simulation of the water withdrawal process is often lower than an actual measurement value, and the huge water body inertia characteristics of a large reservoir are difficult to accurately reflect. Therefore, development of a novel reservoir flood forecasting method capable of effectively fusing future multi-source forced information and having physical interpretability and constraint mechanism is needed. The foregoing is provided merely for the purpose of facilitating understanding of the technical solutions of the present invention and is not intended to represent an admission that the foregoing is prior art. Disclosure of Invention The invention aims to provide a reservoir flood forecasting method based on physical constraint and space-time double-flow coupling, which aims to overcome the technical defects of response lag, physical mechanism deletion, long-duration memory attenuation and the like in sudden flood forecasting of the existing data driving model, and realize real-time rolling forecasting of a reservoir wa