CN-121997812-A - River section flow prediction method and system based on deep learning

CN121997812ACN 121997812 ACN121997812 ACN 121997812ACN-121997812-A

Abstract

The invention relates to the technical field of river flow prediction, in particular to a river section flow prediction method and system based on deep learning. The method comprises the steps of obtaining upstream rainfall data, downstream water level data and weather forecast data of a target river basin, aligning the upstream rainfall data, the downstream water level data and the weather forecast data according to a unified time standard, constructing a time sequence alignment data set, constructing a flow prediction model based on the time sequence alignment data set, wherein the flow prediction model is used for carrying out feature fusion on the time sequence association feature vector and the space dependency vector by extracting the time sequence association feature vector between the upstream rainfall and the downstream water level, identifying the space dependency vector between a multi-branch sink point and a target section, and outputting a feature fusion vector. The method fundamentally overcomes the defect that the prediction result of the pure data driving model violates the basic physical rule, and remarkably improves the prediction reliability under the extreme rainfall event and the data loss scene.

Inventors

ZHOU BO
MOU YUN
MEI JUNYA
ZHI WEI
MAO BEIPING
DENG SHAN
HU JINGZHE
JIA ZHIWEI
ZHANG LI
HAN RUN

Assignees

长江水利委员会水文局
河海大学

Dates

Publication Date: 20260508
Application Date: 20260106

Claims (10)

1. The river section flow prediction method based on deep learning is characterized by comprising the following steps of: s1, acquiring upstream rainfall data, downstream water level data and weather forecast data of a target river basin; Step S2, aligning upstream rainfall data, downstream water data and weather forecast data according to a unified time standard, and constructing a time sequence alignment data set; s3, constructing a flow prediction model based on a time sequence alignment data set, wherein the flow prediction model performs feature fusion on the time sequence association feature vector and the space dependency vector by extracting the time sequence association feature vector between upstream rainfall and downstream water level, identifying the space dependency vector between a multi-branch inflow point and a target section, and outputting a feature fusion vector; Step S4, introducing mass conservation constraint and momentum conservation constraint in the training process of the flow prediction model; S5, generating a second flow predicted value of a target section in a future appointed period by using the trained flow predicted model, performing physical consistency verification on the second flow predicted value, and outputting predicted data comprising the predicted flow value and a verification result; And S6, transmitting the predicted data to a hydrologic monitoring management platform through a data interface.
2. The river cross-section flow prediction method based on deep learning of claim 1, wherein constructing the flow prediction model in step S3 comprises: Step S31, extracting time sequence correlation characteristic vectors between upstream rainfall data and downstream water level data based on the time sequence alignment data set, wherein the time sequence correlation characteristic vectors comprise accumulated rainfall characteristics of the upstream rainfall, change rate characteristics of the downstream water level and hysteresis correlation characteristics between the upstream rainfall and the downstream water level; Step S32, identifying a spatial dependency relation vector between the multi-branch inflow point and the target section, wherein the spatial dependency relation vector comprises the distance characteristic of each branch inflow point and the target section, the water collecting area characteristic of each branch inflow point and the elevation characteristic of each branch inflow point; And step S33, carrying out feature fusion on the time sequence associated feature vector and the space dependent relation vector, and outputting a feature fusion vector through vector splicing operation and full connection layer processing.
3. The deep learning-based river section flow prediction method of claim 2, wherein extracting a time-series correlation feature vector between upstream rainfall data and downstream water data comprises: Calculating cross-correlation values of upstream rainfall data and downstream water level data at a plurality of time lag points, and selecting the time lag with the largest cross-correlation value as dominant lag time; based on the dominant lag time, performing time migration on the upstream rainfall data, and calculating the accumulated quantity of the upstream rainfall after migration in a preset time window as a rainfall accumulation characteristic; Extracting the water level change rate of downstream water level data in the dominant lag time, wherein the water level change rate is obtained by calculating the average slope of a water level rising section and a water level falling section and is used as a water level response characteristic; The rain accumulation feature and the water level response feature are combined to form a time sequence correlation feature vector.
4. The depth learning based river profile flow prediction method of claim 2, wherein identifying a spatial dependency vector between a multi-tributary sink and a target profile comprises: acquiring a digital elevation model of a target river basin and water system network data; Identifying all tributary inflow points which are converged into the target section based on the digital elevation model of the target river basin and the water system network data; For each tributary converging point, calculating the straight line distance and the river-along distance between the tributary converging point and the target section as distance characteristics; Calculating the water collecting area of each branch inflow point, normalizing the water collecting area to be a percentage of the total water collecting area of the target section, and taking the percentage as the water collecting area characteristic; calculating the elevation difference between each branch inflow point and the target section, and calculating the average gradient based on the elevation difference and the distance to be used as elevation characteristics; combining the distance characteristic, the water collecting area characteristic and the elevation characteristic of each branch inflow point into a space characteristic subvector; And carrying out weighted average processing on the space characteristic sub-vectors of all the tributary merging points, and outputting a space dependency relation vector, wherein the weight is based on the historical flow contribution proportion of each tributary.
5. The river section flow prediction method based on deep learning of claim 1, wherein introducing mass conservation constraint in step S4 comprises: Predicting the feature fusion vector through a flow prediction model, and outputting a first flow prediction value; Acquiring a first observed water level corresponding to the first flow predicted value and a corresponding observed flow value; Calculating a first theoretical relation deviation between a first flow predicted value and an observed water level according to a preset continuous equation formula, and carrying out square sum calculation on the first theoretical relation deviation to output a continuous equation constraint term; Calculating a model prediction error term based on the first flow prediction value and the observed flow value; The continuous equation constraint term and the model prediction error term are combined in a weighted form to form a total loss function.
6. The deep learning-based river section flow prediction method of claim 1, wherein introducing momentum conservation constraints in step S4 comprises: performing time difference calculation on the first flow predicted value to obtain a predicted flow change rate; calculating the change rate of the downstream water level of the target section in a preset first period to obtain a downstream water level gradient; calculating a second theoretical relationship deviation between the predicted flow rate change rate and the downstream water level gradient according to a preset momentum equation formula, and performing square sum calculation on the second theoretical relationship deviation to output a momentum equation constraint term; And combining the momentum equation constraint term with the continuous equation constraint term and the data fitting error term in the total loss function in a weighted form to form the multi-constraint joint loss function.
7. The method for predicting river cross-section flow based on deep learning of claim 1, wherein the verifying of the physical consistency of the second flow prediction value in step S5 comprises: acquiring a second observed water level corresponding to the second flow predicted value; Calculating a third theoretical relationship deviation of the second observed water level corresponding to the second flow predicted value based on a preset continuous equation formula; If the third theoretical relation deviation exceeds a preset continuous equation threshold, marking the second flow predicted value as a first abnormal value; performing time difference calculation on the second flow predicted value to obtain a second predicted flow change rate; Calculating the change rate of the downstream water level of the target section in a preset second period to obtain a second downstream water level gradient; collecting historical observed flow data of a target section, and calculating 99% fraction and 1% fraction of the flow based on the historical observed flow data; If the second flow predicted value exceeds 99% fraction or is lower than 1% fraction of the historical observed flow, marking the second flow predicted value as a second abnormal value and eliminating the second abnormal value; And eliminating the first abnormal value and the second abnormal value from the predicted result.
8. The river cross-section flow prediction method based on deep learning of claim 1, wherein the predicted data outputted in step S5 further comprises: independently predicting the feature fusion vector for a plurality of times by starting the randomness of the Dropout layer in the flow prediction model to generate a plurality of groups of predicted flow results; counting the discrete degree of a plurality of groups of predicted flow results, and calculating the variance of the predicted flow; Calculating the average value of a plurality of groups of predicted flow results as a central predicted value; determining a distribution range of the predicted flow according to the mean value and the variance; Based on the distribution range, calculating a confidence interval range covering 95% probability, and outputting upper and lower limit values of the confidence interval.
9. The depth learning-based river section flow prediction method of claim 2, wherein the splicing operation of the time-series-associated feature vector and the spatial dependency vector comprises: calculating the association strength between each dimension in the space dependency relation vector and the target section; converting the association strength into normalized attention weight through a full connection layer and a Softmax function in the flow prediction model to form an attention weight matrix; weighting the space dependency relation vector by using the attention weight matrix, and outputting a weighted space dependency relation vector; Splicing the weighted spatial dependency relation vector and the time sequence associated feature vector to form a high-dimensional fusion feature vector; And reducing the dimension of the high-dimension fusion feature vector through the full connection layer, and outputting the feature fusion vector.
10. A deep learning-based river profile flow prediction system for performing the deep learning-based river profile flow prediction method of claim 1, the deep learning-based river profile flow prediction system comprising: the collecting module is used for obtaining upstream rainfall data, downstream water level data and weather forecast data of the target river basin; The alignment module is used for aligning the upstream rainfall data, the downstream water level data and the weather forecast data according to a unified time standard to construct a time sequence alignment data set; The fusion module is used for constructing a flow prediction model based on the time sequence alignment data set, wherein the flow prediction model is used for carrying out feature fusion on the time sequence association feature vector and the space dependency vector by extracting the time sequence association feature vector between the upstream rainfall and the downstream water level, identifying the space dependency vector between the multi-branch inflow point and the target section and outputting a feature fusion vector; The constraint module is used for introducing mass conservation constraint and momentum conservation constraint in the training process of the flow prediction model; The verification module is used for generating a second flow predicted value of the target section in a future appointed period by using the trained flow predicted model, carrying out physical consistency verification on the second flow predicted value, and outputting predicted data containing the predicted flow value and a verification result; and the pushing module is used for transmitting the predicted data to the hydrologic monitoring management platform through the data interface.

Description

River section flow prediction method and system based on deep learning Technical Field The invention relates to the technical field of river flow prediction, in particular to a river section flow prediction method and system based on deep learning. Background During a storm flood in a middle and small river basin, monitoring personnel need to rapidly calculate the flow of the downstream section so as to support flood control scheduling. Traditional hydrologic predictions are based mainly on conceptual hydrologic models and physical equations (such as the san-valan equation set), rely on accurate topographical parameters and boundary conditions, are difficult to model in complex flow domains and are computationally expensive. In recent years, the deep learning method has made remarkable progress in the field of flow prediction by virtue of its strong nonlinear fitting capability. However, the existing pure data driving model lacks physical mechanism constraint, and the prediction result often violates basic physical rules such as mass conservation, momentum conservation and the like, so that the prediction reliability is insufficient under rare events or data loss scenes. In addition, the prior art fails to fully integrate the spatial dependence relationship between upstream rainfall-downstream water level time sequence association and multiple tributaries, and has limited space-time feature extraction capability. Therefore, there is a need for a flow prediction method that organically integrates deep learning and physical mechanism, so as to improve prediction accuracy, ensure that the result meets the physical consistency of hydrology, and provide more reliable decision support for hydrology monitoring management. Disclosure of Invention Accordingly, the present invention is directed to a river section flow prediction method and system based on deep learning, so as to solve at least one of the above-mentioned problems. In order to achieve the above purpose, a river section flow prediction method based on deep learning comprises the following steps: s1, acquiring upstream rainfall data, downstream water level data and weather forecast data of a target river basin; Step S2, aligning upstream rainfall data, downstream water data and weather forecast data according to a unified time standard, and constructing a time sequence alignment data set; s3, constructing a flow prediction model based on a time sequence alignment data set, wherein the flow prediction model performs feature fusion on the time sequence association feature vector and the space dependency vector by extracting the time sequence association feature vector between upstream rainfall and downstream water level, identifying the space dependency vector between a multi-branch inflow point and a target section, and outputting a feature fusion vector; Step S4, introducing mass conservation constraint and momentum conservation constraint in the training process of the flow prediction model; S5, generating a second flow predicted value of a target section in a future appointed period by using the trained flow predicted model, performing physical consistency verification on the second flow predicted value, and outputting predicted data comprising the predicted flow value and a verification result; And S6, transmitting the predicted data to a hydrologic monitoring management platform through a data interface. Preferably, the present invention also provides a river profile flow prediction system based on deep learning, for performing the river profile flow prediction method based on deep learning as described above, the river profile flow prediction system based on deep learning comprising: the collecting module is used for obtaining upstream rainfall data, downstream water level data and weather forecast data of the target river basin; The alignment module is used for aligning the upstream rainfall data, the downstream water level data and the weather forecast data according to a unified time standard to construct a time sequence alignment data set; The fusion module is used for constructing a flow prediction model based on the time sequence alignment data set, wherein the flow prediction model is used for carrying out feature fusion on the time sequence association feature vector and the space dependency vector by extracting the time sequence association feature vector between the upstream rainfall and the downstream water level, identifying the space dependency vector between the multi-branch inflow point and the target section and outputting a feature fusion vector; The constraint module is used for introducing mass conservation constraint and momentum conservation constraint in the training process of the flow prediction model; The verification module is used for generating a second flow predicted value of the target section in a future appointed period by using the trained flow predicted model, carrying out physical consistency verification on the se