CN-122022085-A - Scene flood forecasting method considering similar water coupling deep learning

CN122022085ACN 122022085 ACN122022085 ACN 122022085ACN-122022085-A

Abstract

The invention discloses a scene flood forecasting method considering similar water coupling deep learning, and relates to the technical field of hydrologic forecasting. The method comprises the steps of obtaining target flood and historical flood sets to be forecasted, calculating mechanism similarity weights of the target flood and the historical flood based on feature vectors of each field of historical flood in the target flood and the historical flood sets, calculating effective sample numbers according to the similarity weights, adaptively screening the set of calibration samples, eliminating artificial interference time periods by using a disturbance gating mechanism, constructing weighted objective function calibration model parameters, and carrying out amplitude correction and peak time alignment on a preliminary forecasting sequence of a physical model by using a data driving model configured with a time-shifting layer. The method solves the problems that the traditional similarity matching neglects the physical mechanism and the deep learning is difficult to correct the phase error, and improves the accuracy and the robustness of flood forecasting.

Inventors

ZHANG YE
WANG SHUANG
XUAN SHENGWEI
WANG LEIZHI
LI XITING
SU XIN
LI LINGJIE
LIU YONG
HU JIANWEN
Yun Zhaode
CHEN ZHAOYI

Assignees

水利部交通运输部国家能源局南京水利科学研究院

Dates

Publication Date: 20260512
Application Date: 20260416

Claims (10)

1. A method of session flood forecast considering similar incoming water coupling deep learning, comprising: Acquiring target flood to be forecasted and a historical flood set containing multiple historical floods; Based on the characteristic vector read by each field of historical flood in the target flood and the historical flood set, calculating the similarity characteristic of the target flood and each field of historical flood, and determining the similarity weight corresponding to each field of historical flood according to the similarity characteristic; Screening a plurality of field history floods meeting preset similar conditions from the history floods set according to the similar weights to form a rating sample set; Performing parameter calibration on a pre-configured hydrologic model by using a calibration sample set to obtain a forecast parameter adapting to target flood; driving a hydrologic model to simulate target flood based on the forecast parameters to obtain a preliminary flow forecast sequence; The method comprises the steps of obtaining associated hydrological meteorological data affecting target flood formation, inputting a preliminary flow forecasting sequence and the associated hydrological meteorological data into a pre-trained data driving error correction model, and carrying out error correction processing on the preliminary flow forecasting sequence to obtain a final flood forecasting result.
2. The method of claim 1, wherein calculating the similarity feature of the target flood and each historical flood based on the feature vector read by each historical flood in the set of target flood and historical flood, specifically comprises: Driving a hydrological model by using preconfigured priori parameters, respectively inputting a precipitation sequence of target flood and a precipitation sequence of each field of historical flood in a historical flood set into the hydrological model for forward simulation, and outputting a corresponding simulation flow process and an internal state track set, wherein the internal state track set at least comprises one of a tension water storage track, a free water storage track and a water diversion source outflow track; The method comprises the steps of carrying out statistics calculation on an internal state track set to obtain the mean value and standard deviation of each state variable as statistical characteristics, calculating the duty ratio of each water diversion source outflow component to obtain the water diversion source duty ratio statistical characteristics, respectively constructing normalized mechanism signature vectors for target flood and each field of history flood based on the statistical characteristics of the internal state track set and the water diversion source duty ratio statistical characteristics, and taking the mechanism signature vectors as characteristic vectors for calculating similarity.
3. The method according to claim 2, wherein determining the similarity weight for each field history flood accordingly comprises: Calculating Euclidean distance between the mechanism signature vector of the target flood and the mechanism signature vector of each field of history flood in the history flood set to obtain a similar distance; mapping the similar distance into a normalized value by using an exponential decay function containing a temperature coefficient to obtain a similar weight corresponding to each field of historical flood; Wherein, the smaller the similarity distance, the larger the corresponding similarity weight.
4. The method of claim 3, wherein the step of screening a plurality of field history floods satisfying a preset similarity condition from the history floods set according to the similarity weight to form a calibration sample set comprises the following steps: Calculating the effective sample number reflecting the discrete degree of the weight based on the similar weight of each field of history flood in the history flood set, and rounding the effective sample number to obtain the target sample number; Sorting the historical flood sets from large to small according to the similarity weights, selecting the historical floods which are arranged in descending order of the similarity weights and are positioned in the front and have the number equal to the number of the target samples, and forming a rating sample set; wherein the number of valid samples is inversely proportional to the sum of squares of similar weights.
5. The method of claim 4, wherein the parameter calibration of the pre-configured hydrologic model using the set of calibration samples to obtain the forecast parameters adapted to the target flood comprises: driving a hydrological model by using parameters to be calibrated to simulate each field of historical flood in a calibrated sample set to obtain a simulated flow process, and calculating a difference value between the simulated flow process and a corresponding actually measured flow process to obtain a simulation error; constructing a weighted calibration objective function, wherein the weighted calibration objective function comprises a weighted error term, and the weighted error term is formed by accumulating the product of the simulation error of each field of history flood in the calibration sample set and the corresponding similar weight; Searching a parameter combination minimizing a weighting calibration objective function by adopting a global optimization algorithm, and determining the parameter combination as a forecast parameter; the prediction parameters are used for driving the hydrologic model to generate a preliminary flow prediction sequence of the target flood.
6. The method of claim 5, further comprising, prior to constructing the weighted scaling objective function: Driving a hydrological model to simulate historical flood in a calibration sample set by using a prestored priori parameters, calculating the difference value between the simulated flow and the actually measured flow, and constructing a disturbance residual sequence; carrying out standardization processing on the disturbance residual sequence to generate a gating mask for identifying the artificial scheduling interference time period; The method specifically comprises the steps of calculating simulation errors of the calibration sample set only for the time period of which the gating mask is marked as non-interference, and constructing the weighting calibration objective function by combining similar weights.
7. The method of claim 1, wherein the data-driven error correction model is configured with a micro-time-shift layer, and wherein the error correction processing is performed on the preliminary traffic forecast sequence, and the method specifically comprises: Inputting the preliminary flow forecasting sequence and the associated hydrological meteorological data into a data driving error correction model, and outputting an amplitude correction sequence and a time-lapse variable in parallel; And carrying out translation transformation on the amplitude correction sequence on a time axis according to the time-shift quantity by utilizing the time-shift layer, so as to realize peak time alignment and obtain a final flood forecast result.
8. The method of claim 7, wherein the data-driven error correction model is trained based on a joint loss function, comprising: The joint loss function comprises an amplitude fitting term and a time-shift regularization term; The amplitude fitting term represents an error between a final flood forecast result and an actual measurement flow process, and the time shift regularization term represents an absolute size of a time shift variable and is used for restraining the time alignment amplitude.
9. The method of claim 1, wherein the feature vector comprises a maximum N hours precipitation, a normalized precipitation time-course sequence, and a sub-basin surface average precipitation vector, and wherein the calculating the similarity feature of the target flood and each field of history flood comprises: calculating the similarity of the total amount of strong precipitation based on the difference value of the maximum precipitation amount for N hours; calculating precipitation time similarity based on the pearson correlation coefficient of the normalized precipitation time sequence; Based on the correlation coefficient of the average precipitation vector of the sub-drainage basin surface, the spatial similarity is calculated.
10. The method of claim 9, wherein determining the similarity weight for each field history flood based thereon comprises: Sequencing the historical flood sets in three dimensions of strong precipitation total similarity, precipitation time similarity and spatial similarity respectively to obtain sequencing positions of each set of historical flood in the three dimensions; calculating the difference between the maximum value and the minimum value of the sequencing positions of each field of history flood in three dimensions to obtain sequencing conflict degree; weighting and summing the similarity of three dimensions by using a predetermined dimension weight parameter to obtain a basic score, and correcting the basic score by using a penalty term based on the sorting conflict degree to obtain a similarity weight corresponding to each field of history flood; wherein, the greater the ordering conflict degree, the heavier the penalty the similarity weight receives.

Description

Scene flood forecasting method considering similar water coupling deep learning Technical Field The invention belongs to the technical field of hydrologic forecasting, and particularly relates to a scene flood forecasting method considering similar water coupling deep learning. Background The session flood forecast is a core support for flood control and disaster reduction decisions, and has important significance for reducing flood disaster loss. Currently, a hydrologic model (such as Xin An Jiang Moxing) based on a physical mechanism and a data-driven deep learning model are two main tools for flood forecasting, and a loose coupling mode combining the hydrologic model and the data-driven deep learning model becomes an important means for improving forecasting accuracy. In the existing research, the parameter calibration of a model is often assisted by historical flood data. The common practice is to select a fixed number (e.g. 5 fields) of most similar historical floods as samples based on the statistical characteristics of external observation data such as rainfall total, rainfall correlation coefficient and the like, and determine the current forecast parameters in a weighted or average mode. In the error correction link, a convolutional neural network or a long-short-term memory network CNN-LSTM is directly adopted, the residual error of the simulated flow and the actually measured flow of a physical model is taken as a target, and the point-to-point amplitude regression correction is carried out on the forecasting result. However, the prior art still has limitations in terms of similarity mechanism characterization and phase error correction, and there is still room for improvement in terms of accurate prediction requirements in complex scenes. The method mainly comprises the following deep technical problems that firstly, similarity matching is stopped on an external image statistics layer, consistency of internal state variables such as soil water content, groundwater accumulation and the like is ignored, the evolution consistency of the internal state variables such as the soil water content, the groundwater accumulation and the like cannot be guaranteed only by rainfall characteristic similarity, a selected sample is similar and is different, the reflecting precision of a current production convergence mechanism is required to be improved, secondly, deep learning correction lacks explicit processing capacity on peak time deviation, the existing model is subjected to amplitude fitting based on mean square error MSE loss, when a flood simulated by a physical model has time lag or advances (phase error), forced fitting can lead to model leveling rather than alignment of the flood to cause forecasting failure, and thirdly, robust recognition on artificial disturbance such as reservoir scheduling and the like is lacking, and the process of setting disturbed data pollution parameters is caused. Disclosure of Invention The invention aims to provide a field flood forecasting method considering similar water coupling deep learning, so as to solve the problems in the prior art. According to the technical scheme, the method for forecasting the scene flood by considering similar water coupling deep learning comprises the following steps: Acquiring target flood to be forecasted and a historical flood set containing multiple historical floods; Based on the characteristic vector read by each field of historical flood in the target flood and the historical flood set, calculating the similarity characteristic of the target flood and each field of historical flood, and determining the similarity weight corresponding to each field of historical flood according to the similarity characteristic; Screening a plurality of field history floods meeting preset similar conditions from the history floods set according to the similar weights to form a rating sample set; Performing parameter calibration on a pre-configured hydrologic model by using a calibration sample set to obtain a forecast parameter adapting to target flood; driving a hydrologic model to simulate target flood based on the forecast parameters to obtain a preliminary flow forecast sequence; The method comprises the steps of obtaining associated hydrological meteorological data affecting target flood formation, inputting a preliminary flow forecasting sequence and the associated hydrological meteorological data into a pre-trained data driving error correction model, and carrying out error correction processing on the preliminary flow forecasting sequence to obtain a final flood forecasting result. The method has the beneficial effects that the problems that the traditional similarity matching ignores a physical mechanism and the deep learning is difficult to correct the phase error are solved, and the accuracy and the robustness of flood forecasting are improved. Drawings Fig. 1 is a flowchart of steps of a method for forecasting a scene flood considering sim