Long-term streamflow forecast method and system based on process-data synergic drive

A long-term streamflow forecast system based on process-data synergic drive, comprising:

• a collecting part, configured to collect streamflow data, meteorological data, climate system indices, climate model forecasts, watershed elevation data, soil data and vegetation data;
• a calibrating and validating part, configured to construct a daily-scale Variable Infiltration Capacity (VIC) distributed hydrological model of a target watershed, and perform calibration and validation for model parameters;
• a correcting and disaggregating part, configured to, after the climate model forecasts is spatially interpolated to a grid scale matching the VIC distributed hydrological model, firstly use a linear scaling method to perform bias correction on the climate model forecasts, and then use a k-nearest neighbor algorithm-based model to disaggregate monthly-scale meteorological data forecasted by a climate model into daily data;
• a calibration period runoff yield calculation part, configured to, by using the spatially-interpolated and disaggregated climate model forecasts, drive the VIC distributed hydrological model to perform a gridded runoff yield calculation within a calibration period;
• an improving part, configured to, with observed monthly streamflow values of an outlet hydrological station within the calibration period as dependent variables and with monthly data obtained by aggregating runoff yield outputs of all grids within a watershed in different lead times as independent variables, select a third layer of soil moistures with the runoff yield outputs as daily scale, and use a LASSO regression model to perform feature variable screening, and then take grids corresponding to the screened-out runoff yield outputs as typical grids to construct an improved VIC model;
• a full time period runoff yield calculation part, configured to drive the improved VIC model with the disaggregated climate model forecasts to perform the gridded runoff yield calculation of a full time period;
• a candidate predictor obtaining part, configured to re-aggregate daily-scale gridded runoff yield calculation results of the improved VIC model into monthly-scale data and combine the monthly-scale data with historical streamflows, the climate model forecasts, the meteorological data and the climate system indices to form a candidate predictor set of a data-driven model, wherein different lag times are to be considered for the influence of the historical streamflows, the meteorological data and the climate system indices on the forecast monthly streamflows;
• a composite model forming part, configured to perform normalization on the candidate predictor set and the monthly streamflow data respectively and then screen predictors by using a machine learning model;
• then, train a deep learning model using the screened-out predictors to further obtain a composite model based on process-data synergic drive;
• a streamflow forecasting part, configured to use the composite model based on process-data synergic drive to perform long-term streamflow forecast;
• a controlling part, configured to communicate with the collecting part, the calibrating and validating part, the correcting and disaggregating part, the calibration period runoff yield calculation part, the improving part, the full time period runoff yield calculation part, the candidate predictor obtaining part, the composite model forming part, and the streamflow forecasting part respectively to control operations thereof;
• wherein, in the improving part, the observed daily streamflow values of the outlet hydrological station within the calibration period and the runoff yield outputs of all grids within the watershed in different lead times are firstly extracted;
• then, normalization is performed on monthly data obtained by aggregating both the streamflows and the runoff yield outputs to eliminate a dimensional influence between variables;
• with the monthly streamflow values as dependent variables and with the runoff yield outputs of the grids as independent variables, the LASSO regression model is used to perform feature variable screening, wherein a target function set in the form of least square is:
• min
• ⁢
• {
• ∑
• i
• =
• 1
• m
• (
• y
• i
• -
• β
• 0
• -
• ∑
• j
• =
• 1
• n
• x
• i
• ⁢
• j
• ⁢
• β
• j
• )
• +
• λ
• ⁢
• ∑
• j
• =
• 1
• n
• ❘
• "\[LeftBracketingBar]"
• β
• j
• ❘
• "\[RightBracketingBar]"
• }
• in the formula, x and y correspond to the independent variables and the dependent variables respectively, m corresponds to a dependent variable series length, n corresponds to a number of grids of the target watershed, β0 is a constant item, β is a coefficient of each variable, λ is a penalty parameter;
• the grids corresponding to the screened-out runoff yield outputs are taken as typical grids.