Correction of Reservoir Runoff Forecast Based on Multi-scenario Division and Multi Models

Accurate runoff forecast is very important for reservoir operation. In view of the shortcomings of the existing correction models for runoff forecast, including the influence of the difference of external factors on the forecast results is not considered, and the optimal situation adaptation of different forecast models is not considered, three models, i.e., long and short-term memory neural network model (LSTM), gaussian process regression model (GPR) and support vector machine regression model (SVR), are used to forecast the relative errors of runoff forecast under different scenarios in this paper. The classification of forecast scenarios is determined based on factors such as rainfall, inflow, and foresight period, and two scenario sets are given, i.e., 12 forecast scenarios and 24 forecast scenarios. Then, a multi-model coupled runoff forecast correction method considering forecast error and forecast scenario is proposed. Through the case study of the Three Gorges Reservoir (TGR), it is found that, when the analysis is carried out based on the forecast period, the SVR model should be used for forecast correction when the foresight period is 1-5 days, and the LSTM model should be used for forecast correction when the foresight period is 6 days. The application effect of SVR and LSTM is better than GPR in the scenario set of 12 forecast scenarios. LSTM model has the highest accuracy of forecast correction in the scenario set of 24 forecast scenarios, and the mean value of the coefficient of certainty (R-2) changes from 0.919 of 12 forecast scenarios to 0.931 of 24 forecast scenarios, increasing by 1.31%. The mean value of mean relative error (MRE) changes from 6.80% of 12 forecast scenarios to 5.64% of 24 forecast scenarios, a decrease of 17.06%. Finally, the best model adaptation table corresponding to different forecast scenarios of TGR is established, which has an important guiding role in the actual runoff forecast of TGR.