Adaptive Reservoir Inflow Forecasting Using Variational Mode Decomposition and Long Short-Term Memory

Abstract

Developing an accurate and timely inflow forecast model continues to be an integral part of modern reservoir operation and has been greatly promoted by the combination of data preprocessing and data-driven techniques. This paper integrates variational mode decomposition (VMD) and long short-term memory (LSTM) into a hybrid model named A-VMD-S-LSTM. After dividing data for calibration and validation periods, A-VMD-S-LSTM decomposes the calibration inflow time series into subseries and cast them into a single LSTM forecast model to output final inflow predictands directly. When new inflow data are appended to the calibration time series, the VMD and LSTM models will be updated to adapt to all available inflow information to avoid transmitting future inflow information. The parameter settings of VMD and LSTM and the meaningful lags corresponding to each subseries are well identified by grid search and partial autocorrelation function (PACF), respectively. Finally, the proposed model is applied to the 1-, 3-5-, and 7-day ahead inflow forecasts of the Three Gorges Reservoir (TGR) along with five competitive models. Results show that A-VMD-S-LSTM achieves the best forecast accuracy and fourth computational efficiency among the six candidates, demonstrating its best comprehensive forecast performance. Furthermore, using two or three LSTM layers in A-VMD-S-LSTM is redundant since the change in forecast accuracy is negligible and the increase of run time is remarkable. Therefore, a practical reservoir inflow forecast model with high forecast accuracy and relatively low computational cost is provided in this paper. INDEX TERMSReservoir inflow, adaptive forecasting, variational mode decomposition, long short-term memory, parameter calibration.