Day-ahead inflow forecasting using causal empirical decomposition

Abstract

It is essential to have accurate and reliable daily-inflow forecasting to improve short-term hydropower scheduling. This paper proposes a Causal multivariate Empirical mode Decomposition (CED) framework as a complementary pre-processing step for a day-ahead inflow forecasting problem. The idea behind CED is combining physics-based causal inference with signal processing-based decomposition to get the most relevant features among multiple time-series to the inflow values. The CED framework is validated for two areas in Norway with different meteorological and hydrological conditions. The validation results show that using CED as a pre-processing step significantly enhances (up to 70%) the forecasting accuracy for various state-of-the-art forecasting methods. • Pre-processing step is essential for daily inflow forecasting. • A combination of causal-based inference network with decomposition technique is proposed as CED framework. • Multivariate empirical mode decomposition, transfer entropy and conditional mutual information were used in CED framework. • CED framework improved daily inflow forecasting by 70% compared with no preprocessing scenario.