Dam inflow time series regression models minimising loss of hydropower opportunities

Abstract

Recently, anomalies in dam inflows have occurred in Japan and around the world. Owing to the sudden and extreme characteristics of rainfall events, it is very difficult to predict dam inflows and to operate dam outflows. Hence, dam operators prefer faster and more accurate methods to support decision-making to optimise hydroelectric operation. This paper proposes a machine learning method to predict dam inflows. It uses data from rain gauges in the dam regions and upstream river level sensors from previous three hours and predicts the dam inflow in the next three hours. The method can predict the time of rise to the peak, the maximum level at the peak, and the hydrograph shapes to estimate the volume. The paper presents several experiments applied to inflow time series data for 10 years from the Kanto region in Japan, containing 55 floods events and 20,627 time stamped dam in-flow points measured at 15-min intervals. It compares the performance of four regression prediction models: generalised linear model, additive generalised linear model, regression tree and gradient boosting machine and discusses the results.