Forecasting of Reservoir Inflow by the Combination of Deep Learning and Conventional Machine Learning

Abstract

Recently, forecasting of inflow to a reservoir by employing machine learning techniques is getting attention for maximization of generation of hydroelectricity and prevention of disaster caused by flooding. In this context, forecasting of peak inflow caused by heavy rain or melting of snow is of utmost importance to properly utilize water for hydroelectricity, limit the damage to the reservoir/dam and issue flood advisory. The conventional methods of forecasting, which build a forecasting model by using the reservoir inflow data with observed and/or forecasted weather data by minimizing the overall forecasting errors, fail to predict the peak inflow because the number of peak inflows is very small compared to number of normal inflows. To forecast the peak inflow more accurately, we propose a method called Peak-oriented Forecasting by Model Switching (PForMS) which is a hybrid approach that utilizes the forecasting capabilities of both conventional machine learning and Deep Learning-based methods. To show the effectiveness of the proposed method, we perform experiments with five publicly available dam data sets from Japan and the USA and evaluate the proposed method and existing methods in term of AUC (Area under the ROC Curve) of forecasting peak inflow of various heights. Experimental results suggest that our proposed method is able to forecast the higher peaks of inflow more accurately than the existing methods of conventional machine learning and Deep Learning.