Enhanced Artificial Neural Network Inflow Forecasting Algorithm for Run-of-River Hydropower Plants

Abstract

An improved artificial neural network-based algorithm for short-term water inflow forecasting (STWIF) into run-of-river hydropower plants is presented. Deregulation of the electrical power industry and introduction of competition between power producers have prompted redefinition of their daily operational tasks. In a competitive market environment, accurate short-term production planning and profitable bidding strategies become an important issue, requiring water inflow forecasts for up to 36 h ahead. As a result, short forecasting horizons have been found to be a main drawback of first-generation STWIF. An additional module using forecast precipitation data is developed for enlarging the forecast horizon up to two days ahead. The water inflow forecaster is further enhanced by inclusion of new input variables. With a large number of potential input variables, a new algorithm for selection of input variables using average mutual information and nonparametric density estimation is applied to the specific problem of water inflow forecasting. The performance of the enhanced STWIF is applied to the Soca River cascade hydropower system in Slovenia, and results are presented along with some comparisons with the previous-generation STWIF.