Hourly Discharge Modelling and Forecast for a Run-of-river Dam

Water resources have become a growing concern in society. This is largely due to the scarcity of this natural asset and the realisation that increasing demand could lead to future conflicts. Sometimes, human action limits access to water or alters natural flows. Run-of-river hydropower schemes manage river flows on a short-term basis, altering the natural flow of rivers according to the energy needs of consumers or the risk of flooding. The aim of this work is to show how to model and predict the hourly flow in a run-of-river reservoir, using the Crestuma-Lever dam on the river Douro (Portugal) as a case study. Data collected from 1998 to 2020 will be used. The study focuses on the use of time series models capable of dealing with multiple periodicities, such as the TBATS model. The findings show that the model can be used for 48-hour to weekly forecasting. In general, it captures the large fluctuations in the turbine discharges and most peak discharges. However, it does not capture most zeros and has difficulty in dealing with low flow values. The results of the time-series model are also compared with those obtained using three machine learning algorithms: the Seasonal Naïve, the Neural Network, and the Random Forest.