Predictive analysis of run‐of‐river power plants

Abstract

This thesis explores several methods for short-term forecasting of the power generation of a run-of-river hydropower plant. The goal is to create forecasts of the power generation for one day in advance. With power generation and historical weather data ARIMA(X), linear regression and generalized linear regression models are estimated and compared to each other. We employed feature engineering and derived new features from the weather data. The set of possible inputs is huge, since there are different kinds of weather data and derived time series. Thus we have to determine the inputs most important for predicting the power generation to create useful models. Therefore, stepwise regression, a feature selection method, is used to find the most influential (lagged) variables. To answer the question which models can create the best forecast, the performance of the different models is evaluated using cross-validation and the performance measures (root) mean squared error ((R)MSE), mean absolute error (MAE) and median absolute error (MdAE).