Energy Forecasting Using an Ensamble of Machine Learning Methods Trained Only with Electricity Data

This work presents case studies on forecasting PV power production and electricity demand in Portugal. We study an ensemble of different machine learning methods to exploit the growing collection of energy supply and demand records. The ensemble uses only electricity data to forecast, since this data is available online for any forecasting horizon. The ensemble relies on offline training and online forecasting, by applying the most recent power measurements to trained models. The different machine learning methods perform different non-linear transformations to the same electricity data, thus introducing diversity in the ensemble. To assess the forecasting performance of this system, we consider two forecasting horizons relevant to the Internal Electricity Market, namely 36 hours ahead, relevant to the single day-ahead coupling, and 2 hours ahead, relevant to the single intraday coupling. The forecasting performance using only electricity data compares gracefully with the state-of-the-art and improves the reference accuracy in our case studies. Since the ensemble relies only on energy data, the results show that machine learning methods are useful to exploit energy big data towards efficient energy forecasting systems.