Incremental Ensemble Learning for Electricity Load Forecasting

The efforts of the European Union (EU) in the energy supply domain aim to introduce intelligent grid management across the whole of the EU. The target intelligent grid is planned to contain 80% of all meters to be smart meters generating data every 15 minutes. Thus, the energy data of EU will grow rapidly in the very near future. Smart meters are successively installed in a phased roll-out, and the first smart meter data samples are ready for different types of analysis in order to understand the data, to make precise predictions and to support intelligent grid control. In this paper, we propose an incremental heterogeneous ensemble model for time series prediction. The model was designed to make predictions for electricity load time series taking into account their inherent characteristics, such as seasonal dependency and concept drift. The proposed ensemble model characteristics - robustness, natural ability to parallelize and the ability to incrementally train the model - make the presented ensemble suitable for processing streams of data in a "big data" environment.