A Multi-Phase Ensemble Model for Long Term Hourly Load Forecasting

Abstract

Long-term projection of electricity demand is necessary for strategizing production, transmission, distribution and grid expansion in power systems. In this work, we propose a model for forecasting hourly profile of load data which must be taken into consideration by power system planners to produce cost optimal and realizable solutions. The developed ensemble model is formulated in two phases, with the initial phase primarily centered on stacking of gradient and adaptive boosting regressors. In the subsequent phase, the variance is diminished by bagging Lasso LARS regressor on the stacked dataset. For implementation of the proposed model, we collect real-world data of the Germany electricity market for thirteen years spanning from 2006 to 2018. Electricity demand forecasts have been evaluated for the duration of five-years from 2014 to 2018 and are found to be extremely accurate as well as consistent. The presented model on comparison with five benchmark load forecasting models is observed to surpass all of them with a mean absolute percentage error of 1.59 on the test set. Furthermore, unlike neural network models, the proposed ensemble is computationally inexpensive with a training time of 110s.