Short Term PV Power Forecasting Using ELM and Probabilistic Prediction Interval Formation

Abstract

Accurate prediction of solar power is important to the grid operator for ensuring energy management from multiple sources without jeopardizing stability and to the PV plant owner for scheduling plant maintenance periods and avoiding power imbalance costs. It is evident that meteorological data like solar irradiance is more readily available than the historical PV power output series with hourly samples. In this case, indirect forecasting can be utilized where PV output predictions are obtained using the solar irradiance forecasts. Decisions solely based on point forecasts can be risky considering the sharp variations in solar irradiance patterns. The inherent uncertainty in the point forecasts can be quantified by associating them with a probability distribution to form prediction intervals (PIs) which is a more interpretable representation of uncertainty. This paper presents a probabilistic forecasting approach using a nonparametric PI formation method based on Extreme Learning Machine. No prior assumption on the error distribution is required for the PI formation. Solar irradiance data from NUS geography weather station, Singapore, is analyzed and assembled into two separate sets for better model performance. Coverage probability and interval scores are evaluated for the resulting PIs which show promising results.