Bayesian Deep Learning-based Confidence-aware Solar Irradiance Forecasting System

Abstract

For stable and successful use of grid-connected PV (photovoltaic) plants, it is quite necessary to know the expected power from PV plants in advance. However, forecasting PV output power accurately is difficult in practical cases where uncertainties are unavoidable. In this paper, we propose a confidence-aware forecasting system that produces a point forecast together with its confidence information. Our system classifies forecast outputs into confident forecasts and non-confident forecasts using the confidence information. Then, the confident forecast is used directly and the non-confident forecast is replaced by its lower bound, which is desirable for conservative scheduling of existing power plants. Through the experiments, we show that MAPE (maximum absolute percentage error) of the confident forecasts and the non-confident forecasts are 9.8% and 21.5%, respectively. We also show that the lower bound is lower than actual value in over 95% of the non-confident forecasts. The results show that our approach is good to classify forecasts into confident forecasts and non-confident forecasts and to produce effective lower bounds.