Data-driven short-term forecasting of solar irradiance profile

Abstract

Autoregressive and non-linear models are frequently used for solar irradiance forecasting, and their results are found to vary depending on the forecast horizon, time resolution and location. Given that higher accuracies are often reported at lower time resolutions and areas with less cloud cover, it was found necessary to separately assess the applicability of models for tropical Singapore where high solar irradiance variabilities are experienced.This paper presents a short-range forecasting system of 30-min irradiance averages for 0.5 to 6 hours ahead based on per-min data of solar irradiance and ambient temperature. In addition, it explores the possibility of predicting volatility by looking at the distribution of solar irradiance in the next 30-min period with a novel approach that estimates the proportion of points within each of 21 bands defined to cover the range of irradiance. With it, upper and lower bound predictions for the period are obtained to calculate upside and downside risks posed by photovoltaic (PV) generation. Using persistence models for comparison and assessing accuracy across 8 locations, all models showed marked improvement, especially at longer forecast horizons. On average, MAE of point forecast models decreased by 9% (98 to 89 W/m2) and 58% (299 to 125 W/m2) for the 0.5 and 6-hour horizons respectively. For volatility models, MAE decreased from 4.8 to 3.7% in proportion predictions while errors of making upper and lower bound predictions outside the actual range of per-min fluctuations decreased from 43.0 to 10.4% and 30.6 to 3.4% respectively.