Short-Term Power Forecasting of Solar PV Systems Using Machine Learning Techniques

Roof-top mounted solar photovoltaic (PV) systems are becoming an increasingly popular means of incorporating clean energy into the consumption profile of residential users. Electric utilities often allow the inter-connection of such systems to the grid, compensating system owners for electricity production. As the systems grow in number and their contribution to the overall load profile becomes increasingly significant, it becomes imperative for utilities to accurately account for them while planning and forecasting generation. Additionally, this information is useful for system-owners who want to optimize their production schedule. We use various machine learning and statistical techniques to train models on solar irradiance data and different meteorological parameters to forecast solar irradiance, and therefore power, for different forecasting horizons in the short-term future. We begin with a series of naïve models – linear regression, locally-weighted linear regression and support-vector regression (SVR) – where we only use meteorological information in the present to make future predictions, and then progress to time-series modeling. We find that both linear and locally-weighted linear regression perform rather poorly in the naïve case; we consider only SVR (both the regular and least-squares variation) along with conventional statistical models such as the seasonal auto-regression integrated moving-average (ARIMA) model in the time-series implementation. Our best results, with an RMSE of 40.16 W/m2, are obtained from using least-squares SVR (LS-SVR) with an RBF kernel, trained on solar irradiance data and meteorological features in the 7 hours prior to the present time t, and from 24 and 48 hours prior to the forecasting time t+ tf , where tf is the forecasting horizon. This model performs better than existing models in literature which use the same dataset.