A benchmarking of machine learning techniques for solar radiation forecasting in an insular context

Philippe Lauret, Cyril Voyant, Ted Soubdhan, Mathieu David, Philippe Poggi
2015 Solar Energy  
In this paper, we propose a benchmarking of supervised machine learning techniques (neural networks, Gaussian processes and support vector machines) in order to forecast the Global Horizontal solar Irradiance (GHI). We also include in this benchmark a simple linear autoregressive (AR) model as well as two naive models based on persistence of the GHI and persistence of the clear sky index (denoted herein scaled persistence model). The models are calibrated and validated with data from three
more » ... h islands: Corsica (41.91°N; 8.73°E), Guadeloupe (16.26°N; 61.51°W) and Reunion (21.34°S ; 55.49°E). The main findings of this work are, that for hour ahead solar forecasting, the machine learning techniques slightly improve the performances exhibited by the linear AR and the scaled persistence model. However, the improvement appears to be more pronounced in case of unstable sky conditions. These nonlinear techniques start to outperform their simple counterparts for forecasting horizons greater than one hour. Praene et al., 2012) . Due to this high variability, the insular grids can experience a drop of around 40-50% of the PV power output in minutes. As a consequence, since the end of 2010, the French government has limited by law the total power produced by the instantaneous integration of intermittent renewables (PV and wind) into the insular grids, to 30%. Since 2011, this limit has been reached for Reunion Island and Corsica. In order to assure reliable grid operation and to balance the supply and demand of energy, utilities require accurate forecasts at different granularities and for different forecast horizons. For instance, short term forecasts are needed for operational planning, switching sources or re-scheduling of means of production, programming backup, planning for reserve usage, and peak load matching (Kostylev and Pavloski, 2011). Depending on the forecast horizon, different input data and forecasting models are appropriate. Statistical models with on-site measured irradiance are adequate for the very short-term time scale ranging from 5 minutes up to 6 hours (Lorenz and Heinemann, 2012). Forecasts based on cloud motion vectors from satellite images (Lorenz and Heinemann, 2012) show a good performance for a temporal range of 30 minutes to 6 hours. For forecast horizons from about 6 hours onwards, forecasts based on Numerical Weather Prediction (NWP) models are generally more accurate (Inness and Dorling, 2012; Maini and Agrawal, 2006; Muselli et al., 1998) . In this work, we assess the performance of different models for intraday solar forecasting with a special focus on the hour ahead solar forecast, as it is the most-common operational forecast requested by the French utility company when operating the insular grids. Consequently, in this work, light is shed on the use of statistical models. Indeed, the solar radiation sequence can be seen as a time series, and therefore one can build statistical models to capture the underlying random processes and predict the next values. Several statistical techniques can
doi:10.1016/j.solener.2014.12.014 fatcat:lbj6vvcn4faxhbj3biky5ua6ra