Comparison of Statistical Techniques as Applied to Short Term Wind Power Forecasting

In this study, several machine learning and statistical models are used for short-term wind energy forecasting. The forecasts of these models, which include conditionally weighted least squares, kernel regressions, k-nearest neighbor, neural networks, random forests and support vector machines, are compared and analyzed at the 1-hour forecast horizon. The use of similar data-driven techniques in short-term wind power forecasting has received a lot of attention in recent literature, previous studies have mostly focused on forecasts beyond 3-hour horizons as surveyed in Bessa et al. (2009) and Foley et al. (2012) due to the difficulties that arise at the 1-horizon. Input data, from both online and offline sources, may be noisy at this horizon leading to inconsistent forecasts. Data preprocessing and feature selection help attenuate these problems and improve the quality of the forecasts. We validate these models using different metrics to evaluate their forecasting performance as well their ability in identifying wind power ramp events which are important for wind power operators, utilities and system operators for balancing and grid stability. While the results of these techniques do vary, the forecasts generated by the statistically trained models generally demonstrate performance advantages over the persistence forecast.