Feature selection for accurate short-term forecasting of local wind-speed

There is increasing demand for accurate short-term forecasting of weather conditions at specified locations. This demand arises partly from the growing numbers of renewable energy facilities. In order successfully to integrate renewable energy supplies with grid sources, the short term (e.g. next 24 hrs) output profile of the renewable system needs to be forecast as accurately as possible, to avoid over-reliance on fossil fuels at times when renewables are available, and to avoid deficit in supply when they aren't. In particular, the inherent variability in wind-speed poses an additional challenge. Several approaches for wind-speed forecasting have previously been developed, ranging from simple time series analysis to the use of a combination of global weather forecasting, computational fluid dynamics and machine learning methods. For localized forecasting, statistical methods that rely on historical location data come to the forefront. Recent such work (building localized forecast models with multivariate linear regression) has found that accuracy can gain significantly by learning from multiple types of local weather features. Here, we build on that work by investigating the potential benefits of simple additional 'derived' features, such as the gradient in wind-speed or other variables. Following extensive experimentation using data from sites in Nigeria (primarily), Scotland and Italy, we conclude that the ideal forecasting model for a given location will use a judicious combination of direct and derived features.