Wind speed reconstruction using a novel Multivariate Probabilistic method and Multiple Linear Regression: advantages compared to the single correlation approach

Abstract Multivariate methods can improve wind speed prediction accuracy under specific conditions. In this work, Multiple Linear Regression (MLR) and Multiple Mass Probability (MMP) models are compared to the single correlation approach in selected tests. The novel MMP method can be trained using an unlimited number of inputs. Time series of outcome variables are then reconstructed using multiple probability functions. These functions consist of a main component, obtained from an input discrete event, and marginal components, obtained using a few nearest bins for the input variables; the use of the marginal components allows improving the algorithm's precision. A significant improvement of the wind speed prediction (by up to 52% for MMP) is achieved using four reference points. As rule of thumb, the target point is equidistant from the reference sites. In this condition, both multivariate methods perform better than the single correlation approach. This is also confirmed when estimating the Weibull parameters in a long-term period of ten years and the theoretical energy produced by two kinds of Wind Turbine Generators. Further applications of MMP concern wind power production forecast and, more in general, nonlinear relationships. In these cases, the direct use of MLR can lead to lack of accuracy.