Data-Driven Construction of Local Models for Short-Term Wind Speed Prediction

Currently, there is a growing interest in improving the methods applied to the prediction of wind speed. In this document, we propose to combine physically-based decision rules, inferred through a data-driven process, with local regression models. Specifically, quantitative and qualitative analysis of historical records lead us to define a regression structure with a decision tree at the top and local regression models at each leaf. Specifically, our results suggest that this encoding improves the predictions for wind speed for a number of regression schemes, including radial basis neural networks, binary regression trees, support vector regression, adaptive network-based fuzzy inference systems, and bagging trees. A reduction of about 14% in the RMSE is shown for the latter.