Feature selection in machine learning prediction systems for renewable energy applications

This paper focuses on feature selection problems that arise in renewable energy applications. Feature selection is an important problem in machine learning, both in classification and regression problems. In renewable energy systems, feature selection appears related to prediction systems in the most important sources such as wind, solar and marine resources. The objective of the paper is twofold: first, a review of the most important prediction systems for renewable energy applications involving feature selection is carried out. Analysis and discussion of different feature selection problems in prediction systems are considered. We show that wrapper FSP approaches are those mostly used due to their higher performance. They include a diversity of algorithms, prevailing fast-training approaches. The lack of an uniform framework for FSP and the diversity of tackled problems impede a systematic assessment of the performance and properties of the applied methods. Thus, the simultaneously use of several global search mechanisms should be the preferred option. In a second part of the paper, we explore this possibility, by introducing a novel approach for feature selection based on a novel meta-heuristic, the Coral Reefs Optimization algorithm with Substrate Layer. This approach is able to combine different search mechanisms into a single algorithm, providing a global search procedure of high quality. We use an Extreme Learning Machine for prediction within this novel approach. The performance of the system is evaluated in a problem of wind speed prediction from numerical models input, using real data from a wind farm in Spain, where comparison with alternative regression algorithms is carried out. Improvements up to 20% in hourly and daily wind speed prediction are obtained with the proposed system versus the algorithms without the feature selection process considered.