Feature selection in wind speed prediction systems based on a hybrid coral reefs optimization – Extreme learning machine approach S. Salcedo-Sanz a,⇑ , A. Pastor-Sánchez a , L. Prieto b , A. Blanco-Aguilera c , R. García-Herrera c,d a Department of Signal Processing and Communications, Universidad de Alcalá, Alcalá de Henares, Spain b Department of Energy Resource, Iberdrola, Spain c Department of Atmospheric Physics, Universidad Complutense de Madrid, Spain d IGEO, Instituto de Geociencias (UCM-CSIC), Madrid, Spain article info Article history: Received 28 August 2013 Accepted 15 June 2014 Available online 18 July 2014 Keywords: Short term wind speed prediction Feature selection problem Coral reefs optimization algorithm Extreme learning machines abstract This paper presents a novel approach for short-term wind speed prediction based on a Coral Reefs Opti- mization algorithm (CRO) and an Extreme Learning Machine (ELM), using meteorological predictive vari- ables from a physical model (the Weather Research and Forecast model, WRF). The approach is based on a Feature Selection Problem (FSP) carried out with the CRO, that must obtain a reduced number of predic- tive variables out of the total available from the WRF. This set of features will be the input of an ELM, that finally provides the wind speed prediction. The CRO is a novel bio-inspired approach, based on the sim- ulation of reef formation and coral reproduction, able to obtain excellent results in optimization prob- lems. On the other hand, the ELM is a new paradigm in neural networks’ training, that provides a robust and extremely fast training of the network. Together, these algorithms are able to successfully solve this problem of feature selection in short-term wind speed prediction. Experiments in a real wind farm in the USA show the excellent performance of the CRO–ELM approach in this FSP wind speed pre- diction problem. Ó 2014 Elsevier Ltd. All rights reserved. 1. Introduction Wind power is currently the most important renewable energy source in the world in terms of annual growing and economic impact [1,2]. The installed wind power worldwide by the end of 2013 reached a total of 318 GW, with a few leading countries bet- ting for this technology: China (91 GW), the USA (61 GW), Ger- many (34 GW), Spain (23 GW) or India (20 GW) [3], and many others in which wind energy is considered as the future source of alternative energy out of conventional sources, such as Denmark (25% of wind energy penetration), Portugal (16%), Ireland (12%), Italy (4%) or France (3%). This wind energy booming around the world has brought new problems in the management and mainte- nance of wind farm facilities [4]. One of this important problems is the integration of wind energy in the energy transportation net- work, where the prediction of the generated power in wind farms is a key problem, influenced by the variability of the wind in the short and medium terms. Thus, wind speed prediction is a basic task performed in all wind farms facilities as part of their operation management. There are two types of approaches that have been used to carry out wind speed prediction in wind farm facilities. First, pure statis- tical approaches consider only previous wind speed series in one or several towers to construct a predictor for the wind speed in the near future. These approaches sometimes include meteorological variables measured at the prediction area, in order to enhance the wind speed prediction. In any case, these approaches do not take into account the atmospheric dynamics in order to make the wind speed prediction. In the last few years, many different statis- tical approaches have been applied to wind speed prediction, including linear prediction models [5], classical Box–Jenkins meth- odologies such as auto-regressive models [6] and other time series analysis such as the Mycielski algorithm [7], different clustering algorithms [8,9], and several modern computational approaches such as neural networks [10–13], neural networks ensembles [14], Bayesian methods [15], support vector machines [16,17], or combinations of different statistical models: neural networks and auto-regressive models [18], auto-regressive models and Kalman filtering [19], neural networks and Markov models [20], and wave- lets and neural approaches [21,22]. http://dx.doi.org/10.1016/j.enconman.2014.06.041 0196-8904/Ó 2014 Elsevier Ltd. All rights reserved. ⇑ Corresponding author. Address: Department of Signal Processing and Commu- nications, Universidad de Alcalá, 28871 Alcalá de Henares, Madrid, Spain. Tel.: +34 91 885 6731; fax: +34 91 885 6699. E-mail address: sancho.salcedo@uah.es (S. Salcedo-Sanz). Energy Conversion and Management 87 (2014) 10–18 Contents lists available at ScienceDirect Energy Conversion and Management journal homepage: www.elsevier.com/locate/enconman