Recurrent networks for wind speed forecasting

In recent years, wind power has prompted as a renewable energy source. However, integrating wind power into the electric grid is a major challenge due to the wind speed variations. Then, wind speed forecasting is an alternative for the pre-dispatch of power system. This paper proposes the forecast wind speed from 1 to 24 hours ahead using a multivariate series with several meteorological attributes. We use two recurrent networks models, which are virtually omitted in such problems: Long Short-Term Memory and Echo State Networks. To validate the experiments, we use one time series from northern Chile, a region that needs renewable energy sources to relieve mining sector. The results show that recurrent networks outperform a feedforward network forecasting several steps ahead in terms of two robust measures such as MdSE and MdAPE.