Multi-step forecasting strategies for wind speed time series

Abstract

A time series is a sequence of observations, measured at certain moments in time, ordered chronologically and evenly spaced, so that the data are usually dependent on each other. Currently, time series are used to estimate wind gusts, which are highly non-linear, unknown, and at times unpredictable. A good estimation of wind gusts implies correct planning on the generation of clean wind energy. In this work, we use Artificial Intelligence (AI) techniques such as the use of convolutional neural networks for wind gust estimation. One of the best models for dealing with this type of information is the Large Short Term Memory (LSTM) network because it is a type of recurrent network that specializes in sequence information. In this work, an LSTM prediction model is implemented for five different wind speed data sets using different multi-step forecasting strategies. The strategies used are Recursive, Direct, MIMO (multiple-input to multiple-output), DIRMO (Combination of direct strategy and MIMO), and DirREC (Combination of direct and recursive strategy).