A comparison of deep learning methods for wind speed forecasting

Currently, deep learning methods are being used and proposed to deal with the problem of wind speed time series forecasting. This is since they have good forecast accuracy; however, they also have greater complexity and there is an increase in the computational effort used in comparison with the conventional forecasting methods. This paper reviews the deep learning methods most widely used in time series forecasting, such as convolutional neural networks, long short-term memory networks, and hybrid methods. The results are compared against the autoregressive integrated moving average (ARIMA) method, which is typically used, due to its simplicity and high precision. A benchmark was generated based on a wind speed time series obtained from a meteorological station, obtaining hourly forecasts one step ahead and subsequently obtaining forecasts of several steps ahead. The results show the improvement in the accuracy in the forecast obtained when using the methods based on deep learning, as compared with the ARIMA method.