Nonlinear autoregressive neural networks for forecasting wind speed time series

Increasingly, wind electric generation is becoming indispensable to complement the global energy matrix. In this way, it is necessary to develop forecasting methods for this type of renewable resource, in order to guarantee the appropriate use to be compatible with the needs of the electricity sector. In this paper, two hybrid models are proposed that combine Autoregressive Integrated Moving Average methodologies (ARIMA) with that of artificial neural networks (ANNs) to forecast time series of daily average wind speeds from data collected at 10 meters. The first model is a nonlinear autoregressive neural network, NARNET, of order (3,3,1) and the second a NARNET of order (4,3,1). The training of the models occurred with data collected from two cities in the northeast of Brazil, Acaraú and Guaramiranga, during a period of 100 days and the forecast occurred for a period of 126 consecutive days after the training. The forecast horizon is average, around days for 3 months. Moreover, the choice of the best forecasting model occurred through the analysis of statistical accuracy indices, among which are: the mean absolute error (MAE), the mean square error (MSE), the mean absolute percentage error (MAPE), the model suitability (FIT) and the correlation coefficient (R) and of determination (R2). The forecast proved to be efficient in the localities of Acaraú with MSE errors of 0,2118 m/s, MAPE 10,75%, R 0,8685 and also for the locality of Guaramiranga MSE 0,2234 m/s, MAPE 12,16% and R 0,7233. In addition, the forecast curves for both locations showed good agreement with the observed values. Finally, the possibility of further improving hybrid modeling is discussed by changing the activation functions of the hidden layer.