Evaluation of neural network algorithms, and time-series models and SARIMA-SETAR hybrid model in Monthly wind speed prediction

The aim of this study is modeling and predicting the monthly wind speed of Dezful. Therefore, SARIMA model, threshold nonlinear time series model (SETAR), SARIMA-SETAR hybrid model, and also artificial neural networks were used. In addition, the PMI algorithm was used for selecting the effective input variables in predicting wind speed for the neural network model. Using the Hempel criterion and the AIC information criterion monthly relative humidity in the previous two months RH (t-2), monthly in the previous month evaporation (t-1), the average monthly temperature in the previous three months Tave (t-3) and maximum monthly temperature in the previous month Tmax (t) as effective input variables for modeling and predicting monthly wind speed were identifie .TO validate the SARIMA and SETAR models and the fitted SARIMA-SETAR hybrid model, the functions of autocorrelation, partial autocorrelation and the residual independence test of the model (Ljung-Box) were used. The superiority models ware determined based on the minimum numerical value of Schwartz and AIC statistics. For modeling and predicting wind speed with neural network, linear function was used for input and output layer and different stimulus functions with different training algorithms were used for the hidden layer.The neural network model with topology (1-3-5) with sigmoid tangent stimulus function and Levenberg-Marquardt training algorithm was found to have better performance in predicting the monthly wind speed of Dezful synoptic station compared to SARIMA linear and SETAR nonlinear models. The results of the models showed that the SARIMA-SETAR hybrid model (2, 2, 3) had better performance compared to other models. It has acceptable accuracy in predicting the monthly wind speed of the Dezful synoptic, with a coefficient of determination of 0.91 and a root of mean squares 72.