Innovative hybrid models for forecasting time series applied in wind generation based on the combination of time series models with artificial neural networks

This work shows two innovative hybrid methodologies capable of performing short and long term wind speed predictions from the mathematical junction of two classical time series models the Autoregressive Integrated Moving Average with Explanatory Variable (ARIMAX) and the Holt-Winters (HW), both combined with Artificial Neural Networks (ANN). The first hybrid model (ARIMAX and ANN) is made from the physical relations between pressure, temperature and precipitation with the wind speed, that is, this model is considered as multivariate. The second hybrid model (HW and ANN) is considered as univariate, i.e. allowing only wind speed inputs. By means of statistical analysis of error it is verified that the proposed hybrid models offer perfect adjustments to the observed data at the regions of study, and thus, better comparisons with traditional ones from the literature. It is possible to find in this analysis percentage error of 5.0% and efficiency coefficient (Nash-Sutcliffe) of approximately 0.96. The confirmation of accuracy by the hybrid models reveals that they provide time series that are able to follow the observed time series profiles with similarities of maximum and minimum values between both series. Therefore, it became an important indicative in the representation of characteristics of seasonality by the models.