The technology of airborne wind turbines is rapidly growing, with the purpose of overcoming working limitations of wind turbines at low altitudes. High-altitude wind is strong enough for efficient power generation, but wind conditions vary. Wind-speed forecasting in real time is necessary for power generation or flight stabilization. This study investigates three widely used forecasting models with a single-step and multistep ahead scheme for short-term wind-speed prediction from historical wind measurement data: a persistence model, an autoregressive moving average (ARMA) model, and an artificial neural network (ANN). In the single-step scheme, the accuracy of the persistence model dramatically decreases as the time horizon increases; nevertheless, the persistence model is the simplest algorithm to implement. The ARMA model and the ANN yield a significant accuracy of wind-speed forecasting, compared with the persistence model. The overall mean absolute errors (MAEs) of ARMA and ANN are 19.78% and 22.69% lower than that of the persistence method, respectively. The lowest errors were found in ANNs for most cases of time horizon lengths. Unlike ANNs, the setup of the ARMA model is systematical. A few time horizons can be recommended for short-term wind-speed forecasting for an airborne wind turbine. However, for a long time horizon, the multistep ahead forecasting scheme is recommended since the overall MAEs from the ARMA and ANN are reduced by 4.70% and 11.88%, respectively.
Natapol Korprasertsak, T. Leephakpreeda
Journal name not available for this finding