TY - JOUR
T1 - One day ahead wind speed forecasting
T2 - A resampling-based approach
AU - Zhao, Weigang
AU - Wei, Yi Ming
AU - Su, Zhongyue
N1 - Publisher Copyright:
© 2016 Elsevier Ltd
PY - 2016/9/15
Y1 - 2016/9/15
N2 - Wind speed forecasting plays a vital role in dispatch planning and operational security for wind farms, however, its difficulty is commonly accepted. This paper develops a nonlinear autoregressive (exogenous) model for one-day-ahead mean hourly wind speed forecasting, where general regression neural network is employed to model nonlinearities of the system. Specifically, this model is a two-stage method consisting of the model selection and training stage along with the iterative forecasting and correcting stage. In the former stage, the model is in the series-parallel configuration, and its test error is estimated by the cross-validation (CV) method. With the help of ARIMA identification results, CV errors are minimized by the Fibonacci search method to select the best lag structure and the only adjustable parameter. In the latter stage, the model is in the parallel configuration, and the so-called leave-one-day-out resampling method is proposed to iteratively estimate correction parameters for horizons up to 24 h ahead, which holds out each full-day data segment from the sample of observations in turn to faithfully reproduce the entire process of training, iterative forecasting and correcting in the in-sample period. Finally, the out-of-sample corrected forecasts can be successively obtained by using the model selected and trained in the former stage and the correction parameters estimated in the latter stage. Furthermore, effectiveness of this model is verified with four real-world case studies of two wind farms in China.
AB - Wind speed forecasting plays a vital role in dispatch planning and operational security for wind farms, however, its difficulty is commonly accepted. This paper develops a nonlinear autoregressive (exogenous) model for one-day-ahead mean hourly wind speed forecasting, where general regression neural network is employed to model nonlinearities of the system. Specifically, this model is a two-stage method consisting of the model selection and training stage along with the iterative forecasting and correcting stage. In the former stage, the model is in the series-parallel configuration, and its test error is estimated by the cross-validation (CV) method. With the help of ARIMA identification results, CV errors are minimized by the Fibonacci search method to select the best lag structure and the only adjustable parameter. In the latter stage, the model is in the parallel configuration, and the so-called leave-one-day-out resampling method is proposed to iteratively estimate correction parameters for horizons up to 24 h ahead, which holds out each full-day data segment from the sample of observations in turn to faithfully reproduce the entire process of training, iterative forecasting and correcting in the in-sample period. Finally, the out-of-sample corrected forecasts can be successively obtained by using the model selected and trained in the former stage and the correction parameters estimated in the latter stage. Furthermore, effectiveness of this model is verified with four real-world case studies of two wind farms in China.
KW - Cross-validation
KW - Fibonacci search method
KW - Forecast correction
KW - General regression neural network
KW - Leave-one-day-out resampling
KW - Wind speed forecasting
UR - http://www.scopus.com/inward/record.url?scp=84976636850&partnerID=8YFLogxK
U2 - 10.1016/j.apenergy.2016.06.098
DO - 10.1016/j.apenergy.2016.06.098
M3 - Article
AN - SCOPUS:84976636850
SN - 0306-2619
VL - 178
SP - 886
EP - 901
JO - Applied Energy
JF - Applied Energy
ER -