Abstract: In the recent past, the integration of wind energy generation into smart grids has gained lot of momentum because of its availability. The major hurdle in the integration of wind power in smart electric grids, at present time is the irregularity and unpredictability of wind power. Therefore, in order to deal with these challenges, the superior forecasting tool plays an important role in the planning and execution of the wind energy integration. In the expanding power system, because of increasing wind power penetration, a precise wind power forecasting technique is greatly needed to help system operators and consider wind power production in economic scheduling, unit commitment, and allocation trouble reservation. In this paper, two hybrid recurrent dynamic neural networks have employed hybridizing wavelet transform (WT) for short-term prediction of wind power. The proposed approach consists of wavelet decomposition of wind power and wind speed time series, and NAR and NARX recurrent dynamic neural networks are employed to regress upon each decomposed sub-series. Thereafter, the individual outputs of sub-series are aggregated to achieve final prediction of wind power, with up to 24 hours of forecast horizon. The performance of the proposed method is obtained in terms of MAE, MSE, and MAPE values and compared to the results of the persistence method. The forecast results reveal that WT-NARX model is better in terms of the selected performance criteria as compared to the WT-NAR and persistence models respectively.

Key words: wavelet transform, Smart Grids, Recurrent Neural Network, Dynamic Neural Network, Nonlinear Autoregressive Network (NAR), Nonlinear Autoregressive Network with exogenous input network (NARX).