Multi-step forecasting of wave power using a nonlinear recurrent neural network

Short term forecasting is a vital interest to future implementations of a smart grid, particularly in the reliable integration of renewable energy resources. In this study we focus on multi-step prediction of high resolution wave power. Significant wave height data was first obtained from Belmullet Berth, Ireland and underwent several data preprocessing steps. These include a linear interpolation to fill irregular or missing data points, conversion to power using an interpolated power matrix of a Pelamis Device energy converter, and then exponential smoothing is applied. We utilized a nonlinear autoregressive recurrent neural network for 3, 6, 12 and 24 hour prediction. Our method showed highly accurate results when data has been smoothed, versus raw data, and when compared to previous studies.