On temporal resolution selection in time series wind power forecasting

As a key and popular renewable energy, wind power penetration has increased significantly into the power system worldwide in recent years. In order to solve the stochastic nature of wind power and develop better dispatch plan, wind power forecasting is imperative before integrating it into the system. To achieve better forecasting accuracy, it is necessary to investigate the significance of temporal resolution for model based multi-step forecasting. In this paper, three models, namely Gaussian process (GP), temporally local Gaussian process (TLGP) and persistence are employed for short term wind power forecasting under two temporal forecasting resolutions, e.g. hourly and quarter-hourly. The methods are applied to the data collected from a wind farm located in Texas, USA. A key finding is that the hourly forecasting is more accurate and computationally efficient than the quarter-hour resolution. Another finding is that, to achieve the best forecasting performance regardless of the resolutions, the models need to use data of the same length of time period. However, a combination consideration should be taken regarding the minimum horizon required.