Interval prediction of solar power using an Improved Bootstrap method

The integration of solar energies into power grid requires accurate prediction of solar power. While most previous literature is focused on how to improve the accuracy of point forecast, we consider constructing the prediction interval (PI) for solar power which is more appropriate for its nature of high variability. Traditional theoretical approaches of constructing PIs always require the assumption that forecast errors are normally distributed with zero mean. However, this assumption can be easily invalid for solar power data. In this work, an Improved Bootstrap method is proposed to improve the traditional theoretical approaches. It is especially designed to provide PIs for solar power and the problem of invalid assumption about forecast errors can be addressed. The proposed methodparison with three different types of novel PI methods. With interval width and coverage probability as evaluation measures, our method achieves a more than three times lower interval width than other methods while the coverage probability can be still guaranteed. Two-year photovoltaic data of the University of Queensland is used to validate the methodology and different prediction time frames of 5 min, 30 min, 1 h, 2 h and 6 h are applied.