Normalized Residue Analysis for Deep Learning Based Probabilistic Forecasting of Photovoltaic Generations

Abstract

In this study, probabilistic forecasting schemes of day-ahead photovoltaic (PV) generations are investigated with the auto-regressive recurrent neural network model named DeepAR, and are evaluated based on the normalized residues. For PV generations, probabilistic outcomes should be helpful for efficient grid managements to account uncertainties such as sudden changes in the local weather. The tightness of the prediction interval for local PV generations is investigated with DeepAR models with varying input data like the local weather forecasts of the day and historical records of the PV generations. For performance measure, normalized residue with the mean and standard deviation of the predicted traces is compared to the standard normal distribution. For evaluation, local PV generation data captured at Hadong, Korea is tested by the DeepAR models with optional input of local weather forecasts data. The evaluation results of the PV generation tests show that the local weather data provides extra tightness of the prediction interval with the normalized residues close to the standard normal distribution.