Probabilistic Ramp Rate Forecasts of Aggregate Power of PV Fleets

Forecasts of distributed energy resources play a vital role in reserve planning for electrical grids. We combined multiple feedforward neural networks into a mixture-of-experts model to create a probabilistic forecast of aggregate AC power of a modeled fleet of 100 PV systems. The power forecasts were sampled and combined to create two ramp rate forecasts: a forecast where samples at the start and end of the ramp period were matched by sample rank (i.e. samples were first sorted), and a forecast where the samples were matched at random. The two ramp rate forecasts were evaluated for 15, 30, and 60-minute ramp rates against aggregate power of the fleet using the continuous ranked probability score. The forecast model was trained and tested using Group Shuffle Split cross-validation 63 times. Use of unranked matching significantly reduced error scores for 30-minute and 60-minute ramp rate forecasts (by 6% and 16%), and ranked matching significantly reduced error scores for 15-minute ramp rate forecasts (by 14%). We conclude that temporal correlation between forecast periods is modeled more effectively with ranked matching at the shortest ramp rate, but temporal correlation is overestimated by ranked matching for longer ramp rate periods.