High Resolution Forecasts of Photovoltaic Energy Production
Using Analog Ensemble to forecast photovoltaic energy production on a household level
Key Points
This poster studies the temporal predictability of the small-scale photovoltaic energy production. The location studies is at State College, PA. The study compares the prediction performance between the Artificial Neural Network (ANN) and the Analog Ensemble (AnEn) technique in forecasting short-term (3-day forecasts) photovoltaic energy production.
- The optimal ANN is trained using a brute-force search with 1-20 hidden nodes, and the best one with 10 nodes is chosen. At this specific location, AnEn outperforms ANN in respect to daily averaged errors. ANN tends to under-predict the peak photovoltaic (PV) energy production while AnEn shows good results during peak PV energy time period.
- AnEn can be used to downscale hourly PV forecasts to every 5-minute which, in this case, is the resolution of PV observations. Results show that the downscaling technique with AnEn is fairly accurate when the original forecasts are accurate.
- Both (ANN and AnEn) model are not able to perform well when the underlying meteorological model is inaccurate.