Kalman Filter-based Adaptive Forecasting of PV Power Output

In building a forecasting model, weather conditions are typical inputs. However, PV faults happen unexpectedly and affect PV power output significantly. Considering that reliable forecasting is still required in such cases, a forecasting model needs to be adaptive to such PV faults until PV faults are fixed. In pursuing that adaptive forecasting, we utilize Kalman filter together with forecast history. Observing recent forecast history relevant to the current forecast, we try to adjust forecast output by reflecting the observations through Kalman filter. Through experiments using the real-world data, we show that our approach quickly realizes adaptive forecasting just when relevant forecast history is available while the effectiveness of the re-training approach (i.e., one representative existing approach) is limited by available data about PV faults. Our approach (the re-training approach) increases MAPE from 15.46% to 20.27% (33.3%) as the amount of drop of PV power output increases from 0% to 30%.