Incremental Learning for Fast Shading Adaption in Solar Power Forecasting

This study presents a novel hybrid incremental
learning approach for short-term solar power prediction under
dynamic shading conditions, using time window-based incre-
mental learning to adapt a pre-trained prediction model. The
proposed approach rapidly adapts to changes in shading patterns
without the need for full model retraining. The model is config-
ured to use an estimation period, after which it requires three
additional days to effectively compensate for shading effects.
The results show that the proposed method significantly reduces
forecast errors, improving the RMSE by up to 19.27% and the
MAE by up to 31.81% in strong shading scenarios compared
to transfer learning approaches. Moreover, the method requires
only one tenth of the computation time. The proposed method
provides a scalable, robust, and efficient solution for solar power
forecasting, particularly in scenarios with frequent and strong
shading.