Mohammed Ezzaldin Babiker Abdullah

Domain expertise still trumps brute force: a physics-guided neural network beats self-attention at solar irradiance forecasting, despite being far less complex.

Penalizing over-optimistic Remaining Useful Life (RUL) predictions with an asymmetric loss function yields safer and more reliable industrial prognostics.

By structurally enforcing thermodynamic constraints, this model eliminates the physically impossible nocturnal power generation that plagues even state-of-the-art deep learning approaches to solar forecasting.