Abstract
The rapid integration of renewable energy sources is increasing the volatility and non-stationarity of modern power systems, posing significant challenges for data-driven forecasting models. In particular, concept drift and uncertainty in exogenous inputs such as weather forecasts can severely degrade predictive performance over time. This work proposes a lightweight two-layer forecasting framework that decouples prediction from adaptation. A traditional offline regression model is augmented by an online meta-learner that continuously generates adaptive meta-features, enabling the system to respond to structural changes and noisy inputs without repeated retraining. The framework is evaluated on two real-world case studies. First, concept drift is addressed in nuclear power production forecasting, where abrupt and gradual capacity changes are inferred through an online meta-learner. Second, feature uncertainty is mitigated in day-ahead solar production forecasting by correcting noisy weather forecast inputs. Across both scenarios, the proposed approach consistently outperforms single-layer baselines, reducing root mean squared error by up to 10% and maintaining robust performance over multi-year horizons without retraining. The results demonstrate that meta-learning provides a practical and computationally efficient mechanism for improving forecast robustness in non-stationary energy systems, with applicability to a wide range of power-system forecasting problems.