Towards Continuous Power Forecasting: Practical Continual Learning for Real-World Energy Systems in Nonstationary Time Series

As energy systems evolve due to weather variability, infrastructure upgrades, and changing consumption behaviors, power forecasting models often suffer from performance degradation. Traditional offline retraining is frequently impractical due to strict operational constraints, such as limited data access, privacy requirements, and the need for uninterrupted service. This paper asks whether continual learning (CL) can be pragmatically integrated into industrial power forecasting pipelines to allow models to adapt to nonstationary data streams while maintaining long-term stability.

The authors propose the paradigm of Continuous Power Forecasting (CPF), which treats forecasting as a lifelong adaptive process. They utilize the CLeaR (Continual Learning for Regression) framework, a modular system that combines novelty detection with various CL update strategies. The study systematically evaluates six representative CL approaches categorized into three types: regularization-based (e.g., Elastic Weight Consolidation), replay-based (e.g., Random, Recent, and Decay-based Replay), and pseudo-replay (e.g., Generative Replay). These methods are tested on a real-world regional power grid dataset containing 95 diverse entities, focusing on how they handle distributional drift under bounded memory constraints.

The empirical evaluation demonstrates that continual learning effectively enables forecasting models to self-adapt to distributional drift and accumulate knowledge over time. By using novelty detection to trigger updates only when necessary, the models successfully mitigate catastrophic forgetting without the need for large-scale historical data storage. The study highlights that different CL strategies offer distinct stability-plasticity trade-offs: while replay-based methods are robust for short-term adaptation, regularization and generative approaches provide scalable alternatives when explicit historical data storage is restricted. These findings suggest that CPF is a viable, sustainable solution for long-term deployment in dynamic energy environments.