Stability-Controlled Continual Federated Learning for Energy-Harvesting AIoT Systems

Energy-harvesting (EH) AIoT systems enable long-term autonomous operation but suffer from time-varying energy availability, which makes stable learning difficult. In such environments, federated learning (FL) is prone to energy depletion (blackout), while continual learning is required to handle evolving data distributions, leading to a trade-off between energy stability and catastrophic forgetting. In this paper, we propose a stability-controlled continual federated learning framework that jointly regulates local training intensity and rehearsal usage based on the residual energy state. The proposed method is derived from a Lyapunov drift-plus-penalty formulation and implemented as a lightweight mode-based control policy. Simulation results using real solar energy traces show that the proposed method significantly reduces blackout while improving accuracy and mitigating forgetting compared to existing approaches. These results demonstrate the effectiveness of energy-aware joint control for stable continual federated learning in EH-AIoT systems.