Reactive and Proactive Migration for Edge AI Training on Energy-Harvesting, Intermittently Powered Devices

In energy-constrained environments where edge devices operate on intermittent renewable sources, maintaining continuous AI model training poses significant challenges. This work evaluates two migration strategies, reactive and proactive, for preserving computational progress during power interruptions. The reactive strategy relies on centralized heartbeat monitoring to trigger checkpoint transfers, while the proactive strategy enables devices to offload state or data based on local energy thresholds. To improve reliability and efficiency, load balancing is applied to distribute computation across available devices, and battery buffering is integrated to stabilize power supply. Extensive simulation results based on real-world data demonstrate that load balancing increases the migration success rate to 96% in the reactive setting and reduces migration frequency by 39-45% in the proactive scenario. Battery integration yields the most substantial improvement, reducing total proactive migrations by approximately 96% and limiting reactive migrations to just two. These findings highlight the effectiveness of adaptive migration and power-aware strategies in enabling robust, uninterrupted Edge AI training under intermittent energy availability.