Energy-aware dynamic programming scheduler for TinyML workloads on energy-harvesting CubeSat-IoT platforms: a comprehensive system modelling and performance analysis

Abstract The integration of Tiny Machine Learning (TinyML) algorithms into CubeSatInternet of Things (IoT) platforms presents a transformative opportunity for autonomous space-based sensing and decision-making. However, the energyconstrained nature of CubeSats necessitates sophisticated energy management strategies to sustain computational workloads. This paper presents a comprehensive framework for energy-autonomous TinyML operation on CubeSat-IoT platforms, combining multi-source energy harvesting with provably optimal task scheduling.We develop a detailed system model incorporating photovoltaic (PV) and radio-frequency (RF) energy harvesting, battery storage dynamics, and realistic TinyML workload characteristics. The core contribution is a dynamic programming (DP) scheduler with orbital lookahead that maximizes cumulative science return while enforcing battery safety constraints. The scheduler operates on discretized state-of-charge (SOC) and time domains, with polynomial-time complexity enabling real-time implementation.Through extensive 72-hour Low Earth Orbit (LEO) simulations under both energy-positive and energy-constrained operating regimes, we demonstrate that the DP scheduler achieves superior performance compared to baseline approaches (threshold, greedy, conservative). Under energy-constrained conditions (reduced PV area, elevated base loads), the optimal policy maintains SOC above the safety limit of 20% while maximizing task execution, achieving science return improvements of up to 34% over threshold-based scheduling and 18% over greedy scheduling, while maintaining zero safety violations. Under energy-positive baseline conditions, DP and greedy schedulers achieve near-identical performance (0.3% difference), confirming that lookahead benefits emerge when energy constraints are operationally critical. Analysis of the optimal policy map reveals intelligent task selection patterns that exploit energy-rich periods during sunlit phases while ensuring safe operation during eclipses. Furthermore, we derive an operational design curve relating payload power consumption to sustainable processing time under realistic power margins. This provides practical guidance for electronics designers in sizing power systems for TinyML-enabled CubeSat missions. The framework establishes a foundation for energy-autonomous intelligent nanosatellites capable of persistent onboard inference within green energy constraints.