Toward Learning-Based Power Subsystem Simulation for Scalable Digital Twins in Distributed Space Systems

Simulating the full set of satellite subsystems with high fidelity is essential for both mission design and operational planning. However, conventional simulators are often computationally intensive and difficult to scale, particularly in the context of distributed space systems (DSS) such as constellations, swarms, or formation-flying architectures. This work explores the use of machine learning (ML) as a lightweight, data-driven alternative for modeling the power subsystem, specifically, the prediction of solar array power output, battery input power, and battery output power. Using open-access telemetry from the BEESAT-4 CubeSat mission, Multilayer Perceptron (MLP) models were trained to estimate these quantities based on a wide range of onboard sensor data. The models achieved high accuracy, with mean absolute errors below 2% of the respective power ranges. Permutation Feature Importance analysis revealed that subsystem activity indicators, such as charger currents, sun vector orientation, and communication system states, play a critical role in power behavior. These findings demonstrate the feasibility of using ML to approximate subsystem dynamics with minimal computational overhead, and provide insights into sensor prioritization for future Digital Twin (DT) implementations in space systems.