Context-Augmented NLP for Predicting Nanosatellite EPS Output Activation from Telecommand Scripts

Abstract

Nanosatellite operations increasingly rely on telecommand scripts to automate payload activities, communication sessions, and subsystem mode transitions. Although scripting improves operational efficiency, it complicates power planning because a single script (or short command sequence) may activate multiple regulated Electrical Power System (EPS) outputs and drive concurrent current draw. This paper proposes an NLP-based multi-label prediction framework that infers which EPS outputs become active directly from telecommand script text. The dataset is built from two 3U nanosatellites operated by the University Center for Research in Space Technology at Mohammed V University (Rabat) and combines (i) ground-segment task logs containing both scheduled scripts uploaded for execution outside contact windows and in-pass scripts executed during ground-station visibility, and (ii) EPS telemetry streams reporting timestamped per-output currents. Training instances are created by temporally aligning script execution to telemetry intervals; labels are generated using an idle-aware activation rule that classifies an output as active when its post-interval current exceeds an empirically characterized idle level plus a noise margin. We compare a TF-IDF + MLP baseline against a transformer encoder adapted to command-like domain-specific language. On the held-out test set, the TF-IDF baseline achieves 0.65 micro-F1, and the transformer model achieves 0.91 micro-F1, showing that command semantics captured by NLP can provide operationally meaningful power awareness for planning and verification.