Abstract
Household energy interventions are commonly delivered as uniform messages even though consumption routines, responsiveness and motivation vary markedly between users. This paper presents GREENMIND, an explainable machine-learning framework that converts fine-grained electricity traces into behavioural profiles that can support personalised carbon-aware nudging. Using 2,075,259 one-minute measurements from the UCI Household Electric Power Consumption dataset, the pipeline resamples readings, constructs 30-day windows and derives seven interpretable temporal features capturing habit strength, variability, evening demand, entropy, peak concentration and appliance diversity. A weighted two-layer long short-term memory network and a Random Forest are evaluated for classifying windows as Habitual, Aspiring or Reactive. The Random Forest achieves 84.4% accuracy and 0.492 macro F1, outperforming the weighted LSTM by 21.9 percentage points in accuracy and 0.152 in macro F1. SHAP analysis identifies appliance diversity and temporal variability as the dominant behavioural signals. The results establish a transparent profiling layer for adaptive nudge selection, while also exposing the minority-class and single-household limitations that must be resolved before causal claims about carbon reduction are made. The contribution is therefore not an untested promise of behaviour change, but a reproducible and interpretable foundation for profile-aware environmental engagement systems.