Abstract
Non-intrusive load monitoring (NILM) promises appliance-level intelligence from aggregate smart-meter data, yet many solutions remain server-bound and are evaluated with metrics that conceal class imbalance. This paper presents Energy Detective, an edge-deployable bidirectional long short-term memory (Bi-LSTM) pipeline for appliance-level power estimation, state recognition, and user-facing explanation. Seven houses from the UK REFIT dataset were combined into 5.3 million one-minute samples. A 128-unit Bi-LSTM was trained with 60-minute windows, batch normalisation, dropout, and mean absolute error (MAE) loss; model shards were then executed locally through TensorFlow.js in a React dashboard. On unseen Houses 3, 9, and 10, the normalised MAEs were 0.0018, 0.0147, and 0.0158, respectively (mean 0.0108), while browser inference remained below 15 ms. Crucially, a fixed 20 W state rule produced 94.08% overall accuracy but zero recall for the active class (macro-F1: 0.4847), revealing majority-class collapse and normalisation-induced magnitude suppression. The resulting contribution is both architectural and evaluative: the work demonstrates a practical cloud-edge/XAI deployment pattern while showing why cross-household NILM requires adaptive calibration, class-aware metrics, and explanations that expose rather than hide failure modes.