NILM: One Smart Meter, Full Appliance-Level Insight
Installing smart plugs on every outlet is expensive and impractical. Non-Intrusive Load Monitoring (NILM) reads a single main smart meter and algorithmically disaggregates the aggregate power signal to identify individual appliance consumption — a technique pioneered at MIT in 1992 with modern open-source toolkits now making it practical for homes.
NILMTK: The Standard Toolkit
NILMTK is the primary Python library for NILM research and deployment. The pipeline: collect aggregate power at 1–6 second intervals → detect step-change events (on/off transients) → extract features (ΔP, ΔQ, rise time, steady-state level) → match against appliance signature database → output per-appliance timeseries.
Training Data: REDD and UK-DALE
REDD (6 US households with sub-metered ground truth) and UK-DALE (5 UK households at 1/16,000 Hz) are the standard benchmarks. NILMTK provides built-in loaders for both, enabling algorithm comparison before deploying on your own home data.
Recommended Algorithms
Factorial Hidden Markov Model (FHMM): Each appliance modeled as a 2-state HMM; observed aggregate power = sum of all HMM states. Computationally tractable for ≤15 appliances.
Seq2Point Neural Network: A 1D CNN operating on 599-sample sliding windows of aggregate power predicts the center-point state of a target appliance. State-of-the-art F1-scores: washing machine 0.94, rice cooker 0.91, refrigerator 0.89.
Hardware for Data Acquisition
- Shelly EM (2,500 THB): 2-channel, 120A CT clamps, 1 Hz MQTT — recommended for NILM - CircuitSetup energy monitor (2,800 THB): ESPHome-compatible, 16-circuit Data flows Shelly EM → MQTT → InfluxDB → Python NILM inference script, with results feeding Home Assistant energy dashboards.
Thailand Use Cases
NILM enables detecting failing appliances before failure (refrigerator compressor running unusually long), monitoring children’s A/C usage without confrontation, and optimizing appliance scheduling to align with MEA off-peak TOU rates. Cost of one Shelly EM (2,500 THB) versus 16 individual smart plugs (~12,000 THB) makes the ROI compelling.