Building Energy Analytics with Appliance Disaggregation

The problem with a single electricity bill

A building's electricity bill is one number. To reduce it, you need to know where the energy went: the air conditioner, the water pump, the fridge, lights left on overnight, computers never shut down. Installing a meter on every circuit is expensive and rarely done. Non-intrusive load monitoring (NILM) infers the breakdown from the single meter the building already has.

This matters more as electricity prices rise. In Pakistan, as in many countries, bills have become a major household and business expense, and smart meters are gradually replacing old ones, which makes the data needed for disaggregation more available. Facility managers of offices, hostels and campuses face the same question at a larger scale.

What already exists and the gap

• Sense sells a home energy monitor installed in the electrical panel that learns device signatures from high-frequency data.
• Emporia Vue measures individual circuits with clamps, which is accurate but needs a sensor per circuit.
• Bidgely provides disaggregation analytics to utilities using smart meter data.
• NILMTK is an open source toolkit with dataset converters, baseline algorithms and metrics, supporting public datasets such as REDD, UK-DALE and REFIT.

Consumer devices are sold mainly in North America and Europe and need extra hardware; utility analytics are sold to utilities, not to a campus facilities office. The gap is an open, explainable pipeline that turns existing meter data into a small number of actionable findings with money attached.

What you build

• Ingestion: load a public dataset's aggregate and sub-metered readings, resample, handle gaps.
• Event detection: find step changes in the aggregate signal that mark appliances switching on or off.
• Disaggregation model: a neural network (for example a sequence-to-point convolutional model) per appliance category, compared with NILMTK's classical baselines.
• Waste detection: estimate the always-on baseline (the lowest sustained load, typically overnight) and equipment running outside occupied hours.
• Dashboard: breakdown by category, estimated cost at a configurable tariff, and a ranked list of waste findings with a yearly cost.

The technically interesting part

Disaggregation is genuinely hard. Many appliances have similar signatures, several run at once, and variable loads such as anything with a motor, compressor or inverter do not produce clean steps. Accuracy varies enormously by appliance, so reporting per appliance rather than one average is the honest presentation.

The second hard part is transfer. A model trained on one house usually degrades badly on another, because appliances differ. Train on some buildings, test on a held-out building, and report the drop honestly. That result is more valuable to your supervisor than a strong number on a single building.

How to evaluate it

• Per-appliance metrics: mean absolute error and F1 for on/off state, against sub-metered ground truth.
• Comparison: your model against at least two NILMTK baselines on the same split.
• Transfer: metrics on a held-out building, with the drop per appliance.
• Waste finding accuracy: estimated always-on load against the sub-metered sum of standby appliances.
• Usefulness: 3 people who pay an electricity bill or manage a building read the dashboard and say which finding they would act on.

Suggested plan

1. Weeks 1 to 3: datasets, NILMTK setup, cleaning and baselines.
2. Weeks 4 to 7: event detection and neural disaggregation models.
3. Weeks 8 to 10: transfer experiments and per-appliance evaluation.
4. Weeks 11 to 13: waste detection and dashboard.
5. Weeks 14 to 15: user feedback and report.

Path to a product

The realistic first customer is a campus facilities office, a hostel, or an office building manager who already has a main meter with interval data. Offer an energy report service: monthly findings with money attached, charged per building. The always-on load finding is often the single most actionable result, and a clear number on it sells the service. Utilities and solar installers are later partners, since both want to understand customer load.

Scope warning

Public datasets are the right starting point. Instrumenting a real building is a stretch goal, not the plan. Do not build hardware, a billing system or real-time control of appliances.

Learn what you need on Sefism

• Machine learning roadmap : supervised learning and evaluation before deep models.
• Machine learning cheat sheet : metrics and validation strategies.
• AI engineering roadmap : serving models behind a dashboard.
• Google Colab : free GPU time for training sequence models.
• PostgreSQL Documentation : storing and querying time series readings.

Get help with this idea

This is a demanding project, and the honest per-appliance and transfer results are what make it strong. Tauseef can help you set a realistic appliance list, plan experiments and present the evaluation. If you are new here, create a free account first.

• Discuss this idea with a mentor
• Send a query to Tauseef
• Book a 1:1 session