Smart Meter Appliance Detection

ISEF Category: Embedded Systems

Subcategory: Internet of Things  ·  Difficulty: Intermediate  ·  Setup: Home Setup  ·  Time: 1 to 2 Months

The Hook

Your home already leaks clues about what is running. A tiny current sensor can read those clues from the wall like a detective reading footprints in mud. That means your project can tell a lamp from a laptop without opening a single device. It can also spot hidden standby power waste, the kind that keeps draining energy while you think everything is off.

What Is It?

Non-intrusive load monitoring, or NILM, is a way to guess which appliances are running by watching the total power coming into a circuit. Think of it like hearing a whole band and trying to pick out the guitar, drums, and vocals from one recording. Each appliance has its own power pattern, so the detector looks for changes in current, voltage, and power behavior.

A CT clamp, short for current transformer clamp, measures current without cutting the wire. An ESP32 is a small microcontroller with Wi-Fi that can collect sensor data and send it to a laptop or server. Your project turns those measurements into labels, like kettle, fan, or TV, and then looks for odd steady draws that might mean a device is wasting power in standby mode.

The hard part is not just sensing. You also need a way to turn messy electrical signals into useful categories. That means choosing features, building a simple model, and testing how well it handles different appliances, different houses, or noisy background loads.

Why This Is a Good Topic

This is a strong science fair topic because you can measure real signals, compare clear categories, and test a practical problem. It connects to home energy use, smart homes, and power waste. You can make it your own by choosing the classification method, the appliance set, or the way you detect standby power vampires. You can also collect enough data at home to learn real data cleaning, model testing, and error analysis.

Research Questions

• How does the choice of appliance affect NILM classification accuracy?
• What is the effect of sampling rate on appliance detection performance?
• Does adding voltage features improve classification compared with current-only features?
• To what extent can a simple model detect standby power vampires from daily load traces?
• Which appliance pairs are most often confused by a CT-clamp based classifier?
• How does background household activity change the false positive rate of standby-power detection?
• What is the effect of using time-domain features versus frequency-domain features on model accuracy?

Basic Materials

• ESP32 development board.
• Split-core CT clamp sensor with burden resistor or compatible sensor module.
• AC power adapter or safe sensor interface for the measurement setup.
• Breadboard and jumper wires.
• Laptop with USB cable.
• Household appliances with distinct on-off cycles, such as lamp, fan, and charger.
• Digital multimeter.
• Notebook or spreadsheet for logging labels and results.

Advanced Materials

• ESP32 development board with external ADC if needed.
• CT clamp current sensor and voltage sensing module for power factor estimates.
• Oscilloscope or power quality analyzer for validation.
• Kill A Watt meter or equivalent plug-in power meter.
• Safe isolated test circuit or instrumented outlet box.
• Data logging storage, such as microSD module or server connection.
• Reference loads with known behavior for calibration.
• Laboratory-grade multimeter and clamp meter for cross-checks.

Software & Tools

• Arduino IDE: Programs the ESP32 and handles serial data collection.
• Python: Cleans sensor data, builds features, and trains classification models.
• pandas: Organizes time-series data and labels from each appliance run.
• scikit-learn: Trains and compares simple classifiers for NILM tasks.
• ImageJ: Can help if you graph current traces as images for pattern checks, though it is optional.

Experiment Steps

1. Define the exact sensing problem, such as appliance classification, standby-power detection, or both.
2. Choose one electrical measurement path and decide how you will label each appliance state.
3. Plan a data set with enough repeated runs, background loads, and device combinations to test confusion cases.
4. Select a small group of features or signal summaries that your model will use.
5. Decide on a baseline method first, then compare it against a stronger classifier or a different feature set.
6. Set up validation rules that separate training data from test data so your accuracy estimate stays honest.

Common Pitfalls

• Recording data from one appliance at one time of day, which makes the model memorize the setting instead of the device.
• Mixing up standby draw with low-power active states, which causes false vampire alarms.
• Skipping calibration against a known meter, which leaves the current readings too uncertain to trust.
• Using appliances with nearly identical power patterns, which makes the classifier look worse than it should.
• Training and testing on overlapping runs, which inflates accuracy and hides real errors.

What Makes This Competitive

A stronger project will do more than label a few appliances. You can compare simple features against better ones, test whether the model generalizes across rooms or days, and report confusion patterns instead of one single accuracy number. You can also separate true standby waste from normal low-power behavior, which shows deeper thinking about the problem. A careful error analysis and a clean validation plan can move this far beyond a demo.

Project Variations

• Test whether the classifier still works when appliances are plugged into different outlets on the same circuit.
• Compare current-only NILM with a model that also uses voltage or power factor features.
• Focus only on standby-power detection and rank devices by how much wasted energy they draw overnight.

Learn More

• MIT OpenCourseWare, Introduction to EECS and signals material: Search MIT OpenCourseWare for courses on signals, systems, and embedded systems to learn the basics behind sensor processing.
• USGS Energy and Environment resources: Search the U.S. Geological Survey site for electricity and energy-use background that can help frame home power research.
• NOAA Education resource on data analysis: Search NOAA Education for lessons on working with time-series data and graphs.
• PubMed: Search review articles on non-intrusive load monitoring, appliance classification, and energy disaggregation.
• IEEE Xplore: Search for NILM review papers and case studies, often accessible through school or public library subscriptions.