nilm_appliance_usage_classification

Non-Intrusive Load Monitoring (NILM): Time Series Classification for Appliance Detection

-Fasna Sharaf, Adithya Thonse, Tushar Sharma

---

Overview

Non-Intrusive Load Monitoring (NILM) is an advanced application of machine learning that focuses on identifying and monitoring the usage of individual electrical appliances in a system without requiring separate sensors for each device. Instead, NILM relies on analyzing the overall power consumption data collected from a single point, such as a smart meter.

By processing this time series data, NILM can determine which appliances were active at any given time. For example, it can identify whether a washing machine, refrigerator, or air conditioner was running, based solely on the patterns in the power consumption data.

NILM has practical applications in areas like monitoring energy usage in factories to optimize costs, tracking power distribution in electric vehicles, and analyzing appliance usage patterns in households to promote energy efficiency. It plays a vital role in reducing energy consumption, lowering electricity bills, and contributing to sustainability by minimizing carbon footprints.

About Dataset

We used an open-source dataset from Kaggle, which was part of the ESDA Non-Intrusive Load Monitoring (NILM) Competition 2021. The goal of the competition was to predict the appliances that were active based on the provided dataset.

For this example, we used the vw_train.csv file from that dataset. Upon inspection, the file contained 30 columns:

id, activePower, activePowerDelta, reactivePower, apparentPower, current, voltage, phase, hertz, transient1, transient2, transient3, transient4, transient5, transient6, transient7, transient8, transient9, transient10, harmonicDelta1, harmonicDelta2, harmonicDelta3, harmonicDelta4, harmonicDelta5, harmonicDelta6, harmonicDelta7, harmonicDelta8, harmonicDelta9, timestamp, appliances

By analyzing the unique values in the appliances column, we identified a total of 28 classes. Below is the class distribution:

Class Distribution

Class	Count
+fridge	28,575
+fridge+shower	593
+fridge+microwave	1,969
+fridge+washing_machine+washer_dryer	31,183
+fridge+washer_dryer	3,328
+fridge+tumble_dryer+washer_dryer	15,475
+fridge+tumble_dryer+washer_dryer+microwave	570
+fridge+kettle	405
+fridge+washing_machine+washer_dryer+microwave	1,172
+fridge+kettle+microwave	9
+fridge+vacuum	24
+fridge+kettle+washing_machine+washer_dryer	225
nan	23
+washer_dryer	5
+fridge+kettle+washer_dryer	60
+fridge+shower+washer_dryer	34
+fridge+kettle+tumble_dryer+washer_dryer+microwave	2
+fridge+shower+washing_machine+washer_dryer	90
+fridge+shower+kettle	3
+tumble_dryer+washer_dryer	43
+fridge+shower+tumble_dryer+washer_dryer	48
+fridge+washer_dryer+microwave	148
+fridge+vacuum+washing_machine+washer_dryer	7
+fridge+kettle+tumble_dryer+washer_dryer	75
+fridge+vacuum+tumble_dryer+washer_dryer	11
+fridge+vacuum+washer_dryer	1
+shower	1
+kettle	1

The numbers indicate how many times each combination of appliances was active in the dataset.

Refining the Dataset

Due to the limited data available for many classes, we had to eliminate certain classes to ensure sufficient data for training, validation, and testing. After careful consideration, we selected the following 4 classes for our example:

1. +fridge
2. +fridge+tumble_dryer+washer_dryer
3. +fridge+washer_dryer
4. +fridge+washing_machine+washer_dryer

Additionally, we reduced the number of variables from the original 30 to just 5 variables to focus only on those that are easy to measure directly. The selected variables are:

• Active Power
• Voltage
• Current
• Reactive Power
• Phase

From these selected classes, we created the train, validation, and test datasets and structured the data to make it compatible with Tiny ML ModelMaker for classification tasks. You can access the refined dataset as a ZIP file here.

Usage in Tiny ML ModelZoo

Here is the command to use the refined dataset we produced earlier with Tiny ML ModelZoo:

./run_tinyml_modelzoo.sh examples/nilm_appliance_usage_classification/config.yaml

Users can configure the model pipeline using a YAML configuration file (like shown in the command above), where different stages (dataset loading, data processing and feature extraction, training, testing, and compilation) can be enabled or disabled based on requirements.

Configuring the YAML File

The YAML file is the core configuration file used in Tiny ML ModelMaker to define the pipeline for tasks such as dataset loading, model training, testing, and compilation. In this configuration, we have enabled dataset loading and correctly set the input_data_path. For feature extraction, we applied several transformations to process the data effectively: the dataset is split into smaller windows of size frame_size (set to 32) using WINDOWING, followed by FFT_FE to convert the time-domain data into the frequency domain. The data is then normalized using NORMALIZE, and reduced in size with BINNING, which groups adjacent data points into bins. FFT_POS_HALF selects the DC component and specific FFT samples, while ABS computes the absolute value of the signal. The amplitude is converted to a logarithmic scale using LOG_DB, and finally, the features are concatenated into a single vector using CONCAT, with num_frame_concat set to 8. This feature extracted data is then trained using CLS_13k_NPU model and quantization is set to 2 which means we use TI NPU quantization. Here we are compiling the model using TI NPU hardware acceleration with full INT8 quantization pipeline. This is the default preset in our toolchain and hence we don't have to specify it in YAML file. The compilation uses skip_normalize=true and output_int=true. flags, meaning the hardware accelerator expects pre quantized INT8 inputs and produces quantized INT8 outputs. All operations, weights, activations, and intermediate tensors are processed in INT8 format on the NPU hardware.

Results

On running the YAML file, the following accuracies were observed:

I. Float Train

• Best Epoch: 14
• Accuracy: 100%
• F1-Score: 1.000
• AUC ROC Score: 1.000

Confusion Matrix:

Ground Truth	Predicted: +fridge	Predicted: +fridge+tumble_dryer+washer_dryer	Predicted: +fridge+washer_dryer	Predicted: +fridge+washing_machine+washer_dryer
+fridge	83	0	0	0
+fridge+tumble_dryer+washer_dryer	0	41	0	0
+fridge+washer_dryer	0	0	3	0
+fridge+washing_machine+washer_dryer	0	0	0	91

---

II. Quant Train

• Best Epoch: 9
• Accuracy: 100%
• F1-Score: 1.000
• AUC ROC Score: 1.000

Confusion Matrix:

Ground Truth	Predicted: +fridge	Predicted: +fridge+tumble_dryer+washer_dryer	Predicted: +fridge+washer_dryer	Predicted: +fridge+washing_machine+washer_dryer
+fridge	83	0	0	0
+fridge+tumble_dryer+washer_dryer	0	41	0	0
+fridge+washer_dryer	0	0	3	0
+fridge+washing_machine+washer_dryer	0	0	0	91

---

III. Test Data

• Evaluation Accuracy: 95.95%
• AUC ROC Score: 0.992

Confusion Matrix:

Ground Truth	Predicted: +fridge	Predicted: +fridge+tumble_dryer+washer_dryer	Predicted: +fridge+washer_dryer	Predicted: +fridge+washing_machine+washer_dryer
+fridge	83	0	0	0
+fridge+tumble_dryer+washer_dryer	0	39	0	4
+fridge+washer_dryer	0	1	4	0
+fridge+washing_machine+washer_dryer	0	4	0	87

You can find the compiled model at: tinyml-modelmaker/data/projects/{dataset_name}/run/{date-time}/{model_name}/compilation

Running on Device

After successfully running Modelmaker, you will get four main files:

1. Artifacts:

   • mod.a and tvmgen_default.h are generated and stored in:

     data/projects/{dataset_name}/run/{date-time}/{model_name}/compilation/artifacts
     

2. Golden Vectors:

   • user_input_config.h and test_vector.c are stored in:

     data/projects/{dataset_name}/run/{date-time}/{model_name}/training/quantization/golden_vectors
     

These four files will be needed while running on device.

In this example, we will use the following setup:

• Device: LAUNCHXL-F28P55X
• C2000Ware Version: 6.00
• Code Composer Studio (CCS): Version 20.3.0

Steps to run this example on-device can be found by following this guide: Deploying Classification Models from ModelMaker to Device

Upon flashing and running the project we can see the model output matches the golden vectors.

---

Performance Metrics

Here are the key performance metrics for the model running on the device:

Metric	Value
Device Name	F28P55x
AI Model Cycles	165629
Inference Time	1104.19 µs
Results Match	TRUE
Model Size	13187 bytes
Model Flash	11967 bytes
Model SRAM	1220 bytes
Application Size	1300 bytes
Application FLASH	976 bytes
Application SRAM	324 bytes

Flash Usage

SRAM Usage

---

Update history: [24th Nov 2025]: Compatible with v1.3 of Tiny ML Modelmaker