ML-Based Smart Appliance Energy Monitoring System

Abstract

Modern homes only get a combined electricity consumption value. In other words, they have no information about which of their appliances have used the electricity a house has consumed. This paper introduces the design of an ‘ML-Based Smart Appliance Energy Monitoring’ system for recognizing the electricity consumption of a smart home. The Non-Intrusive Load Monitoring system uses an ESP32 development board and a PZEM-004T V3.0 energy meter and current transformer (CT) to measure the voltage, current, power, frequency, and power-factor data of a particular circuit of the home’s electrical circuit. These are then transmitted over Wi-Fi/HTTP to a Flask-based inference server where a 50-feature input (constructed in identical fashion between training and inference from a 13,439-row labeled dataset) is classified by a joint ensemble (HistGradientBoosting (45%), ExtraTrees (35%), and Random Forest (20%)) of regressors into one of eight appliance-combination classes representing three target devices (a phone charger, a laptop charger, and a hot glue gun). An additional individual-device model enables ON-time tracking, per-device energy estimation, and electricity cost calculation for the devices. Classification results are then locally visualized on a 16 × 2 I2C LCD and remotely on the Blynk IoT platform while storing relevant device statistics in the ESP32 Preferences. In contrast to the recent deep-learning-based approaches, which typically require heavy computation and a large amount of annotated data to train on, we propose an explainable ensemble pipeline for low-end single-household appliances.

Introduction

The text presents a machine-learning-based Non-Intrusive Load Monitoring (NILM) system designed to monitor household electricity consumption and estimate the energy usage of individual appliances using a single measurement point. Traditional electricity meters provide only total household consumption, making it difficult to identify which appliances consume the most energy. NILM addresses this limitation by analyzing the overall electrical measurements without requiring separate sensors for each appliance.

The proposed system uses an ESP32 microcontroller, a PZEM-004T energy monitoring module, a current transformer (CT), and an LCD display to measure voltage, current, active power, power factor, and frequency from the household's main supply line. The collected data is transmitted through Wi-Fi to a backend server, where it is processed using a feature-extraction pipeline that generates 50 electrical features to improve appliance identification.

For appliance-combination classification, the system employs a weighted ensemble of three traditional machine-learning algorithms: Histogram Gradient Boosting (45%), ExtraTrees (35%), and Random Forest (20%). The ensemble identifies eight appliance-combination classes based on the collected household data. A separate individual-appliance estimation model is also used to estimate the power consumption of individual devices.

Previous NILM research has explored deep-learning architectures, including CNN-BiLSTM-Transformer models, attention mechanisms, and transfer learning. Although these methods can achieve high accuracy, they often require substantial computational resources and large training datasets. In contrast, the proposed approach emphasizes computational efficiency, interpretability, and practical implementation on resource-constrained devices.

The system also provides local and remote visualization interfaces, allowing users to monitor electrical parameters and appliance-level consumption information. Overall, the research aims to develop an affordable, practical, and scalable smart energy monitoring solution that supports household energy management and demand-side optimization.

Conclusion

An innovative ML-Based Smart Appliance Energy Monitoring System is proposed that is non-intrusive in terms of using the power aggregate measurements at one point to extract the operating conditions of the appliances. The ESP32 acquisition module, PZEM-004T sensing module, Wi-Fi communication, engineered electrical features and machine learning models are utilized for the appliance-level monitoring of the connected loads without using any dedicated sensor for the appliances. In total, 50 engineered electrical features were considered for the eight-class appliance-combination classification problem. The highest accuracy was achieved by the HistGradient boosting classifier among the individual classifiers at 99.58%. On the contrary, the proposed weighted ensemble model of HistGradient boosting, Extratrees and Random forest achieved 99.55% accuracy with 99.41% macro precision, 99.48% macro recall and 99.44% macro F1-score for all classes. On an individual basis, the accuracies achieved were as high as 99.40%, 99.99%, and 99.97% for the phone charger, laptop charger and hot glue gun respectively. The achieved results confirm that engineered electrical characteristics, such as the features related to power-current relationship, power consistency, power ratios, and reactive-power related features, carry useful information for the differentiation of the operating conditions of different appliances. The proposed system also provides an estimate of the appliances’ power, energy consumption, electricity cost, as well as real-time monitoring and interval-based reporting via the implemented dashboard and reporting interface. However, it should be noted that the obtained individual appliance power values at simultaneous operation of multiple appliances are based on the system’s mathematical model using the single-point aggregate electrical consumption. This explains the relatively higher complexity of simultaneous appliance combinations compared with a single appliance, reflected in the achieved lower recall for the multiple appliance operating classes. Future work would consider the increase in the number of appliances under monitoring as well as the consideration of higher-resolution current or voltage waveforms to exploit possible transient content in order to further improve the performance. Alternatively, deep-learning-based approaches could be considered if sufficient labeled data becomes available for training such models. The proposed system could also be utilized for intelligent demand-side energy management and appliance scheduling solutions to enable smarter electricity consumption patterns.In summary, the developed solution demonstrated a certain degree of promise in its capacity to allow for smart monitoring of appliance energy by means of single point electrical sensing plus lightweight ensemble machine learning.

References

[1] \"Machine learning models for non-intrusive load monitoring: A systematic review and meta-analysis,\" MDPI, Mar. 2026. [2] \"Lightweight and scalable transfer learning framework for load disaggregation,\" arXiv:2603.04998, 2026. [3] X. Kong, Z. Gui, M. Wu, C. Miao, and Z. Luo, \"A hybrid deep learning framework for non-intrusive load monitoring,\" Electronics, vol. 15, no. 2, p. 453, Jan. 2026. [4] \"Performance enhancement of non-intrusive load monitoring based on adaptive multi-scale attention integration module,\" Electronics, vol. 15, no. 3, Jan. 2026. [5] \"NILMFormer: Non-intrusive load monitoring that accounts for non-stationarity,\" arXiv:2506.05880, 2025. [6] J. Kelly and W. Knottenbelt, \"The UK-DALE dataset, domestic appliance-level electricity demand and whole-house demand from five UK homes,\" Scientific Data, 2015. [7] G. W. Hart, \"Nonintrusive appliance load monitoring,\" Proc. IEEE, vol. 80, no. 12, pp. 1870–1891, 1992. [8] NILMTK: An open-source toolkit for Non-Intrusive Load Monitoring research.

Copyright

Copyright © 2026 Divya Dongre, Astha Bavistale, Atharva Wasnik, Rohan Hedau, Sanchit Nardelwar, Krishnakumar Bhabhda, Ms. Priya Meshram. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.