A comprehensive review of machine learning and deep learning models for non-intrusive load monitoring - performance, analyses, practical insights, and emerging trends

Aug 5, 2026

Non-intrusive load monitoring (NILM) enables energy disaggregation by estimating the energy consumption and operating time of individual appliances from aggregated meter data. Recent progress comes from integrating machine learning (ML) and deep learning (DL) models that improve estimation accuracy. This review evaluates ML/DL methods for NILM across key research questions: addressed NILM problems, model–dataset performance, interpretability approaches, real-world operational requirements, and remaining dataset/model/implementation challenges, concluding with recommendations and future research directions.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages36

Reading Time1 h 31 min

AccessFree to download

Abstract

Introduction

Energy demand growth and monitoring needs

Intrusive vs. non-intrusive load monitoring

NILM history and core methodology

Early feature-based approaches and limitations

ML/DL-oriented directions mentioned in the review

A comprehensive review of machine learning and deep learning models for non-intrusive load monitoring: performance, analyses, practical insights, and emerging trends Muhammad Hammad Saleem1,2,3 · Muhammad Taha4 · Muhammad AsifAli Rehmani1,5 · Shafiqur Rahman Tito6 · Snjezana Soltic1 · Pieter Nieuwoudt1 · Neel Pandey7 · Mollah Daud Ahmed8 Received: 3 April 2024 / Accepted: 6 September 2025 © The Author(s) 2025 Abstract Non-intrusive load monitoring (NILM) is widely regarded as a key method for energy disaggregation to estimate the energy consumption and operating time of individual equipment and appliances. Over the past few years, NILM has achieved substantially high accuracy through the incorporation of several machine learning (ML) and deep learning (DL) algorithms. This article reviews ML/DL-based approaches that have been applied to address NILM-related problems. This review focuses on five research questions: First, which problems associated with NILM systems have been addressed in recent studies using ML and DL models, and how well have the ML and DL models addressed NILM-related problems? Second, which ML and DL models have been the most explored for NILM, and how do the ML and DL algorithms perform on different NILM datasets? Third, how has model interpretability been addressed in ML/DL-based NILM methods, and what techniques have been proposed to improve the transparency of predictions? Fourth, what are the practical requirements of ML-based NILM systems in real-world settings? Finally, what are the main challenges of ML-based NILM systems in terms of datasets, models, and their implementation? This review also provides recommendations for addressing the identified research gaps and emerging trends and technologies in NILM and presents directions for future research. Keywords Non-intrusive load monitoring (NILM) · Machine learning · Deep learning · Demand side management · Convolutional neural network · Recurrent neural network 􀀍 Muhammad Hammad Saleem [m.h.saleem@salford.ac.uk](m.h.saleem@salford.ac.uk)[ ](m.h.saleem@salford.ac.uk)Muhammad Taha [engr.muhammadtaha95@gmail.com](engr.muhammadtaha95@gmail.com)[ ](engr.muhammadtaha95@gmail.com)Muhammad AsifAli Rehmani[asif.rehmani@outlook.com](asif.rehmani@outlook.com) Shafiqur Rahman Tito [shafiq.tito@waikato.ac.nz](shafiq.tito@waikato.ac.nz) Snjezana Soltic[snjezana.soltic@manukau.ac.nz](snjezana.soltic@manukau.ac.nz)[ ](snjezana.soltic@manukau.ac.nz)Pieter Nieuwoudt [pieter.nieuwoudt@manukau.ac.nz](pieter.nieuwoudt@manukau.ac.nz) Neel Pandey [neel@nzseg.com](neel@nzseg.com) Mollah Daud Ahmed [daud.ahmed@xtra.co.nz](daud.ahmed@xtra.co.nz) 1 School of Professional Engineering, Manukau Institute of Technology, Auckland 2104, New Zealand 2 School of Science, Engineering, and Environment, University of Salford, Salford M5 4WT, UK 3 Data Science and Artificial Intelligence Hub, University of Salford, Salford M5 4WT, UK 4 Department of Electrical Engineering, National University of Computer and Emerging Sciences, Karachi, Pakistan 5 Hamilton City Council, Urban and Spatial Planning Unit, Hamilton, New Zealand 6 Ahuora-Centre for Smart Energy Systems, School of Computing and Mathematical Sciences, University of Waikato, Hamilton, New Zealand 7 New Zealand Skills & Education Group, Auckland, New Zealand 8 Business Dynamics Management Consultants Limited, Auckland, New Zealand 1 3 1 Introduction The global energy demand has significantly increased by more than 45% in the last two decades [1] and is expected to increase by another 50% by 2050 [2] . To meet the energy demand, the electricity industry has taken many steps, including but not limited to moving towards renewable energy [3], promoting energy-efficient devices/equipment [2, 4], implementing energy storage technologies [5–8], and monitoring load consumption. Smart monitoring systems play a key role in reducing electricity consumption and improving awareness of electrici

FAQs

What is non-intrusive load monitoring (NILM) used for?

NILM estimates the energy consumption and operating time of individual appliances using only aggregated smart-meter power data, supporting energy disaggregation for end-user and grid awareness.

Why are ML and DL models important for NILM performance?

The abstract states that integrating ML/DL algorithms has led NILM to substantially high accuracy over recent years, enabling improved handling of NILM-related problems across different datasets.

Which topics does the review cover about ML/DL-based NILM methods?

It examines which NILM problems have been tackled, which ML/DL models are most explored and how they perform on different datasets, how interpretability is addressed, practical requirements in real-world settings, and the main challenges in datasets, models, and implementation.

Aug 5, 2026

Upload by Clementine