IMG-NILM: A Deep learning NILM approach using energy heatmaps

TL;DR

This paper introduces IMG-NILM, a novel deep learning approach that transforms electricity time series data into heatmaps for CNN-based appliance disaggregation, achieving high accuracy across datasets.

Contribution

The paper presents a new image-based representation of electricity data for NILM, leveraging CNNs to improve robustness and accuracy in appliance detection.

Findings

Achieves up to 93% accuracy on UK-Dale dataset within a single house.

Maintains an average of 85% accuracy across different houses.

Demonstrates robustness with various appliance types and states.

Abstract

Energy disaggregation estimates appliance-by-appliance electricity consumption from a single meter that measures the whole home's electricity demand. Compared with intrusive load monitoring, NILM (Non-intrusive load monitoring) is low cost, easy to deploy, and flexible. In this paper, we propose a new method, coined IMG-NILM, that utilises convolutional neural networks (CNN) to disaggregate electricity data represented as images. Instead of the traditional approach of dealing with electricity data as time series, IMG-NILM transforms time series into heatmaps with higher electricity readings portrayed as 'hotter' colours. The image representation is then used in CNN to detect the signature of an appliance from aggregated data. IMG-NILM is robust and flexible with consistent performance on various types of appliances; including single and multiple states. It attains a test accuracy of up to 93% on the UK-Dale dataset within a single house, where a substantial number of appliances are present. In more challenging settings where electricity data is collected from different houses, IMG-NILM attains also a very good average accuracy of 85%.