The techniques of Non-Intrusive Load Monitoring (NILM) aim to label or disaggregate loads into time series of electrical energy consumption, mainly in the context of residential consumers, considering the optimization of electricity usage in demand response programs. In this sense, many researches have been developed with different methodologies of consumption time series transformation into images, which consider different window sizes. Thus, this work compares and analyzes different transformation techniques for window sizes ranging from 1 to 10 minutes, verifying how the Convolutional Neural Network models behave for different loads from the UK-DALE (The United Kingdom - Domestic Appliance-Level Electricity dataset). Through the proposed methodology, it was possible to obtain f1-scores higher than 90% for three of the five considered loads, demonstrating the robustness of the time-series transformation into images.
Corrêa et al. (Tue,) studied this question.