What is it about?
During years of research in event-based NILM, we noticed that many publications lack crucial details about their implementations. The reasons for this can be manifold, e.g., page limitations at the respective venue. However, these missing details make reproducing their work almost impossible. Hence, our work details important implementation steps, pitfalls to consider and workarounds. Further, we provide a couple of empirical studies on parameters.
Featured Image
Photo by Spencer Davis on Unsplash
Why is it important?
Since the general field of NILM and the event-based approach within it has not agreed on one common baseline set of, e.g., ground-truth to compare their algorithms against, it is not possible to reconstruct the experiments from the sparse documentation that is provided. We aim to change this by publishing one way of assigning a correct event timestamp, extracting transients, and calculating features on the aggregate data with these known transients and timestamps. This should help new researchers get a foothold in this very active research area.
Perspectives
We wanted to avoid new researchers having to go through the same struggles of having to figure out how to solve the necessary preprocessing steps before evaluating an event-based NILM classifier. By providing a detailed description and evaluation of these processing steps, we aim to lower the entry-barrier in this field of research.
Justus Breyer
Read the Original
This page is a summary of: Practical Insights from Implementing Event-Based NILM Systems, June 2025, ACM (Association for Computing Machinery),
DOI: 10.1145/3679240.3734643.
You can read the full text:
Resources
Contributors
The following have contributed to this page
