Edge computed NILM
Shamim Ahmed, Marc Bons
Abstract
In the context of residential Non-intrusive load monitoring (NILM), the usual service deployment process consists of collecting data from a metering device to the cloud, run algorithms on the cloud and then display results in a Web front or in an App. This approach comes with two major problems: on the one hand, important resources are allocated to the cloud process (including maintenance) while selling the solution on a substantial subscription basis is still a challenge. On the other hand, end-users are more and more reluctant to see their personal data being uploaded. In order to propose an alternative, this research has focused on edge computed NILM, namely the possibility to run NILM algorithms on existing devices on the end-user side, such as a smart phone. A two-stage model development has been carried out to obtain good disaggregation accuracy with lower model size. In the first stage, an efficient deep learning algorithm (MobileNet) has been adopted to obtain an accurate and light weight model. In the second stage, TensorFlow Lite has been used to compress further, in order to reduce edge device memory usage and computing time. To deal with real-life diversity, we have built large and diverse training and testing sets based on a combination of HES, UKDALE and REFIT datasets. Disaggregation performance has been assessed for both models: before and after TensorFlow Lite compression. Comparative analysis has been performed to facilitate implementation choices.