LightNILM: Lightweight neural network methods for non-intrusive load monitoring

The aim of non-intrusive load monitoring (NILM) is to infer the energy consumed by the appliances in a house given only the total power consumption. Recently, literature have shown that deep neural networks are the state-of-the-art approaches for tacking NILM. For example, both sequence-to-sequence (seq2seq) and sequence-to-point (seq2point) learning models are the popular frameworks with typical network architectures such as convolutional neural networks (CNNs). However, these deep neural network approaches are computationally expensive and require huge storage for the purpose of prediction, and consequently would not be capable of deploying on mobile/edge devices. This paper addresses these issues for seq2point learning models by employing specifically designed network architectures which can be processed by using TensorFlow Lite to deploy on mobile phones. We show that our models only require 0.5% number of the parameters used in original seq2point models, whilst achieve comparable accuracy. Our models are then successfully tested on mobile phones with reasonable accuracy performance.