Harnessing Temporal Awareness: Dual-CNN Model for Non-Intrusive Load Monitoring

Thoroughly monitoring the energy consumption of each appliance in a household is essential to assist users in better engaging in power-saving practices. To get per-appliance electrical profile, non-intrusive load monitoring (NILM), also known as energy disaggregation, is a mix source separation task, that decomposes the total household electricity consumption into the load profile of each individual appliance. In recent years, deep learning models have demonstrated superior performance in NILM tasks. However, when encountering an unseen household, the model performance degrades seriously. In order to improve the performance on unseen households, we observe that temporal information is beneficial for the model generalization ability, as the ON/OFF patterns of electrical appliances follow a similar time schedule across different households. Thus, in this paper, we propose a time-aware dual-CNN architecture for NILM (TimeNILM), which fully exploits the temporal information and employs feature fusion based on attention mechanism to improve model performance. Experimental results on two datasets (REDD and UK-DALE) demonstrate that our model outperforms state-of-the-art models, achieving 8%-27% mean absolute error (MAE) gain and 8%-35% signal aggregated error (SAE<inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">${}_{\delta } $ </tex-math></inline-formula>) gain for unseen households. As appliance usage patterns are private data, in order to protect user privacy, we extend TimeNILM to federated settings and propose a scheme for aggregation during the learning process. Furthermore, we have successfully deployed TimeNILM on edge devices, and performance evalutation indicates that the model is capable of real-time load monitoring on a Raspberry Pi.