A Heterogeneous Multiple-Experts Approach to Low-Frequency Nonintrusive Load Monitoring

Many nonintrusive load monitoring (NILM) approaches to decomposing household-level energy consumption into appliance-specific usage patterns using aggregated smart-meter data in conjunction with advanced deep learning techniques are able to leverage the complementary information existing between the power consumption estimation and appliance state detection tasks using multi-task frameworks. However, existing multitask frameworks suffer from three primary limitations, including a reliance on a single shared neural network that is unable to extract features for both tasks fully, particularly under long sampling intervals on the order of 15 min, the application of homogeneous neural network architectures that are unable to extract the diverse characteristics of NILM data, and the reliance on static weights to balance the competing task objectives. The present work addresses these issues by developing a multiple and heterogeneous deep learning expert-based approach to NILM conducted using low sampling frequency smart-meter data. Specifically, the proposed approach leverages a multi-gate mixture-of-experts architecture comprising five different neural network architectures, to capture both local and global time-series features effectively. Additionally, a novel multi-objective optimization strategy is designed in conjunction with the gradient descent algorithm to manage tradeoffs between the appliance state detection and power consumption estimation tasks dynamically by optimizing the weights for each task. The performance of the proposed approach is compared with that of existing state-of-the-art methods. Generalizability is validated through cross-household testing on unseen homes in the public UK-DALE and REDD datasets, while robustness to realistic data sparsity is demonstrated on a real-world dataset sampled at challenging 15-minute intervals. The results obtained for the proposed approach demonstrate significant improvements in state classification accuracy and power estimation error, particularly for appliances with complex power consumption patterns.