Transfer capabilities of Seq2Seq and Seq2Point CNN architectures in Non-intrusive Load Monitoring with unseen appliances

In the Non-Intrusive Load Monitoring context, Seq2Seq and Seq2Point Convolutional Neural Network architectures have demonstrated state-of-the-art performance. However, as these methods suffer from high computational costs and the need for large volumes of training data, their transfer capabilities to different domains are essential for real-world implementation. This paper analyzes the drop in performance of Seq2Seq and Seq2Point architectures in the presence of appliances not seen in the aggregated power used for training. A theoretical analysis based on a first-order Taylor expansion is performed to analyze the structure of the additional error incurred. The experimental results showed a significant decrease in the performance of the methods when the noise increases, especially for monitored appliances with low-power states or complex patterns. The study reveals a strong dependence on the aggregated power structure in the training set and suggests that future methods should focus on learning robust appliance-specific signatures rather than directly regressing from the aggregated signal.