Investigating Deep Learning Generalization Capabilities for Non-Intrusive Load Monitoring from Synthetic to Real-World Datasets

This thesis addresses the Non-Intrusive Load Monitoring (NILM) task with the aim of accurately reconstruct and identifying the individual active power consumption of domestic appliances, directly from the aggregated load signature. Given the predicted consumption, the NILM model also classifies the operational status (‘On’/’Off’) of the considered device, at each time step. The focus of this work is to assess the generalization capability of the BERT model, specifically adapted for NILM, on the UK-DALE, an open-access dataset containing real-world power consumption data. For this purpose, a public and synthetic dataset SynD has been tailored to effectively train and validate the BERT4NILM model, before assessing its performances on the real-world dataset. The analysis will concentrate only on five common household appliances: fridge, washing machine, dishwasher, kettle, and microwave. The NILM model performs well on synthetic dataset, achieving high accuracy, F1 score, and low relative error when trained and tested on different portions of the data. However, when evaluated on a real-world dataset, the model is not able to correctly generalize, especially for less frequently used devices like microwaves and kettles. This result can be mainly attributed to the differences between synthetic and real-world data, both in terms of consumption patterns and overall data distribution. Synthetic datasets, like SynD, are artificially generated and often simplify the complexities that characterize real consumption patterns, leading the model to fail in correctly recognizing devices.