TY - THES A1 - Kronberger, Moritz T1 - Context-Aware NILM: Disaggregating Electrical Load Profiles with Multimodal Co-Learning and Textual Metadata N2 - Non-Intrusive Load Monitoring (NILM) uses machine learning to predict the power consumption of individual electrical consumers from an aggregate source of power data, removing the need for costly sensor hardware. This thesis introduces a multimodal co-learning approach that integrates consumer-specific textual metadata as an additional modality for NILM models. Specifically, metadata are encoded into fixed-length text embeddings, which are fused with the aggregate power data sequence in a Transformer-based deep learning model. This additional input specifies which consumer’s power consumption to predict, allowing one model to be trained to disaggregate arbitrary consumers. Representing consumer names and types as word embeddings allows the multimodal NILM model to identify the target consumer and achieve disaggregation performance comparable to models trained individually for each consumer. However, encoding additional contextual information into consumer representations did not improve performance. Future work on adapting the training strategy to better exploit the textual modality could increase the model’s ability to utilize contextual information, enabling more generalizable NILM models with better applicability in real-world energy monitoring. Y1 - 2025 U6 - https://doi.org/10.60524/opus-2827 ER -