Multi-task Learning for Non-intrusive Load Monitoring

Non-Intrusive Load Monitoring (NILM) is the process of inferring household electrical energy consumption patterns through access to aggregated signals. Compared to intrusive detection, NILM reduces costs by avoiding the need to install detection equipment for each electrical appliance. While sequence-to-sequence deep learning is a standard method for solving NILM problems, this paper proposes a multi-task-based architecture-Multi-Task Fusion (MTF). MTF goes beyond simply modeling energy consumption sequences and introduces subtasks to model the operating mode of household appliances accurately and establish correlations between them. Furthermore, compared to other traditional multi-task architectures, MTF introduces a dynamically balanced branch learning capability mechanism. According to this mechanism, MTF will perform balance fine-tuning during the training process, which brings performance improvement in prediction accuracy. Experimental results demonstrate that MTF outperforms traditional multi-task architectures and several other methods.