BayesSeg: Bayesian Optimization Framework for Electricity Consumption Segmentation

There's a new paper on arXiv, titled 2608.00513, introducing a framework named BayesSeg. It aims to improve Non-Intrusive Load Monitoring (NILM) by adaptively segmenting electricity usage data. The framework merges time-series segmentation with multidimensional assessment and automatic parameter optimization. For segmentation, it uses a dual steady-state criterion based on previous subsequence averages and tail values, allowing it to divide segments of steady and transition states without needing supervision. The evaluation part converts these segments into binary sequences and combines the event-level F1 score with Normalized Mutual Information (NMI) to measure event precision and state segmentation quality. This work addresses challenges in current methods, like parameter tuning and sensitivity to boundaries. You can find it at arXiv:2608.00513.