Event-steady State Mixed Non-invasive Load Identification Method based on Linear CRF

To address the challenges of high feature similarity, extensive pre-sampling data requirements, and real-time constraints in current non-intrusive load monitoring (NILM) methods based on steady-state characteristics, this paper proposes a hybrid approach that integrates load event and steady-state NILM utilizing a linear chain Conditional Random Field (linear-CRF). This method employs steady-state active power and reactive power as key load features. Specifically, a power matching technique and an enhanced multi-feature genetic algorithm are utilized for load event and steady-state recognition, respectively. Several recognition outcomes that meet the specified criteria are selected to establish candidate sets for both load events and load steady states. Subsequently, the logical relationships between adjacent load steady states and load events are analyzed, and based on these relationships, an event-steady state joint linear-CRF model is constructed. The conditional probability form of the linear-CRF is expressed as a combination of the objective function values for load event recognition and load steady-state recognition. Various candidate event-steady state combinations are optimized to determine the final recognition results. Finally, case studies conducted on multiple datasets demonstrate that this method can significantly enhance overall load recognition accuracy and improve the recognition of steady-state loads, thereby validating the feasibility and superiority of the proposed approach.