A New Nonintrusive Load Monitoring Method via Transient-Steady-State Feature Fusion

Non-intrusive load monitoring (NILM) is an essential approach to promoting intelligent energy management. However, existing methods are limited by low event detection sensitivity and high false alarm rates for similar appliances in multi-appliance scenarios. To address these issues, this paper proposes a household load identification method based on the fusion of transient and multiple steady-state features. First, based on the high dynamic nature and sudden changes in current signals during load switching, a two-stage cumulative sum (CUSUM) event detection method is proposed to provide more accurate event detection and ensure that transient characteristics are not lost. Secondly, considering the diverse electrical characteristics of loads, a convolutional neural network (CNN) is used to fully explore the local correlations among different steady-state features, while the time-series features of transient active power are extracted using a bidirectional long short-term memory (Bi-LSTM) network. The importance of extracted steady-state and transient features is evaluated and adjusted through a channel attention mechanism. Then, the adjusted multi-dimensional features are fused to achieve accurate identification and classification of load types through the fully connected layer. Finally, experimental results on a real-world test platform and the PLAID public dataset demonstrate accuracy rates of 0.90 and 0.98, respectively, validating the effectiveness and superiority of the proposed method.