Low-Frequency Non-intrusive Load Identification Based on Two-Stage Event Detection Method

The randomness of loads and the diversity of power consumption patterns result in multiple events that may coincide or close to each other, making it difficult for a single event detection method to achieve the desired reliability and accuracy. In addition, most existing event detection algorithms use a fixed threshold to determine whether an event occurs, which has certain limitations. Therefore, this paper proposes a two-stage event detection method. The method uses an adaptive threshold and a sliding window based on the mean absolute deviation in the first stage, where the adaptive threshold is adjusted according to the load fluctuation to ensure that events of different magnitudes can be accurately detected. For the pseudo-aliasing phenomenon during steady-state operation, the cumulative sum of the changes in the differences of neighbouring detection points is introduced in the second stage to judge and filter the event points detected in the first stage. Simulating the REDD dataset with the self-built dataset shows that the method proposed in this paper can provide higher accuracy, flexibility, and reliability in event detection compared with the existing state-of-the-art algorithms. Subsequently, load features are extracted, and load recognition is performed using four classifiers. The average recognition accuracy of the REDD dataset is 98.80%, and the comparison with the existing literature reveals that the method of this paper provides higher accuracy in the field of load recognition. Meanwhile, the average recognition accuracy of the self-built dataset is 98.83%, further validating the method’s effectiveness in load recognition.