Event-based non-intrusive load monitoring methods (NILM) have been extensively researched and implemented owing to their numerous advantages. Detecting load events constitutes the foundational stage of event-based NILM techniques. Accurate detection of the entire power transient process of events can directly improve the ultimate effectiveness of NILM. However, it’s a challenge for the existing event detection methods to achieve excellent performance in diverse scenarios. On the one hand, this is because events from appliances with different operating principles have significant differences in time scales, amplitudes, and power waveforms. On the other hand, prevalent load fluctuations, along with background noise, can adversely affect the detection results. To solve the problems exposed by existing studies, this paper proposes a temporal characterization load event detection method in NILM. The proposed method first calculates the adaptive threshold for each power sample, based on which the events can be located. Subsequently, the steady-state search sliding windows are moved twice to locate the steady-state periods before and after the events to determine the areas where the events exist. Finally, the exact start and end points of events can be accurately detected in these areas. In this paper, comparison experiments on private datasets and BLUED datasets demonstrate that the proposed method surpasses the effectiveness of two other state-of-the-art event detection methods.