In Non-Intrusive Load Monitoring (NILM) systems, event detection is used to ascertain the switching behavior of electrical loads, and its precision directly impacts the efficacy of power load identification. Conventional event detection methods focus exclusively on variations in line power fluctuations prior to and following the switching of electrical loads, neglecting the investigation of the line’s harmonic characteristics during load switching. This can lead to difficulties with low identification accuracy. This paper proposes an adaptive sliding filtering technique for NILM event detection, which uses a bilateral composite sliding window algorithm to filter sampling data. It introduces Total Harmonic Distortion (THD) as a secondary criterion alongside power, utilizing the complementary information from both harmonic distortion and power to adaptively determine the type of electrical load. This method effectively identifies when load switching events occur. Through comparison experiments with three typical power load switching, the adaptive composite sliding filter improved event detection by more than 25%, which shows that this method is effective.