Non-intrusive load monitoring based on event decomposition is widely application for less storage and fast calculation speed. Firstly, methods on extracting load steady state feature were given, typical household appliance identification feature library was concluded. From the library, load harmonic features were important for air conditioner and other small power appliance which have even harmonics. Then, influencing factors of harmonic feature extraction in engineering application were researched. Based on the principle of spectrum leakage of FFT algorithm, the influence of grid frequency dynamics and electrical harmonic phase angle jitter on harmonic feature quantity extraction were studied. A Multi-point mean solution is proposed to solve the problem of non-synchronous sampling caused by grid frequency fluctuations, and the method of extreme value difference is proposed to solve the influence of appliance harmonic phase angle. With the two methods, the base harmonic error can be reduced to less than 1%, and the even harmonic error to 2%~4%. Finally, with the experimental platform and engineering, the effectiveness of the harmonic improvement extraction method is verified, and the load identification accuracy can be effectively improved by more than 5%.