According to huge difference in the waveform characteristics between inrush current and internal fault current, the paper proposed a new method to identify inrush current based on wavelet packet and improved BP network. Decompose and reconstruct inrush current and fault current signal using wavelet packet, extracted wavelet packet reconstruction coefficients, calculate the energy of each band and normalized to construct energy feature vectors as input sample for BP network training and testing, and finally propose protection criterion. Through a large number of samples simulation using PSCAD / EMTDC and MATLAB software, it proves that the program can quickly and accurately identify inrush current and internal fault current.