The accuracy of harmonic current compensation on UPQC parallel side directly affects the quality of power supply. The traditional ip-iq method for extracting instruction current signals uses low-pass filters and coordinate transformation algorithms, which make the low-frequency harmonics unable to be filtered and there is a beat delay in compensation. Aiming at the problem of accuracy of UPQC instruction current signal extraction, a harmonic current extraction method based on RBF neural network and ip-iq method is proposed, which increases low frequency harmonic extraction, and the radial basis neural network is trained to fit the time-lag error in the ip-iq compensation process, generates harmonic current compensation module to compensate for the ip-iq method harmonic current extraction deviation. The current compensation effect of the traditional ip-iq method and the proposed instruction current extraction scheme is analyzed. After improvement, the harmonic content of the grid current decreases from 6.64% to 4.25% of the ip-iq method, which verifies the effectiveness of the proposed instruction current extraction strategy.