A new improved variable step-size LMS algorithm is proposed in order to enhance the tracking velocity and accuracy of harmonic detection by variable step-size LMS adaptive algorithm in the low signal-noise rate (SNR) conditions of APF. Based on the MVSS-LMS algorithm, the improved algorithm adds the estimation of forgetting- weighted sum of historical errors to control the step-size updating and dynamically regulate the range of step-size. Also, sliding forgetting-weighted window is introduced to reduce computational complexity. The convergence speed and steady state performance of the algorithm are analyzed. According to experimental results, this algorithm possesses faster dynamic response speed and decreases steady state errors in low SNR conditions of APF. It has the high anti-jamming capability and better effect of harmonic current detection.