The time-varying parameter identification of load models has attracted broadly attention when large amounts of intermittent distributed generations (DGs) and stochastic loads are integrated in active distribution network (ADN). However, in traditional time-varying parameter identification approaches, the load model is usually developed with the composite load model (CLM) or synthesis load model (SLM), which are not suitable for representing the dynamics of DGs. Moreover, the plateau phenomenon and the continuous low-quality data further reduce the performance of traditional load models with time-varying parameter. Therefore, to address these issues, a novel time-varying parameter identification approach with extended Kalman filter (EKF) is designed for the load modeling. To represent the behavior of modern loads, an improved load model contains a parallel SLM and voltage source converter (VSC) is developed. Then, the target parameters, whose changes produce larger variations in model outputs, are selected with trajectory sensitivity to avoid plateau phenomenon. Also, the Chi-square test and the proposed weighted suppression strategy are utilized to suppress the continuous low-quality data. The simulation results on a system-level ADN model show that the proposed approach could accurately identify the parameters of time-varying load models.