This paper proposes a local polynomial modeling (LPM) approach and variable bandwidth selection (VBS) algorithm for identifying time-varying linear systems (TVLSs). The proposed method models the time-varying coefficients of a TVLS locally by polynomials, which can be estimated by least squares estimation with a kernel having a certain bandwidth. The asymptotic behavior of the proposed LPM estimator is studied, and the existence of an optimal local bandwidth which minimizes the local mean-square error is established. A new data-driven VBS algorithm is then proposed to estimate this optimal variable bandwidth adaptively and locally. An individual bandwidth is assigned for each coefficient instead of the whole coefficient vector so as to improve the accuracy in fast-varying systems encountered in fault detection and other applications. Important practical issues such as online implementation are also discussed. Simulation results show that the LPM-VBS method outperforms conventional TVLS identification methods, such as the recursive least squares algorithm and generalized random walk Kalman filter/smoother, in a wide variety of testing conditions, in particular, at moderate to high signal-to-noise ratio. Using local linearization, the LPM method is further extended to identify time-varying systems with mild nonlinearities. Simulation results show that the proposed LPM-VBS method can achieve a satisfactory performance for mildly nonlinear systems based on appropriate linearization. Finally, the proposed method is applied to a practical problem of voltage-flicker-tracking problem in power systems. The usefulness of the proposed approach is demonstrated by its improved performance over other conventional methods.