In recent years, forest wildfires have occurred frequently, yet effective identification methods are lacking in accident investigations to determine whether wildfires beneath power lines are caused by power system faults. By constructing a full-scale experimental platform, zero-sequence component signals from line-to-ground discharge faults caused by flames, as well as tree-line and broken-line faults, are obtained, and corresponding transitional impedance models for different fault scenarios are established. Based on three categories of fault features, namely, the fluctuation of total harmonic distortion (THD) of zero-sequence voltage, the proportion of wavelet energy across different coefficients, and long-duration waveform characteristics of zero-sequence current, a total of 14 fault features are extracted. A fault identification model is proposed using a support vector machines (SVM) guided by multidimensional feature F-score values, achieving an identification accuracy of 95.46%. By integrating post-fire electrical data, this approach can effectively reveal the causes of wildfires and provide a technical basis for determining whether such incidents are triggered by distribution network faults such as tree-line discharges.