The complex circuit structure inside the electric vehicle has the problem of poor connection point contact, which will cause the series arc failure, making the electric vehicle a safety hazard. In this paper, an electric vehicle arc fault experimental platform was designed to simulate the occurrence of arc faults at different loads and circuit locations under real working conditions. By analyzing the current signal changes before and after the arc, it was found that the current amplitude of the dry circuit current waveform is reduced and the waveform fluctuation is intensified before the arc is ignited, which provides a basis for the arc prediction. In this paper, a series arc fault prediction model based on Ghostnet network was constructed, which predicted the arc fault directly with the dry circuit current as input, avoiding the influence of artificial feature extraction on detection accuracy in traditional methods and improving real-time performance. The prediction accuracy of arc faults under various operating states is up to more than 90%, demonstrating good performance. The network generalizability analysis further verifies the stability and adaptability of the model on new data, proves the effectiveness of the model on series arc fault prediction, and provides a new technical means for series arc fault prediction.