Accurate prediction of electric vehicles charging loads is essential to enable fine-grained energy scheduling and ensure grid stability under large-scale vehicle-grid interaction. This study addresses this task by proposing a fine-tuning method based on TimesFM, a large-scale pure decoder-based model for time series prediction. The study first extracts charging behavior segments of typical vehicles from seven major cities in China to construct a minute-level load dataset, and introduces meteorological information and holiday factors as covariates. In the fine-tuning stage, three key strategies are adopted: (1) periodic window segmentation based on Fourier spectral analysis; (2) introduction of covariates to enhance the model’s ability to model non-stationary behaviors; and (3) multi-step rolling prediction mechanism based on a sliding window for overlaying long prediction periods. The results show that the fine-tuned model achieves an accuracy of about 85% in hourly monthly prediction and 90% in 15-min weekly prediction, which verifies the adaptability of the method in multiple time scales and complex scenarios, and provides a scalable solution for intelligent load management under large-scale vehicle-grid integration.