Abstract: To address the issues of FFT transformation error, phase information loss, and insufficient nonlinear fitting ability in traditional frequency-domain harmonic source modeling, a TUNet time-domain harmonic source modeling method integrating UNet and an improved Transformer is proposed. This method uses UNet to achieve multi-scale feature extraction and waveform reconstruction, and embeds an improved Transformer to capture long-distance temporal dependencies and enhance the extraction of local distortion features, thereby realizing end-to-end time-domain voltage-current mapping. Through practical calculation examples, it is verified that this method can significantly improve the accuracy of harmonic waveform fitting, providing an efficient technical solution for the refined modeling of harmonic sources.