Aiming at the existing neural network fault location algorithms for ground faults on VSC-HVDC lines, there are too many training samples, long training time, and no effective verification of robustness is proposed. A method based on ST and PSO optimizes the line fault location algorithm of GRNN. From the perspective of the fault traveling wave energy spectrum, the ST is used to extract the fault transient voltage signal energy spectrum, and the energy representing each frequency interval is summed to achieve accurate extraction of the energy characteristic samples; and then normalized the subsequent energy samples and input to the neural network for training, and the PSO algorithm is used to optimize the smoothing factor of the GRNN to improve the network convergence speed and training accuracy. Finally, the electromagnetic transient simulation proves that the method has high positioning accuracy and is not easily affected by the transition resistance. In the case of input samples with measurement errors and external noise interference, the maximum error is still less than 1. 5%, which has certain engineering application value.