Addressing the problem of sparse samples with large delay times and limited quantity of delay data in train operation data, this paper proposes a train delay data generation method based on an improved Conditional Tabular Generative Adversarial Network (CTGAN). The method constructs train operational scenario features from the train data as input information and utilizes a Transformer to improve the generator in the CTGAN, enabling it to better recognize the spatiotemporal correlation features in the input information. Statistical similarity and distribution similarity metrics are used to evaluate the quality of the generated data. Results indicate that the generated delay data has high similarity with the original delay data. Furthermore, compared to the original data, there are more samples with large delay times, which can compensate for the sparsity problem of large delay data in the original delay data.