Research on the Application of Generative Adversarial Network in Forecasting of Public Works

To solve the problem of “window periods” in railway line maintenance and the small-sample issue caused by the habit of recording partial data on various engineering machines in paper form, as well as to meet the requirements for predicting work completion and material consumption through machine learning, this study explored the application of generative networks in engineering work prediction. It also aims to meet the requirements for predicting work completion and material consumption through machine learning. The application of generative networks in engineering work prediction was explored. Five typical development models of GAN models were summarized and reviewed, and a three-level model selection method was established. First, the characteristics of the railway line maintenance work completion dataset were screened for data adaptability. Second, the models were further screened based on loss trends and the similarity of the generated data to actual data in terms of statistical properties. Finally, the optimal generative model was obtained using machine learning prediction results based on the study objectives. The WGAN model was identified as the optimal model, demonstrating compatibility with the characteristics of engineering data. The model's training loss reached equilibrium, and the statistical similarity between generated and actual data was the highest among all GAN models. Although machine learning prediction performed well, it was slightly inferior to GAN models prone to overfitting. The study determined the best model selection method and provided an optimal GAN model, offering a new solution for missing data in public engineering projects.