A GAN-LSTM Based Framework for Dynamic Project Scheduling and Risk Prediction in Engineering Management

Abstract

As the complexity of large-scale engineering projects increases, traditional schedule management methods face the dual challenges of insufficient prediction accuracy and lagging risk response in a dynamic environment. This study proposes a hybrid framework that integrates a conditional generative adversarial network (cGAN) with a bidirectional long short-term memory network (Bi-LSTM) to achieve dynamic project progress generation and real-time risk warning. The model was trained on a real-world dataset from a cross-sea bridge project, comprising 12 months and over 50,000 process records. The Bi-LSTM module captures temporal dependencies among construction processes, while the cGAN, trained with Wasserstein distance and gradient penalty, generates diverse and plausible progress scenarios conditioned on real process states. The adversarial training uses minimax loss with Adam optimization. Experimental results indicate that the proposed model reduces the mean square error (MSE) by 23.7% compared to a standalone LSTM in progress prediction, and achieves a risk identification accuracy of 91.4%, surpassing a traditional Bayesian network by 18.2%. In dynamic disturbance simulations, the model achieves an early-warning response time of 2.3 hours for emergencies such as resource shortages or extreme weather—67% faster than the baseline. Furthermore, the adversarial generation of 12,000 virtual progress samples mitigates overfitting in small-sample settings and improves the test-set F1-score to 0.89. This study demonstrates the robustness of the cGAN-BiLSTM framework in complex engineering environments and offers a data-driven solution for dynamic schedule optimization and risk management.