An Empirical Evaluation of the Performance of Deep Neural Networks on Delay Risk Prediction in Urban Flexible Pavement Projects in Iraq
Ongoing time overruns in urban Flexible Pavement Projects (FPP) highlight the inadequacy of traditional risk forecasting techniques, which often overlook nonlinear and project-specific delay factors. While recent Artificial Intelligence (AI)-based approaches have been proposed, most remain at a descriptive level, demonstrating only a few mathematically expressible and experimentally validated models suitable for urban road networks. This study addresses these gaps by developing a closed-form Artificial Neural Network (ANN) model using nine carefully selected predictors drawn from recent engineering practices and project data in Najaf, Iraq. The model incorporates advanced preprocessing, including robust outlier detection and min–max scaling, and is trained on a newly compiled dataset covering 35 major projects, thereby improving on previous studies' shortcomings in terms of both data quality and methodological transparency. Empirical results demonstrate that the ANN substantially outperforms baseline models, achieving an R2 of 0.847 and a Mean Absolute Percentage Error (MAPE) of 7.10%, with all improvements being statistically significant (p < 0.001). Additionally, feature sensitivity analysis identified payment delay and contractor experience as the most influential risk factors, underscoring the model's practical relevance. Importantly, the modular mathematical structure of the ANN facilitates transparent benchmarking and direct transferability to other urban regions, while creating a sound and replicable paradigm for impact-based, data-driven decision-making and planning infrastructure. Thus, the proposed model constitutes a benchmark for future research on predictive modelling of time overruns in urban pavement projects.