Applied Artificial Intelligence for Delay Risk Prediction of BIM-Based Construction Projects

Abstract

Construction projects are complex endeavours requiring intricate coordination among stakeholders, resources, and timelines. However, delays are a pervasive issue, causing financial losses, reputational damage, and compromised outcomes. This research addresses this problem by integrating systematic reviews, expert insights, and cutting-edge artificial intelligence (AI) techniques. The study begins with a systematic review adhering to Systematic Reviews and Meta-Analyses (PRISMA) guidelines to identify common drivers of project delays. Analyzing scholarly articles from various regions and project types, the research develops a conceptual framework categorizing delay risk drivers into nine distinct groups. To validate these findings, an expert survey involving industry professionals is conducted, ensuring the research reflects real-world insights. This results in an empirically validated framework for assessing delay risks. Concurrently, the study reviews AI applications in construction, identifying supervised learning and deep learning as the most impactful technologies for predictive modelling. Using the validated delay risk drivers as features, the research develops hyperparameter-optimized AI predictive models through a process of feature engineering, model development, optimization, and evaluation. Among the evaluated models, the Fully Connected Neural Network (FCNN) demonstrates superior performance. To enhance model interpretability, the study employs SHapley Additive exPlanations (SHAP), providing transparent explanations for model predictions. This transparency fosters trust among stakeholders and enables targeted interventions to address critical delay risk drivers. The development and validation of the FCNN model represent a significant advancement in anticipating and mitigating project delays in construction. The integration of SHAP enhances the model's transparency and interpretability, empowering professionals with a powerful tool for proactive delay risk assessment and mitigation. This research makes substantial contributions to academic knowledge and industry practice by providing a robust predictive model and enhancing model interpretability. Acknowledging limitations such as data scope, industry dynamics, potential biases, and inherent machine learning constraints, the study suggests future research opportunities. These include exploring diverse datasets, incorporating new AI techniques, improving interpretability, integrating decision support systems, and leveraging synergies with emerging technologies.