Implementation of Machine Learning-Based Risk Prediction Models for Large-Scale Infrastructure Construction Projects in Urban Environments

Large-scale infrastructure projects often use reactive approaches to manage construction risks. This can result in expensive delays and increased budgets. This study creates and tests a risk prediction framework that uses machine learning, specifically Gradient Boosting Decision Trees (GBDT), to help identify and address risks early in urban infrastructure construction. Data from 220 infrastructure projects, spanning from 2015 to 2024 and located in North America, Europe, and Asia, were analyzed. These projects had values between $50 million and $2 billion USD. The approach combined Principal Component Analysis and GBDT, handling 47 variables related to risk across six different risk areas. To test the model, 5-fold cross-validation was used, along with temporal validation, which involved setting aside the most recent 20% of projects. The GBDT model reached an overall prediction accuracy of 87.3%. It outperformed traditional methods by 23%. The ability to detect risks early on improved significantly, from 45% to 78%, and this led to an average cost reduction of 12.4%. Technical risks had the highest prediction accuracy, at 89.4%, while resource optimization saw a 25.7% improvement in equipment use. This machine learning-based framework is considered to significantly improve construction risk management. It offers better accuracy, earlier risk detection, and cost savings, suggesting it could be widely used in urban infrastructure construction.