Construction Delay Prediction Model Using a Relationship-Aware Multihead Graph Attention Network

Abstract

Existing machine learning (ML) delay prediction models cannot process dependencies among the construction progress records. This study investigates graph attention networks (GAT) incorporating multihead attention mechanisms for predicting construction delays. Leveraging an attention mechanism, GAT emphasizes differential node significance in networks and demonstrates the capability to learn input configurations. The data set was configured into six networks that linked records based on contractual alignment and spatial proximity dependency criteria. Under contractual alignments, predictions for electrical and concrete tasks achieved 65% and 76%, respectively, outperforming spatial-based predictions. However, multihead GAT with spatial networks delivered 77% for insulation tasks, overtaking 67% of contractual networks, underscoring model sensitivity to task dependencies and its applicability across a range of decision making contexts. Recognizing the dependencies and shared aspects among construction records, the proposed GAT model better reflects human understanding of construction progress reports, shifting the focus from mere predictive accuracy to representative modeling of construction delay. The proposed model effectively models and learns dependencies between various construction activities, providing early warnings of potential delays. It is designed to provide early warnings of potential delays, empowering project managers to take swift and informed actions to mitigate negative impacts. For instance, if a specific task, such as electrical work, is predicted to fall behind due to an underestimation of person-hours or material quantities, the GAT model generates an alert. This early warning enables construction professionals to reallocate resources or adjust schedules in real time, ensuring that delays in one activity do not cascade into subsequent tasks. This flexibility makes it useful for diverse project environments, like assessing schedule plans and performance indices based on contractual or spatial dependencies. Our experiments demonstrated that different connectivity structures in the model result in varied accuracies, learning unique activity associations depending on the criteria—whether contractual, where tasks are linked by agreements, or spatial, focusing on site logistics. Beyond delay prediction, the GAT model is invaluable for continuous performance monitoring, enabling real time issue identification and resolution. Its flexibility across various project types and scales is vital for modern construction management, promoting a data-driven approach that reduces disruptions and enhances overall project success.