Risk-informed construction progress forecasting with spatio-temporal machine learning model

This study represents the first application of spatio-temporal machine learning (ML) models for construction progress prediction, addressing gaps where network-based methods used in planning cannot be reflected in ML models. Existing temporal ML models struggle with short-sequence, high-frequency construction activities. To overcome these challenges, 40 spatio-temporal models were evaluated, and among them, 14 GatedGNN-based models were identified as the most accurate and robust. The study utilized a dataset comprising over 2 million records obtained from 175 weeks of construction progress reports. The GatedGNN, GatedGNN-RNN, and GatedGNN-GRU models outperformed others, achieving accuracies between 80% and 95% in cost and schedule performance indicators. Data quality issues were modeled like data poisoning, label flipping, feature manipulation, and FGSM attacks. The GatedGNN models-maintained accuracy above 80% under these attacks; however, in small datasets, FGSM attacks reduced accuracy. The prominent models can assist project managers in better predicting outcomes and making strategic decisions.