Bidirectional spatio-temporal networks for predicting cost performance in construction using deep learning extension of earned value management

Network-based construction planning methods are often not incorporated into machine learning (ML) models, whereas existing temporal models struggle to effectively capture the short- and high-frequency nature of construction activities, as well as the spatial dependencies inherent in construction tasks. This paper presents a dynamic spatio-temporal ML architecture —bidirectional recurrent neural networks with graph convolutional networks (Bi-RNN-GCN)—that forecasts the cost performance index (CPI) within earned value management (EVM) control cycles. The system encodes spatial connectivity via work breakdown structure (WBS) logic and temporal sequencing through delivery scheduling. By structuring the construction progress data into a network format, where nodes represent individual activities and edges reflect both spatial and temporal dependencies, thereby better representing the functional realities of construction dynamics. In this way, both the temporal evolution and the spatial relationships among tasks are learned. The model was trained on a large dataset of over 200,000 progress records collected over 53 weeks, and different configurations were evaluated, including Bi-RNN-GCN, RNN-GCN, and GCN-RNN, against static benchmark models. The findings clearly indicated that the Bi-RNN-GCN model, which prioritizes bidirectional temporal learning followed by spatial learning, achieved 15 % higher accuracy than static models such as GCN.