A comparative analysis of LSTM, GRU, and Transformer models for construction cost prediction with multidimensional feature integration

Construction cost prediction remains a complex challenge due to the multidimensional nature of construction data and external factors. The objective of this study is to identify the most effective deep learning model for accurately predicting construction costs by comparing the performance of LSTM, GRU, and Transformer models. Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Transformer are advanced machine learning regression models widely utilized for data prediction tasks. This study investigates these models’ performance for construction cost prediction using a multidimensional feature framework. Through comprehensive evaluation and comparison, the Transformer model demonstrated superior performance, particularly excelling in handling complex feature interactions and long-sequence data. The LSTM model, while effective in capturing temporal dependencies, shows reliable performance but lags behind the Transformer in accuracy. The GRU model, although faster in training, proved less accurate and is less effective in handling outliers. Key features such as Total Area (TA), Site Area (SA), and Number of Floors (NF) were identified as significant predictors across all models, with the Transformer model proving particularly adept at capturing complex interactions. By integrating these features, this study contributes to improved cost management, thereby enhancing prediction accuracy and reliability.