Predicting Construction Schedule Delays Using Large Language Models

초록

This study investigates the performance of Large Language Models (LLM), specifically generative pretrained transformer, in predicting construction schedule delays through in-context learning approaches. Predictions were conducted both on unstructured data and on data transformed into structured formats, allowing for a comparative analysis of prediction accuracy across different levels of data preprocessing. The results showed that the prediction accuracy of the zero-shot approach was limited to 38.3%. In contrast, the few-shot approach achieved a prediction accuracy of up to 48.8% when trained on unstructured data and up to 73.9% when trained on structured data. These findings demonstrate the critical role of data structuring and prompt design in enhancing model performance. This study contributes to advancing the use of LLM in construction delay analysis and offers practical implications for improving schedule management and decision-making processes in the construction industry.

키워드