Knowledge-based cross-modal fusion for long-term forecasting of grouting construction parameters using large language model

Abstract

Accurate long-term forecasting of grouting construction parameters is essential for foundation safety and the advancement of grouting automation. Existing methods have limited generalization due to diverse equipment and complex geological conditions. This paper addressed these challenges by proposing the Knowledge-based cross-modal fusion for long-term forecasting of Grouting parameters using Large Language Model (KG-LLM). This method captured the variations and relationships among grouting parameters by integrating domain-specific knowledge through construction knowledge and cross-prompt. A cross-modal fusion method combined knowledge-driven prompts with multi-scale time embedding into the frozen LLM, ensuring high prediction accuracy and generalization. Case studies on three projects validated the predictive performance and cross-engineering generalization of KG-LLM, with notable improvements in the prediction of parameters. KG-LLM quickly adapted to other projects without further training and was not constrained by equipment type. Moreover, this method was compatible with any LLM, offering a scalable solution for advancing the intelligent of grouting construction.