Innovative LSTM-transformer hybrid models for predicting consolidation settlement

Abstract This study utilized advanced deep learning algorithms to predict consolidation settlement in deep soft clay deposits, with a specific focus on the construction design phase of port facilities. Two innovative hybrid models, namely, the sequence long short-term memory (LSTM)-transformer (SLT) and the parallel LSTM-transformer (PLT) models, were introduced to generate accurate time-settlement predictions by incorporating geotechnical and construction information from sites where preloading was applied. The models were developed and tested using a dataset from study sites in Busan Newport, South Korea. This dataset was constructed through 3D interpolation, which provided a detailed and accurate representation of subsurface conditions. A case study was conducted to evaluate the performance of the model in real-world scenarios. The accuracy of the proposed models was compared with that of traditional methods, including the Hansbo method and a basic transformer model. Results indicated that the proposed models outperformed these traditional methods by producing more accurate predictions. In addition, a parametric study highlighted the effectiveness of the model in capturing the effects of critical factors, such as step loading period, maximum fill height, and clay layer thickness. The SLT and PLT models demonstrated significant potential for enhancing settlement prediction accuracy during the design phase. This improvement in accuracy aids in planning and increases cost effectiveness in projects involving soft deposits.