Time-series forecasting of consolidation settlement using LSTM network

Consolidation settlement refers to the deformation of soil due to external forces resulting in a reduction in the soil volume, posing a significant challenge for construction on soft ground due to the high compressibility of the soil. Methods, such as preloading and prefabricated vertical drains, have been applied to enhance the strength of the ground and accelerate the consolidation process. However, measurement-based methods, which are commonly used in practice, tend to produce inaccurate results when the measurement records are limited, and the prediction results may have a large variance depending on an engineering judgment. This study aimed to overcome these limitations by developing a long short term memory (LSTM) based model for predicting consolidation settlement with improved accuracy. The model was evaluated through 120 cases with different amounts of training data (10%–70%) and was compared with practical methods such as the hyperbolic and Asaoka methods. The LSTM model outperformed the practical methods, with an average RMSE and MAPE of less than 0.1m and 2%, respectively. The model was also capable of predicting the final settlement with an average MAPE of less than 3% regardless of the amount of training. Statistical analysis have also indicated that the LSTM model had the highest probability of accurately predicting the settlement. The results of this study indicate that deep learning algorithms can be successfully used to predict consolidation settlement and provide highly accurate predictions for construction projects on soft ground. This study has the potential to assist the design and construction of civil engineering projects.