MODELING OF THE AXIAL COMPRESSION BEHAVIOR OF BORED PILE USING LONG SHORT-TERM MEMORY NETWORK

• Main Author: Gratio Deo Warouw, Anry
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/74839
• Institution: Institut Teknologi Bandung

Determining the bearing capacity of pile foundations and predicting the deformation behavior due to axial compressive loading are challenges for geotechnical and foundation engineers. Give the diverse soil conditions and complex interactions between the pile and the soils, the actual behavior of the pile likely always differs from predictions, including those using several methods available in foundation engineering. In recent years, advancements in artificial intelligence in the field of geotechnics have yielded highly accurate predictions. In this study, a deep learning method with Long Short-Term Memory (LSTM) techniques is used to predict the load-settlement curve for bored pile foundations. The pupose of this research is to develop an effective LSTM model for quicly predicting the load-settlement curve based on soil type and Standard Penetration Test (SPT) data, which can be used as a tool in planning and evaluating bored pile capacity berfore performing a static load test. A total of 47 soil data and 71 pile load test results were successfully collected covering various areas in DKI Jakarta. These static load data detail 47 data point in South Jakarta, 9 in West Jakarta, 10 in Central Jakarta, 4 in North Jakarta, and 1 in East Jakarta. The spread of equivalent soil types from the 71 pile and 47 soil data collected shows that fine-grained soils dominate the behavior of pile foundations in the DKI Jakarta area. This study has also sucessfully built a LSTM model to predict the load-settlement curve for bored pile foundations. The model development process includes preparation of pile catalog, data pre-processing, experiments, and testing. The optimal model configuration from 38 model experiments, consisting of Adam Optimizer, 380 epochs, 3 LSTM layers, and 20 LSTM neurons per layer, has a Root Mean Square Error (RMSE) performance of 519.9 kN and 618.6 kN for training and testing data, respectively, a Mean Absolute Error (MAE) of 423.3 kN and 492.4 kN for training and testing data, respectively, an R2 of 0.98 for training data and 0.96 for testing data, and a Mean Absolute Percentage Error (MAPE) of 11.1% for training data and 13.1% for testing data. The developed LSTM model exhibits good performance in terms of sensitivity analysis.