Towards Practical and Accurate Prediction of Earth-Retaining Wall Deformation from Geotechnical Time Series Using Attention-Based LSTM

Predicting the deformation of earth-retaining walls (ERWs) is crucial for assessing the stability of excavation sites. Traditional numerical methods often fall short in capturing real-field measurement dynamics and demand considerable modeling effort and computational resources. Recently, machine learning (ML) algorithms have gained traction for predicting the deformation of ERWs. However, a major limitation of many ML models is their reliance on extensive variable sets to describe the complex behavior of these structures, which restricts their practical application. In this research, we introduce a deep learning-based methodology for conducting time series analysis using geotechnical inclinometer data to efficiently predict the deformation profiles of ERWs. We employed traditional statistical ARIMA (Auto-Regressive Integrated Moving Average), deep learning-based LSTM (Long Short-Term Memory), and an attention-enhanced LSTM (ATLSTM) for analyzing time series data from geotechnical inclinometers, assessing the models' suitability under varying test conditions. This paper details the preprocessing steps tailored for the inclinometer data to optimize the evaluation of ERWs. It compares experimental results across different data lengths and input variable configurations. Our findings indicate superior prediction performance of the LSTM model over ARIMA. Additionally, the ATLSTM model, which integrates LSTM with an attention mechanism, demonstrates enhanced accuracy in both short-term and long-term predictions and shows resilience against domain-specific changes at drilling sites. The study concludes that effective prediction of wall deformation can be achieved through streamlined time series analysis of geotechnical inclinometer measurements, advocating for a reduction in input variables to improve field applicability.