Author(s): Farooque Rahman; Rutuja Chavan
Linked Author(s):
Keywords: Scour depth; Time-series; BiLSTM; Deep learning; Bridge
Abstract: Scour refers to the removal of sediment surrounding hydraulic structures, such as bridge piers, due to rapidly flowing water. Local scour is a significant risk to the stability and durability of bridge foundations. Precise prediction of scour depth is crucial for risk evaluation, construction, and management of hydraulic infrastructure. Traditional empirical and semi-empirical scour prediction formulae use simplified physical assumptions and site-specific calibration. Data-driven modeling has become popular in hydraulic engineering because it can capture nonlinear relationships from measured data. Chen et al. (2020) employ a machine learning methodology to identify water leaks in canal segments by integrating deep learning with canal assessment expertise. Seyedian et al. (2022) conducted a comparative analysis of several soft computing models, including GEP, ELM, least squares support vector machine (LSSVM), and GMDH, to estimate scour depth downstream of grade-control structures. Recurrent neural networks (RNNs), especially Long Short-Term Memory (LSTM) networks, can learn long-range temporal relationships within sequential data, making them ideal for scour prediction and has not been explored much in previous studies. Hence, this study aims to achieve primarily two objectives: (1) to generate robust deep learning models for predicting scour depth in time-series data, and (2) to conduct a comparison analysis of LSTM and BiLSTM models regarding their predictive accuracy and generalization capabilities.
Year: 2025