Author(s): Zhanchen Li; Dawei Guan
Linked Author(s):
Keywords: Pier scour; Unsteady flow; Machine learning; MLP; Transformer
Abstract: In mountainous regions, the velocity of river flow undergoes significant fluctuations between flood and low-water periods. The pronounced variations in flow velocity during flood events lead to substantial alterations in the riverbed topography, often resulting in considerable local scour around pile foundations. This phenomenon poses a significant threat to the stability of hydraulic structures. Hence, the development of an accurate prediction model for foundation scour under variable flow conditions is of paramount importance for enabling real-time early warning systems for hydraulic infrastructure. Currently, research on fundamental scour under unsteady flow predominantly relies on established empirical formulas derived from steady flow scour analysis. The predictive accuracy is closely linked to the precision of these empirical formulas, particularly in simulating scour processes, notably in the initial stages. With the advent of machine learning, a plethora of models and methodologies have emerged. Regarding the fusion of machine learning with engineering applications, neural network approaches have previously been employed in forecasting final scour depths. In this study, machine learning techniques are applied to investigate non-uniform flow scour, with specific emphasis on comparative analysis between the MLP and Transformer models. Focused on pier scour, the dataset utilized exhibits a declining segment. It is observed that the MLP model demonstrates remarkable accuracy in predicting scour scenarios with ascending flow, yet displays oscillations in more intricate scour processes. The Transformer model serves to address this issue, although its predictive consistency for individual scour events varies. This study proposes enhancements to the Transformer positional encoding to mitigate these challenges. The improved Transformer model achieves an RMSE of 0.105 for bridge abutment scour and 0.061 for bridge pier scour. This methodology facilitates rapid experimental training of scour evolution within short timeframes, offering a novel approach for developing time series models to predict scour evolution.
Year: 2024