Prediction of flow-induced local scour depth at the uniform bridge pier using masked attention neural network. (15th December 2022)

• Record Type: Journal Article
• Authors: Lu, Bingjing; Petukhov, Valeriy; Zhang, Minxi; Wang, Xuhong; Yue, Shaolin; Zhou, Huan; Kholodov, Aleksei; Yu, Guoliang
• Is Part Of: Ocean engineering. Volume 266(2022) Part 3
• Publication Date: 2022-12-15
• Subjects: Bridge pier -- Scour depth -- Time series prediction -- Masking technology -- Attention neural network; Ocean engineering -- Periodicals; Ocean engineering; Periodicals; 620.4162
• DOI: 10.1016/j.oceaneng.2022.113018
• Languages: English
• ISSNs: 0029-8018

Abstract

Abstract: Evaluation of the local scour at a uniform bridge pier has become an important issue for hydraulic engineers. However, it is still difficult to accurately predict the local scour depth at the bridge pier using empirical formulae. Reasons underlying this could be the limitation of the statistical fitting technique and the complexity of flow and scouring mechanisms. In this study, a target-modified attention neural network called masked attention neural network (MATN) has been proposed to estimate the flow-induced local scour depth at the uniform cylindrical bridge pier. The experimental data used for training, validation and testing in the MATN model are obtained from several published references. Two laboratory experiments were also carried out to demonstrate the reliability and performance of the MATN model. This study shows that MATN model yields reliable agreements with the historical and test data sets by using small but important data within 10% of equilibrium scour time, which is better than those obtained from empirical scour depth formulae and two classical neural network models Back Propagation Neural Network and Recurrent Neural Network. It further indicates that the MATN model can realize a fast and cost-effective scour depth prediction based on short period of initial scour process. Highlights: A Masked Attention Neural Network (MATN) was designed to predict the temporal scour depth at a uniform bridge pier. Through capturing long-range dependencies and interactions of scour depth data, MATN predicts the scour depth reliably. Observed data within 10% equilibrium scour time are enough to train the model.