Application of long short-term memory and convolutional neural networks for real-time bridge pier scour prediction

Purpose Scour around bridge piers poses a major threat to structural safety, yet existing empirical and machine learning approaches are limited in accuracy, generalisation and real-time applicability. This paper aims to evaluate the effectiveness, efficiency and generalisability of deep learning models for real-time bridge pier scour forecasting. Design/methodology/approach The study formulates scour forecasting as a multivariate time-series prediction problem using real-world sensor data from monitored bridges in Alaska and Oregon (2006–2021). Multiple variants of long short-term memory (LSTM) and temporal convolutional neural network (CNN) models are designed and systematically evaluated. Random-search heuristics and sequential evaluation are used to optimise hyperparameters, improve robustness and assess computational efficiency. Findings Results show that LSTM models achieve high predictive accuracy for larger data sets, with a mean absolute error as low as 0.024 m for one-week-ahead forecasts. Fully convolutional networks deliver comparable accuracy with substantially lower training and inference times. Feature analyses indicate that historical scour time series dominate predictive performance, with flow-related features contributing only marginal or site-specific improvements. Research limitations/implications Model performance depends on data availability, noise levels and site-specific scour behaviour, particularly for smaller data sets. Future work should explore broader geographic validation and hybrid or physics-informed extensions. Practical implications The proposed models enable efficient real-time scour forecasting, supporting early-warning systems and improved decision-making for bridge monitoring and infrastructure risk management. Originality/value This paper presents one of the first systematic comparisons of LSTM and CNN models for real-time scour prediction using long-term field data, highlighting accuracy–efficiency trade-offs for operational deployment.