A data-driven approach to integrated equilibrium-temporal scour forecasting at complex-pier structures using hybrid neural networks

Prediction of pier scour at complex structures has long been challenging due to various complex factor sets. This study proposed a hybrid neural network framework, including a multi-module multi-layer perceptron (MLP) network for predicting equilibrium scour depth and a long-short term memory (LSTM) network for temporal scour evolution forecasting, with extra data exchange between the two component networks for ensuring physical consistency. The multi-module MLP network passes 19 inputs into sub-networks representing different structural and flow factors for generating one output. The results show significantly better accuracy than existing empirical methods and standard simple neural networks. The LSTM network component provides multi-stepahead scour evolution forecasting based on monitored time sequences combined with physics-related sequences from the interim and final outputs of the multi-module MLP model. Sensitivity analysis showed that an antecedent period of 48 h and a forecast horizon of 6 h may yield optimal model performance. For the capacity of conducting extended recursive forecasting, errors may accumulate after certain stages with scour-rate fluctuation given an inappropriate selection of antecedent period and forecast horizon; further efforts can be taken to enhance this capacity. In general, the proposed data-driven model shows superior performance to traditional methods and good potential for integration with smart digital platforms.