Artificial neural network modelling in simulation of complex flow at open channel junctions based on large data sets

The flow characteristics in open channel junctions are of great interest in hydraulic and environmental engineering areas. This study investigates the capacity of artificial neural network (ANN) models for representing and modelling the velocity distributions of combined open channel flows. ANN models are constructed and tested using data derived from computational-fluid-dynamics models. The orthogonal sampling method is used to select representative data. The ANN models trained and validated by representative data generally outperform those by using random data. Sobols' sensitivity analysis is performed to investigate contributions of different uncertainty sources to model performance. Results indicate that the major uncertainty source is from ANN model parameter initialization. Hence an ANN model training strategy is designed in order to reduce the main uncertainty source: models are trained for many runs with random model parameter initializations and the model with the best performance is adopted. ANN models can capture complex flow structures at combined open channels.ANN models should be trained with representative training and validation data.Representative data can be selected using spatial information of data points.Sobols' analysis reveals major uncertainty source in ANN parameter initialization.