Predicting Long-Term Coastal Conditions in San Francisco Bay and Other Estuaries with the Use of Supervised Neural Networks

Neural networks were applied to predict long-term tidal currents in the San Francisco Bay in lieu of typical hydrodynamic simulations. Conventional numerical modeling can require significant computational power and calculation times; however, trained neural networks can provide near-instantaneous calculation of coastal conditions to supplement or replace traditional hydrodynamic modeling efforts. For this study, supervised networks were developed to forecast and hind-cast tidal currents in San Francisco Bay. Two different artificial neural networks were created to predict bay-wide tides and tidal currents throughout all of San Francisco Bay: a multi-layer perceptron (MLP) neural network, and a presumably more accurate long short-term memory (LSTM) recurrent neural network (RNN). Both neural networks were able to accurately forecast and hindcast long-term tides and tidal currents at any given time throughout all of San Francisco Bay. Using a trained neural network, long-term hydrodynamic model results can be obtained within seconds. Potential applications of the trained San Francisco Bay neural network include derivation of boundary conditions to drive smaller and more efficient nested hydrodynamic models, real-time prediction of hydrodynamics for navigation safety evaluations, and sediment or tracer transport for flushing studies.