Prediction of Submarine Turbidity Currents Using Stacked LSTMs with Tree-Structured Parzen Estimator Optimisation
Farid Fazel Mojtahedi, Negin Yousefpour, Shiao Huey Chow, Mark Jason Cassidy · 2023
Turbidity currents are density flow of sediments that occur intermittently in deep seas. Large and fast turbidity currents can cut and erode continental margins and damage vital infrastructures such as seabed telecommunications cables. This paper aims to evaluate the application of Deep Learning method in the forecast of submarine turbidity currents, using long short-term memory networks (LSTM). The proposed LSTM models were trained on turbidity current datasets gathered by Ocean Network Canada, incorporating oceano-graphic, hydrographic, and sedimentological data. The proposed forecasting model for turbidity currents demonstrated high accuracy and reliability by utilising a Bayesian optimisation algorithm for hyperparameter tuning, along with Monte Carlo dropout and deep ensemble techniques for uncertainty assessment. Velocity predictions exhibited a 6.5% and 9% error, while turbidity forecasts exhibited a 5% and 8% error, based on evaluation criteria of mean absolute percentage error and normalised root mean square error. These findings suggest that the proposed deep learning model could serve as a valuable tool for monitoring and predicting turbidity currents.