Forecasting Sediment Accumulation in the Southwest Pass with Machine-Learning Models

Abstract

Connecting the Mississippi River and the Gulf of Mexico, the Southwest Pass (SWP) is one of the most highly utilized commercial waterways in the United States. Hard-to-predict accumulation of sediments in the SWP affects the access of deep-draft vessels to four of the nation’s top 15 ports measured by tonnage. The U.S. Army Corps of Engineers (USACE) spends approximately 100 Million USD annually on dredging operations to maintain SWP at a 14.2-meter (50-ft.) depth. Presently, USACE project managers rely on rules-of-thumb with seasonal river stage trends and thresholds to get 10–14 days of lead time for shoaling conditions at the SWP. This work presents the development of a machine learning modeling framework to increase lead times and accuracy of shoaling forecasts in the SWP. Within a multivariate multistep timeseries forecasting framework, several regression models, input variables, and forecasting days are explored. All multivariate machine learning models outperformed an univariate ARIMA model used as baseline. A multilayered perceptron regressor implemented on a 60-day in-lag scenario was found to be the best model to forecast shoaling in the upcoming 45 days. The proposed model may be applied to forecast dredging needs at other critical waterways.