Estimating Reservoir Sedimentation Using Deep Learning

Abstract

Several reservoirs across the US are filling with sediment, which jeopardizes their functionality and increases maintenance costs. The US Army Corps of Engineers (USACE) developed the Reservoir Sedimentation Information (RSI) system to assess reservoir aggradation and track dam operation suitability for water-resource management and dam safety. The RSI dataset contains historical elevation-capacity data for approximately 400 dams (excluding navigation structures) which correspond to less than 1% of dams across the US. Thus, there is a critical need to develop methods for estimating reservoir sedimentation for unmonitored sites. The goal of this project was to create a generalized method for estimating reservoir sedimentation rates using reservoir design information and watershed data. To meet this objective, geospatial tools were used to build a refined composite dataset to complement the RSI system’s data with precipitation and watershed characteristics. Nine deep learning models were then used on the benchmark dataset to determine its accuracy at predicting capacity loss for the RSI reservoirs: four supervised machine learning models, four deep neural network (DNN) models, and a multilinear power regression model. A DNN model, containing a progressively increasing node and layer construction, was deemed the most accurate, with R2 values from its calibration and validation datasets being 0.83 and 0.70, respectively. The best model was recalibrated over the entire dataset, which showed greater accuracy on the prediction of RSI reservoir’s capacity loss, with an R2 of 0.81. This predictive model could be used to evaluate the capacity loss of unmonitored reservoirs, forecast sedimentation rates under future climate conditions, and identify reservoirs with the highest risk of losing functionality.