Water Resources Research Act Program

Details for Project ID 2020DC142B

Development of a Real-Time Low Flow Forecast System Based on Machine Learning Methods

Institute: District of Columbia
Year Established: 2020 Start Date: 2020-03-01 End Date: 2021-02-28
Total Federal Funds: $10,000 Total Non-Federal Funds: Not available

Principal Investigators: Dr. Nian Ashlee Zhang

Project Summary: Potomac River basin streamflows are near or below normal, influenced by much below normal precipitation and near normal groundwater levels. If below average rainfall continues then further degradation in conditions is expected to occur. Daily monitoring of Point of Rocks and Little Falls flows within Potomac River basin began in September 2019 and will continue to prepare for the possibility that more serious drought conditions develop in the future. Therefore, it is imperative to provide an effective drought early warning system which uses the historical data to make prediction of the probability of flows dropping below drought trigger levels. The overarching goal of this project is to develop a real-time low flow forecast system that can forecast the probability of streamflows dropping below certain drought thresholds using the historical precipitation, streamflows, groundwater levels, and the palmer drought index time series data for the Potomac River basin. The state-of-the-art machine learning methods and their novel combinations will be explored to increase the prediction accuracy of the probability of current conditions dropping below the drought trigger levels. The objectives of the project are: a) Thoroughly investigate several state-of-the-art machine learning methods, including artificial neural network, support vector machine, naïve Bayesian, decision trees, radial basis function, k-nearest neighbors, and deep learning methods; b) Tailor these models to the streamflow prediction problem; c) Creatively build hybrid algorithms combining the most accurate models from the above machine learning algorithms; d) Test these hybrid intelligent methods using the historical time series data; e) Compare those ensembles to the individual machine learning methods in terms of the forecast accuracy.