Comparison of short-term streamflow forecasting using stochastic time series, neural networks, physical, and bayesian models

Authors: WAGENA, MOGES - Virginia Tech; GOERING, DUSTIN - National Weather Service; COLLICK, AMY - University Of Maryland Eastern Shore (UMES); BOCK, EMILY - Virginia Tech; FUKA, DANIEL - Virginia Tech; Buda, Anthony; EASTON, ZACHARY - Virginia Tech

Submitted to: Journal of Environmental Modeling and Software
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 2/14/2020
Publication Date: 2/17/2020
Citation: Wagena, M.B., Goering, D., Collick, A., Bock, E., Fuka, D.R., Buda, A.R., Easton, Z.M. 2020. Comparison of short-term streamflow forecasting using stochastic time series, neural networks, physical, and bayesian models. Journal of Environmental Modeling and Software. (126):1-10. https://doi.org/10.1016/j.envsoft.2020.104669.
DOI: https://doi.org/10.1016/j.envsoft.2020.104669

Interpretive Summary: Reliable streamflow forecasts are critical for decision making in water resources management. In this study, we used data from a long-term experimental watershed to assess the accuracy of short-term (lead times of 1 to 8 days) forecasts of streamflow made by two statistical forecasting models and a physically-based model that simulated the hydrological processes involved in streamflow generation. Results showed that the two statistical models generally outperformed the physically-based model in terms of forecast accuracy, with one of the statistical models providing insight into forecast uncertainty. Findings from the study provide insight into the most appropriate techniques for streamflow forecasting in headwater basins.

Technical Abstract: Reliable streamflow forecasts are essential for water resources planning and management. Although there are many proposed methods for forecasting streamflow, real-time streamflow forecasts remain challenging. The objective of this study is to compare and evaluate streamflow forecasts using four different techniques: a physical model (the Soil and Water Assessment Tool-Variable Source Area model-SWAT-VSA), a stochastic model (an Artificial Neural Network -ANN), an Auto-Regressive Moving-Average (ARMA) model with exogenous covariates, and a Bayesian ensemble model that utilizes the SWAT-VSA, ANN, and ARMA model results. Using these models, we forecast streamflow from 1 to 8 days, forced with the Quantitative Precipitation Forecasts (QPFs) precipitation data from the US National Weather Service, Weather Prediction Center. Of the individual models, SWAT-VSA and the ANN provide better predictions of total streamflow and peak flow during the calibration and evaluation periods, but the ARMA model captured baseflow better. During the forecast period the ANN model generally had the highest predictive power of the three individual models, however all three models tended to underpredict peak flow. The Bayesian ensemble model forecast streamflow with the most skill for all forecast lead times and provided a quantification of prediction uncertainty, something not possible with the individual models.