|Van Liew, Michael|
Submitted to: American Society of Civil Engineers
Publication Type: Proceedings
Publication Acceptance Date: December 1, 2004
Publication Date: July 18, 2005
Citation: Garbrecht, J.D., Schneider, J.M., Van Liew, M.W. 2005. Linking climate forecasts and watershed runoff prediction using a neural network approach. In: Moglen, G.E., editor. Proceedings of the American Society of Civil Engineers 2005 Watershed Management Conference. Managing Watersheds for Human and Natural Impacts, July 19-22, 2005, Williamsburg, Virginia. 2005 CDROM. Interpretive Summary: Abstract Only.
Technical Abstract: NOAA's seasonal precipitation forecasts have been issued since 1995, but have not been widely used in watershed runoff applications. These seasonal forecasts are probabilistic in nature and define the odds of every possible precipitation outcome over the forecast period. The collection of precipitation outcomes is commonly referred to as an ensemble. Developing a corresponding ensemble of watershed responses requires numerical simulation for several hundred possible precipitation outcomes. The computational effort of repeated evaluation of precipitation ensembles using traditional hydrologic models is often impractical in engineering applications. An Artificial Neural Network (ANN) approach has comparatively fewer inputs and performs the same simulation in a fraction of the time. Here, the use of an ANN is illustrated for a 795 km2 watershed in Central Oklahoma that supplies runoff to a multi-purpose reservoir. For this application, National Weather Service precipitation observations were available at 4 stations outside the watershed boundaries, and reservoir inflow estimates were provided by the Bureau of Reclamation and the U.S. Army Corps of Engineers. The ANN was trained (calibrated) using 32 years of precipitation and runoff data, and validated against 8 additional years of data. Inputs consisted of monthly precipitation and storm type. The ANN was able to account for over 70% of the monthly runoff variability, and reproduced the monthly runoff distribution near identically. Using the trained ANN, watershed response to both a wet and a dry precipitation forecast (with magnitudes consistent with past forecasts) was evaluated for wet, average and dry antecedent conditions. The ensemble of watershed responses was displayed as probability of exceedance curves, which are a convenient way to capture the change in runoff odds associated with varying antecedent and forecast conditions. With respect to watershed runoff prediction, antecedent wet or dry conditions were shown to be a stronger predictor than seasonal precipitation forecasts which are typically of modest magnitude. As seasonal climate forecasts continue to improve in size of departure from average and in skill, the utility of seasonal forecasts for watershed applications should increase. Overall, the ANN approach presents a practical alternative to predict watershed response to seasonal precipitation forecasts that are issued on a monthly basis.