Robust Long-term Streamflow Forecasting

Project ID: 7294
Principal Investigator: Roger Hansen
Research Topic: Water Supply Forecasting
Priority Area Assignments: 2012 (Climate Change and Variability Research), 2013 (Climate Change and Variability Research)
Funded Fiscal Years: 2012
Keywords: long-term forecasting, water supply, modeling

Research Question

In the western U.S., water resource managers face a growing challenge to meet water demands for a wide variety of purposes (municipal, irrigation, environmental issues, hydropower, recreation, etc.) under the stress of increased climate variability (Grantz et al. 2005) and decreased water availability. Hoerling et al. (2009) estimated the Colorado River flow may decline 5-20 percent by 2050 (Reclamation 2011). Also, it is estimated that 44 percent of renewable water supplies are consumed annually in the western U.S. as compared with 4 percent in the rest of the country (el-Ashry and Gibbons 1988; Grantz et al. 2005). In this context, accurate and reliable anticipation of future streamflow could be extremely valuable to water managers in making the best possible decisions for management of water resources.

The objective of this research project is to develop and demonstrate a new data-driven modeling approach to provide long-term forecasts of streamflow. The modeling approach will incorporate wavelet-based analysis techniques used in statistical signal processing and a multivariate relevance vector machine (MVRVM) that uses a Bayesian regression method. We will develop a methodology that detects patterns in changes in Pacific sea surface temperature (SST), snowpack, and streamflow using wavelet decomposition. This information will then be used to improve the forecasting potential of the MVRVM.

The research question is: Can the combined modeling approach of wavelet decomposition and MVRVM Bayesian regression capture sufficient information from available climate data at meaningful temporal scales to achieve statistically significant improvements in long-term forecasts of streamflows for streams in Bureau of Reclamation (Reclamation) regions?

Need and Benefit

A number of studies in the western U.S. have used winter snowpack as inputs for streamflow forecasting modeling because much of the runoff is derived from snowmelt (McCabe and Dettinger 2002; Pagano et al. 2004; Grantz et. al 2005; Lettenmaier et al. 2008). Also, researchers have been studying the link between large-scale phenomena (e.g., El Niño-Southern Oscillation [ENSO] and the Pacific Decadal Oscillation [PDO]), and the hydroclimatology of the western U.S. (Jain and Lall 2000; Grantz et al. 2005; Gillies et al. 2010). These large-scale climatic phenomena have been added as inputs to improve streamflow forecasting (Clark et al. 2001; Grantz et al. 2005; Soukup et al. 2009) and are more useful for long-term forecasts.

However, the links between large-scale climate indices (e.g., ENSO, PDO, etc.) and hydroclimatology often fail to provide forecast improvement in every individual basin (Grantz et al. 2005). For example, while streamflow during ENSO has been observed to be below average in the Pacific Northwest and above average in the desert Southwest, use of ENSO information provides limited forecast improvement for basins outside these regions (McCabbe and Dettinger 2002; Grantz et al. 2005). It has been found that relatively minor changes in large-scale atmospheric patterns can result in large differences in surface climate (Yarnal and Diaz 1986; Grantz et al. 2005). Development of methods to capture these relationships is extremely critical for successful streamflow forecasting.

While researchers search for which Pacific Ocean patterns would be suitable to improve streamflow forecasts for a specific basin, they also face the unavoidable problem that these patterns (i.e., ENSO variability of 2 to 10 years) may change (increase or decrease) due to climate change (Collins et al. 2010). Therefore, there is an urgent need for techniques that effectively assimilate important information from new trends of the climate data into models that learn these patterns in order to produce improved streamflow predictions.

The data (i.e. streamflow, SST, snow water equivalent) tend to be periodic in nature, with clear seasonal patterns. However, in addition to periodic behavior, the data contain some extreme events with a frequency spectrum that changes in time. Therefore, traditional Fourier analysis, which decomposes signals into infinite sine and cosine components, is not a good fit. Dynamic Fourier analysis, sometimes referred to as Windowed Fourier analysis, might be of some help, but the series is assumed to have an approximately constant frequency spectrum within the windows used to break up the series. Furthermore, the Dynamic Fourier analysis is based on local sinusoids, which in turn, can introduce artifacts attributable to oversmoothing the signal. It has inherent spectral and time resolution problems, especially with limited datasets. More importantly, time-frequency decomposition of the series is needed in order to use it as inputs for the MVRVM model. For these reasons, there is a need to use wavelet-based techniques to deal with nonstationary data, especially for isolating climate change effects.

Contributing Partners


Research Products

Conference Presentation
Due: Sep 28, 2013
Comments: An abstract will be submitted to American Geophysical Union (AGU) Conference once the model is developed.

Journal Publication
Due: Sep 28, 2013
Comments: A peer reviewed journal article will be developed from the material in the final report.

Final Report in Reclamation format
Due: Sep 30, 2013
Comments: Final Report will include the results of the modeling effort, long term forecasts evaluation, together with a quantification of the uncertainty in the forecasts and recommendations for future research.

Last Updated: June 29, 2015