Erkyihun S.T., E Zagona, B. Rajagopalan, (2017). “Wavelet and Hidden Markov-Based Stochastic Simulation Methods Comparison on Colorado River Streamflow,” Journal of Hydrologic Engineering 2017, 22(9): 04017033 1-12, DOI: 10.1061/(ASCE)HE.1943-5584.0001538. © 2017 American Society of Civil Engineers.
Abstract
Wavelet and hidden Markov-based modeling frameworks were developed to better capture the nonstationarity and non-Gaussian characteristics of streamflow that linear models cannot. Climate-based conditional streamflow simulation techniques recently have been shown to perform even better in capturing the spectral characteristics of streamflow coupled with block bootstrap simulation. This paper presents a comparison of three recently developed time series models in these frameworks: the climate wavelet autoregressive model (CWARM), the climate hidden Markov model (CHMM), and the climate wavelet-based k-nearest neighbor (K-NN) time series bootstrap (CWKNN) model. The purpose is to determine their applicability in water resources planning and management. These three methods incorporate two large-scale climate forcings, Atlantic multidecadal oscillation (AMO) and Pacific decadal oscillation (PDO)—recognized as the drivers of underlying nonstationarity—to condition the streamflow simulation. Comparisons are made of performance in both simulation and projection modes using the Lees Ferry (Arizona) flow in the Colorado River basin (CRB). The three methods are generally very good at capturing the distributional statistics and nonstationary features of the historical data in simulation mode. For short-term projections (1–8 years), important for midterm reservoir operations and planning, the CHMM appears to perform slightly better than the other two models. For longer-term projections (∼20 years), useful for decadal and multidecadal water resources planning, the CWKNN performs much better.