Short lead-time streamflow forecasting by machine learning methods, with climate variability incorporated

• Main Author: Rasouli, Kabir
• Language: English
• Published: University of British Columbia 2010
• Online Access: http://hdl.handle.net/2429/27090

Streamflow fluctuates as a result of different atmospheric, hydrologic, and morphologic mechanisms governing a river watershed. Variability of meteorological variables such as rainfall, temperature, wind, sea level pressure, humidity, and heating, as well as large scale climate indices like the Arctic Oscillation, Pacific/North American Pattern, North Atlantic Oscillation, and El Niño-Southern Oscillation play a role on the availability of water in a given basin. In this study, outputs of the NOAA Global Forecasting System (GFS) model, climate fluctuations, and observed data from meteohydrologic stations are used to forecast daily streamflows. Three machine learning methods are used for this purpose: support vector regression (SVR), Gaussian process (GP), and Bayesian neural network (BNN) models, and the results are compared with the multiple linear regression (MLR) model. Lead-time for forecasting varies from 1 to 7 days. This study has been applied to a small coastal watershed in British Columbia, Canada. Model comparisons show the BNN model tends to slightly outperform the GP and SVR models and all three models perform better than the MLR model. The results show that as predictors the observed data and the GFS model outputs are most effective at shorter lead-times while observed data and climate indices are most effective at longer lead-times. When the leadtime increases, climate indices such as the Arctic Oscillation, the North Atlantic Oscillation, and the Niño 3.4 which measures the central equatorial Pacific sea surface temperature (SST) anomalies, become more important in influencing the streamflow variability. The Nash-Sutcliffe forecast skill scores based on the naive methods of climatology, persistence, and a combination of them for all data and the Peirce Skill Score (PSS) and Extreme Dependency Score (EDS) for the streamflow rare events are evaluated for the BNN model. For rare events the skill scores are better when the predictors are the GFS outputs plus locally observed data compared to cases when only observed data or any combination of observed and climate indices are chosen as the predictors.