Short-term streamflow forecasting by ensemble Bayesian neural network models with climate variability

For daily streamflow forecasting at lead times of 1 to 7 days, statistical/machine learning models were used with three types of predictors -- (i) numerical weather prediction model output (from the NOAA Global Forecasting System (GFS) reforecast), (ii) local observations and (iii) indices of climate variability. To model the nonlinear relation between streamflow and the predictors, Bayesian neural network (BNN), a nonlinear machine learning method, was used to perform ensemble forecasting, with the results compared against multiple linear regression (MLR). Forecast skill scores were calculated relative to (a) climatology, (b) persistence, and (c) the optimal linear combination of both. Based on the contingency table for binary forecasts, two skill scores -- the Peirce skill score and the extreme dependency score -- were used for assessing the forecast skills of extreme streamflow events.

The results from a small coastal watershed, the Stave River basin in British Columbia, Canada, showed that ensemble BNN outperformed MLR. At shorter lead-times of 1-4 days, the most effective predictors were the local observations plus the GFS model outputs, whereas at longer lead-times of 5-7 days, the most effective predictors were the local observations and the indices of climate variability. For the longer lead-times, climate indices such as the Arctic Oscillation (AO), the North Atlantic Oscillation (NAO), and the equatorial Pacific Nino 3.4 region sea surface temperature were useful in streamflow forecasting, likely due to the interannual fluctuations of the accumulated snow in the river basin from climate variability.