Daily streamflow forecasting by machine learning methods with weather and climate inputs

SummaryWeather forecast data generated by the NOAA Global Forecasting System (GFS) model, climate indices, and local meteo-hydrologic observations were used to forecast daily streamflows for a small watershed in British Columbia, Canada, at lead times of 1–7 days. Three machine learning methods – Bayesian neural network (BNN), support vector regression (SVR) and Gaussian process (GP) – were used and compared with multiple linear regression (MLR). The nonlinear models generally outperformed MLR, and BNN tended to slightly outperform the other nonlinear models. Among various combinations of predictors, local observations plus the GFS output were generally best at shorter lead times, while local observations plus climate indices were best at longer lead times. The climate indices selected include the sea surface temperature in the Niño 3.4 region, the Pacific-North American teleconnection (PNA), the Arctic Oscillation (AO) and the North Atlantic Oscillation (NAO). In the binary forecasts for extreme (high) streamflow events, the best predictors to use were the local observations plus GFS output. Interestingly, climate indices contribute to daily streamflow forecast scores during longer lead times of 5–7 days, but not to forecast scores for extreme streamflow events for all lead times studied (1–7 days).

► Use observations, weather forecasts and climate indices in streamflow forecasting.
► Nonlinear machine learning methods generally outperform linear regression.
► Climate indices improve forecasts at 5–7 day lead, but not for high extreme events.