There is an increasing need for skillful streamflow forecasts that extend beyond a 12-month lead time (βyear-2 forecastsβ), particularly in basins with highly variable interannual streamflow, large storage capacity, proclivity to droughts, and many stakeholders such as the Colorado River Basin (CRB). Forecasts that go beyond seasonal timescales are highly sought after by water managers in the CRB in the hopes that they will aid planning when faced with multi-year droughts that strain storage capacity. Ensemble Streamflow Prediction (ESP) is a method currently used by the Colorado Basin River Forecasting Center to project flows that the Bureau of Reclamation use in their operational forecast models, but it defaults to pure climatology between 5- and 16-months into the future, depending on the month a forecast is made. We developed a machine learning approach using a random forest (RF) algorithm trained on variables including antecedent basin conditions, observed climate teleconnections, temperature, runoff efficiency, and climate model projections of future temperature and precipitation. Our random forest approach was tested against ESP in a 35-year hindcast between 1983-2017 for lead times spanning 0- to 18-months. ESP outperforms both climatology and random forest forecasts for lead times up to 3-months, but for lead times of 7-months or longer, ESP performs worse than both climatology and random forest. Conversely, random forest outperforms climatology at all lead times, including at 18-months where over half of hindcast years performed better than climatology. Machine learning approaches such as random forests have the potential to improve flow forecasts at longer than seasonal time scales and aid water management.
Graduate Student CEAE, CU ΒιΆΉΣ°ΤΊ
Μύ