Long-Lead Forecasting of Runoff Season Flows in the Colorado River Basin Using a Random Forest Approach

There is an increasing need for skillful runoff season (i.e., spring) streamflow forecasts that extend beyond a 12-month lead time for water resources management, especially under multiyear droughts and particularly in basins with highly variable streamflow, large storage capacity, proclivity to droughts, and many competing water users such as in the Colorado River Basin (CRB). Ensemble streamflow prediction (ESP) is a probabilistic prediction method widely used in hydrology, including at the National Oceanic and Atmospheric Administration (NOAA) Colorado Basin River Forecasting Center (CBRFC) to forecast flows that the Bureau of Reclamation uses in their water resources operational decision models. However, it tends toward climatology at 5-month and longer lead times, causing decreased skill, particularly in forecasts critical for management decisions. We developed a modeling approach for seasonal streamflow forecasts using a machine learning technique, random forest (RF), for runoff season flows (April 1–July 31 total) at the important gauge of Lees Ferry, Arizona, on the CRB. The model predictors include antecedent basin conditions, large-scale climate teleconnections, climate model projections of temperature and precipitation, and the mean ESP forecast from CBRFC. The RF model is fitted and validated separately for lead times spanning 0 to 18 months over the period 1983–2017. The performance of the RF model forecasts and CBRFC ESP forecasts are separately assessed against observed streamflows in a cross validation mode. Forecast performance was evaluated using metrics including relative bias, root mean square error, ranked probability skill score, and reliability. Measured by ranked probability skill score, RF outperforms a climatological benchmark at all lead times and outperforms CBRFC’s ESP hindcasts for lead times spanning 6 to 18 months. For the 6- to 18-month lead times, the RF ensemble median had a root mean square error that was between $\sim 410$- and $\sim 620\text{-}\text{thousand}$ acre-feet lower than that of the ESP ensemble median (i.e., RF reduced ensemble median RMSE by $-9\%$ to $-12\%$ relative to ESP). Reliability was comparable between RF and ESP. More skillful long-lead cross-validated forecasts using machine learning methods show promise for their use in real time forecasts and better informed and efficient water resources management; however, further testing in various decision models is needed to examine RF forecasts’ downstream impacts on key water resources metrics like robustness, reliability, and vulnerability.