Machine Learning Generated Streamflow Drought Forecasts for the Conterminous United States (CONUS): Developing and Evaluating an Operational Tool to Enhance Sub-seasonal to Seasonal Streamflow Drought Early Warning for Gaged Locations

John C Hammond, Phillip Goodling, Jeremy Diaz, Hayley Rikert Corson-Dosch , Aaron Heldmyer, Scott Hamshaw, Ryan McShane, Jesse C Ross, Roy Sando, Caelan Simeone, Erik Smith, Leah Staub, William David Watkins , • Michael Wieczorek, Kendall Wnuk, Jacob Zwart 

Forecasts of streamflow drought, when streamflow declines below typical levels, are notably less available than for floods or meteorological drought, despite widespread impacts. To address this gap, we apply machine learning (ML) models to forecast streamflow drought 1-13 weeks into the future at > 3,000 streamgage locations across the conterminous United States (CONUS). We applied two ML methods (Long short-term memory (LSTM) neural networks; Light Gradient-Boosting Machine - LightGBM) and two benchmark model approaches (persistence; Autoregressive Integrated Moving Average - ARIMA) to predict weekly streamflow percentiles with independent models for each forecast horizon. To explore whether a training focus on dry weeks improved performance, both ML models were trained using all percentiles (LSTM-all, LightGBM-all) and only percentiles below 30% (LSTM<30, LightGBM<30). We evaluated model performance regionally and nationally for drought occurrence (the classification performance for a future date) and for drought onset/termination (performance identifying drought starts and ends). ML models generally performed worse than the persistence model for discrete classification (moderate, severe, extreme drought) of drought occurrence but exceeded the benchmark models for onset/termination. ML models outperformed benchmarks in predicting continuous streamflow percentiles below 30%. Occurrence performance was better for less intense droughts and shorter forecast horizons, with the ML models having predictive power at 1-4 week horizons for severe droughts (10th percentile threshold). All models struggled to forecast onset, though the best ML model was the LSTM<30 (sensitivity of 22%). Termination performance was greater, with the drought termination performance greatest for the LightGBM-all model. When estimating model uncertainty, the LSTM<30 model had the narrowest 90% percentile interval with closest to optimal capture. This work highlights the challenges and opportunities to further advance hydrological drought forecasting and supports an experimental operational streamflow drought assessment and forecast tool.