Search

Improved forecasts of atmospheric river (AR) events, which provide up to half the annual precipitation in California, may reduce impacts to water supply, lives, and property. We evaluate quantitative precipitation forecasts (QPF) from the High-Resolution Rapid Refresh model version 3 (HRRRv3) and version 4 (HRRRv4) for five AR events that occurred in February–March 2019 and compare them to quantitative precipitation estimates (QPE) from Stage IV and Mesonet products. Both HRRR versions forecast spatial patterns of precipitation reasonably well, but are drier than QPE products in the Bay Area and wetter in the Sierra Nevada range. The HRRR dry bias in the Bay Area may be related to biases in the model temperature profile, while integrated water vapor (IWV), wind speed, and wind direction compare reasonably well. In the Sierra Nevada range, QPE and QPF agree well at temperatures above freezing. Below freezing, the discrepancies are due in part to errors in the QPE products, which are known to underestimate frozen precipitation in mountainous terrain. HRRR frozen QPF accuracy is difficult to quantify, but the model does have wind speed and wind direction biases near the Sierra Nevada range. HRRRv4 is overall more accurate than HRRRv3, likely due to data assimilation improvements, and possibly physics improvements. Applying a neighborhood maximum method impacted performance metrics, but did not alter general conclusions, suggesting closest gridbox evaluations may be adequate for these types of events. Improvements to QPF in the Bay Area and QPE/QPF in the Sierra Nevada range would be particularly useful to provide better understanding of AR events. [ABSTRACT FROM AUTHOR]

The partitioning of rain and snow during atmospheric river (AR) storms is a critical factor in flood forecasting, water resources planning, and reservoir operations. Forecasts of atmospheric rain–snow levels from December 2016 to March 2017, a period of active AR landfalls, are evaluated using 19 profiling radars in California. Three forecast model products are assessed: a global forecast model downscaled to 3-km grid spacing, 4-km river forecast center operational forecasts, and 50-km global ensemble reforecasts. Model forecasts of the rain–snow level are compared with observations of rain–snow melting-level brightband heights. Models produce mean bias magnitudes of less than 200 m across a range of forecast lead times. Error magnitudes increase with lead time and are similar between models, averaging 342 m for lead times of 24 h or less and growing to 700–800 m for lead times of greater than 144 h. Observed extremes in the rain–snow level are underestimated, particularly for warmer events, and the magnitude of errors increases with rain–snow level. Storms with high rain–snow levels are correlated with larger observed precipitation rates in Sierra Nevada watersheds. Flood risk increases with rain–snow levels, not only because a greater fraction of the watershed receives rain, but also because warmer storms carry greater water vapor and thus can produce heavier precipitation. The uncertainty of flood forecasts grows nonlinearly with the rain–snow level for these reasons as well. High rain–snow level ARs are a major flood hazard in California and are projected to be more prevalent with climate warming. [ABSTRACT FROM AUTHOR]