An Evaluation of Numerical Weather Prediction Models in Forecasting Atmospheric Rivers

Atmospheric rivers (ARs), narrow corridors of high atmospheric water vapor transport, influence large regions of the western coast of North America, from southern California to British Columbia and Alaska. Regardless of location, areas influenced by landfalling ARs face various threats and disruptions from excessive rainfall and associated runoff. Therefore, improving forecasts of AR occurrence and characteristics is of great importance to those responsible for protecting life and property. When providing the public with outlooks and warnings related to ARs, forecasters must confront the challenge of assessing the output of different numerical weather prediction (NWP) models. Specifically, forecasters must understand how performance varies across different time scales, geographical regions, and individual models. Prior work, such as Wick et al. (2013), has examined the forecast skill of several NWP models at different lead times, yet as models are continuously updated, a fresh perspective on AR forecast performance is desired.

This study aims to assess how different weather forecast models perform at varying lead times and for distinct regions of the west coast of North America. Hindcasts from several operational NWP models, obtained from the International S2S Project Database, are run out to approximately 60 days. An atmospheric river detection algorithm is applied to the model output in order to quantify how the models handle such features. The study examines atmospheric river "re-forecasts" for three non-overlapping regions that span a large portion of North America's western coast. The first region covers the Pacific Northwest from southern Oregon to the northern extent of Vancouver Island. The second region extends from southern California to the Oregon border. The third and final region consists of the coasts of British Columbia and southeastern Alaska. Together, these regions represent a large fraction of the AR landfall locations for western North America.

Model performance is studied through the lens of AR occurrence, location, intensity, and geometry. Results indicate three possible sources of variation in model performance. First, differences in forecast skill between individual models is discussed. Second, for individual models, forecast skill varies by season, with performance generally decreased for periods during which atmospheric rivers are climatologically less-favored. Third, forecast performance varies by geographic region. For example, during the NDJF time period, models are more successful with short and medium-range AR forecasts for the Pacific Northwest and California compared to those for British Columbia and Alaska. A desired near-term outcome of this work is an increased awareness of both the utility and limitations of NWP models in the prediction of atmospheric river events at short, medium, and long-range leads. A desired long-term outcome is the use of these results as a bridge to understanding what gives rise to the differing characters of atmospheric rivers over the northeast Pacific and how models can improve their depiction of such features.