The societal benefit of accurate operational prediction of atmospheric rivers (ARs) is well established, especially for the western United States. A customized high-resolution AR prediction system (West-WRF) has been developed for the US West Coast, complete with a 32 water-year reforecast. The objective of this work is to analyze errors in the reforecast dataset with the goal of identifying synoptic-scale patterns associated with higher or lower predictive skill. This is a first step in documenting model error sources, and ultimately addressing forecast deficiencies. Using self-organizing maps (SOMs), we sort thousands of AR forecasts at two different forecast lead times (72 and 144 hours) and classify them based on different training variables, creating two-dimensional matrices of synoptic-scale patterns. We then compute several metrics of forecast skill for each SOM node to identify patterns of higher and lower skill. Initial efforts to train SOMs only using integrated vapor transport (IVT) failed to show strong skill differences between nodes. However, SOMs trained on IVT differences between forecasts and reanalyses (European Center Reanalysis Version 5, ERA5) elucidate different meteorological patterns of common forecast errors. Primary IVT error patterns are related to errors in AR translation speed or intensity. These error patterns are consistent across full datasets and subsets of higher and lower skill forecasts as determined by the Model Evaluation Toolkit (MET) Method for Object Based Diagnostic Evaluation (MODE) calculated statistics per SOM node. Furthermore, analyses subset by ENSO and MJO phase illustrate a slight seasonal bias toward some of the modes.

Meteorological fields associated with translation speed and intensity biases are also analyzed, with representative cases examined individually. More of the 144-hour cases are associated with initialization or western boundary errors, whereas the 72-hour cases more often appear related to physical process-based errors. However, even the 144-hour case studies delineated some important under-forecasting of IVT. Understanding the low-intensity errors will help us identify choices for convective or microphysics schemes, as well as the importance of resolving some mesoscale features such as frontal waves, when forecasting atmospheric rivers.