An Evaluation of Numerical Weather Prediction Model Forecasts of Sensible Weather Parameters

Predictions of sensible weather elements are arguably the most valuable and challenging components of a weather forecast. The presence or absence of precipitation and its type and amount in forecasts routinely drives critical decisions by businesses and individual private consumers of these forecasts as disseminated on various media platforms. Extensive information regarding the comparative skill of precipitation forecasts from various global end-forecast providers exists and is widely available. ForecastWatch, a well-known independent forecast verification provider, has published verification of precipitation forecasts from a multitude of forecast providers for nearly fifteen years.

Today, many forecast providers, including The Weather Company, use some variation of an automated consensus forecasting system to drive their operational forecasts. Consensus weather forecasting systems consider multiple inputs, in particular numerical weather prediction (NWP) forecasts, which are blended to provide an objective forecast that outperforms any of the ingredient forecasts. For continuous parameters such as temperature and wind, forecasts are compared to observations and a machine learning approach is applied to optimally combine the ingredient forecasts and minimize error. However, in the case of sensible weather or precipitation forecasts, the blending is often less objective owing in part due to the inherent difficulty in defining robust optimization methods and criteria for these types of parameters. For example, at The Weather Company, the blending of multiple model inputs is generally empirical and not necessarily optimized based on ingredient model skill.

Sensible weather forecast verification information available in the literature to date focuses on the final, or end user forecast. In the context of the consensus forecasting system, this generally means the blended forecast where the component, or input model forecast data, has been abstracted via the blending process.

Yet, there is value in understanding the skill of sensible weather predictions from NWP forecasts. This information can be prove crucial in making informed decisions around which NWP inputs should be included in a consensus forecast system. Further, it can provide insight into the appropriate weights to apply to each model on a regional or site-by-site basis to optimize the skill of the sensible weather forecasts from a consensus system. In this study, we will derive forecasts of sensible weather from NWP forecasts and verify them against ground-based observations. Results of this verification comparison will be shared, highlighting geographical and temporal skill characteristics and the many challenges in verification of these types of parameters.