Multi-model Ensembles with Spatial and Model Dependencies

The aggregation of weather forecasts in multi-model ensembles is a useful approach to enhance forecast accuracy and better understand forecast variability. The component models in a multi-model ensemble may exhibit correlations between models and/or locations being forecast, due to common data sources and approaches. Detecting and understanding these correlations will be critical to avoid overconfidence and biases in the multi-model forecast. With multiple forecast points from the same ensemble of models there is more information available to assess the presence and impacts of correlations between the forecasts, both between models and locations being forecast. Traditional spatial correlation models estimate correlations from observed values. In the case of ensemble forecasts the observed values are not precisely known apriori, but instead, there is a collection of forecasts for these points. A likelihood model is proposed to simultaneously estimate model and spatial correlations in a multi-model ensemble. Bayesian techniques are used to estimate the model for a multi-model ensemble of weather forecasts. The impacts on forecast accuracy and calibration of this approach are compared to more traditional ensemble aggregation approaches.