De praeceptis ferendis: Air Quality Multi-model Ensembles

Ensembles of air quality models have been shown to outperform single models in many cases. Starting from the theoretical evidence behind this empirical ascertainment, we present the conditions granting an ensemble superior to any single model. As those conditions are not systematically met, we also investigate two additional ensemble estimators for which a sound mathematical framework exists. In view of producing a single improved forecast out of the ensemble, the three candidate ensemble estimators, namely the unconditional ensemble mean, the weighted ensemble mean and the mean of the sub-ensemble with the right trade-off between accuracy and diversity, are evaluated against data generated in the context of AQMEII (Air Quality Model Evaluation International Initiative). The pitfalls of training such ensembles are investigated. Overall, following a proper training procedure, the sophisticated ensemble averaging techniques were shown to have higher skill compared to solely ensemble averaging forecasts.