Statistical post-processing of multiple meteorological elements using the multimodel integration embedded method

Abstract

Statistical post-processing of systematic errors is required for numerical weather predictions to obtain accurate and credible forecasts. Traditionally, this is accomplished separately with different individual models and for one specific element of focus. Here, a promising new method is proposed for the post-processing of meteorological elements output by the European Centre for Medium-Range Weather Forecasts (ECMWF), based on the integration of several different models. For 24-h precipitation, 2-m temperature, and 10-m wind speed, our new method, called the Multimodel Integration Embedded Method (MMIEM), outperformed the single models in terms of several skill scores, while being computationally more convenient. The mean average error of the MMIEM post-processed daily maximum 2-m temperature, minimum 2-m temperature, and maximum 10-m wind speed, was 18%, 26% and 29% lower than that of the raw ECMWF forecast, respectively. Also, compared with ECMWF, the threat score of the rainstorm forecast was improved by 9%. Key attributions to this improvement were the use of multiple models containing different advantages, with help from the embedding process and the interaction of multiple features in the model training. Furthermore, MMIEM can be extended to other statistical and forecast problems. It is anticipated that, with ever-increasing amounts of data, machine learning methods can transform the post-processing of numerical weather forecasts by gaining insight into the importance of different meteorological elements.