Enhancing the Skill of Medium Range Forecasts with a Machine Learning Based Multi-Model Super-Ensemble (MMSE)

In recent years, medium-range AI weather forecasting models have improved significantly, now offering forecasting accuracy comparable to classical numerical weather prediction (NWP) models, while also being faster and (once trained) less computationally demanding.

Due to inherent assumptions and limitations, all weather prediction models exhibit some degree of persistent systematic errors, also called biases, in their forecast output, with certain models performing better than others for specific variables and regions.

To address these persistent biases, we introduce a machine learning-based multi-model super-ensemble (MMSE), which collectively reduces model biases by combining the complementary strengths of each model. The MMSE assigns optimized weights to each model's forecast based on its historical performance to leverage each model’s strengths under different conditions (both spatial and temporal) rather than equally weighting models as in a simple ensemble mean.

In this work, we developed two regional MMSE models tailored to specific regions, seasons, and variables of interest. One model targets 2-meter air temperature and 10-meter wind components in Germany’s winter season, while the other targets Indian summer monsoon rainfall.

We trained the MMSE using an Extreme Gradient Boosting framework (XGBoost) to capture spatiotemporal features more effectively. The training data consisted of past forecasts from multiple AI models (FourCastNet, Pangu-Weather, GraphCast) and relevant climatology and topology data. ERA5 reanalysis served as the ground truth. The details of MMSE development will be presented.

Our MMSE developed for 2-meter temperature over Germany showed approximately a one-day improvement in forecast gain time compared to the best-performing individual model. In other words, the MMSE’s 11th-day forecast matched the accuracy of the 10th-day forecast from the best-performing model, effectively adding an extra day of reliable lead time. These findings suggest that the proposed MMSE offers a promising, computationally efficient alternative to traditional ensembles for real-time weather forecasting, with potential applications in domains requiring high-precision predictions. With a view to make these results interpretable and to identify the relative strengths of participating models, we will also present the analysis of SHAP values for various variables and regions.