Automatic model calibration to improve the performance of numerical weather prediction models

Automatic calibration of a numerical weather prediction (NWP) model refers to the process in which NWP model parameters are tuned using mathematical optimization methods to minimize the difference between model predictions and corresponding observations. This approach is not widely practiced so far because of difficulties related to model complexities such as high-dimensionalities of model parameters and model outputs, and the extraordinary demand of computational resources. This paper presents a platform specifically designed to tackle those difficulties, called Uncertainty Quantification Python Laboratory (UQ-PyL). The key functions contained in UQ-PyL include design of experiment (DoE), uncertainty analysis, global sensitivity analysis, surrogate modeling, and multi-objective optimization. We intend to demonstrate how UQ-PyL can be used to improve the predictive skill of the NWP model with a case study involving 5-day weather forecasting in the Greater Beijing region using the WRF model. Through numerous calibration and validation experiments, we found that automatic model calibration can improve predictive skill of the WRF model significantly.