Application of Artificial Intelligence Technology to Numerical Weather Prediction
Sun Jian,Cao Zhuo,Li Heng,Qian Simeng,Wang Xin,Yan Limin,Xue Wei
DOI: https://doi.org/10.11898/1001-7313.20210101
2021-01-01
Abstract:Numerical weather prediction technology plays an increasingly important role in improving accuracy and service level of modern weather forecast. With the development of observation system and higher resolution and complexity of the numerical weather prediction model, the products of numerical weather forecast have been greatly improved in quantity and quality, and can offer rich information at high spatial-temporal frequency. However, such a large amount of prediction data are not fully explored. Artificial intelligence has achieved great success in many fields, such as pattern recognition and natural language processing, which provides an opportunity for further improving numerical weather prediction. It's also employed in initialization, numerical model and production of weather forecast service, involving observation system, data assimilation, model integration, ensemble forecast and high-performance computing methods. Both the accuracy of forecast results and computational efficiency have been improved by using error correction, parameter estimation, local surrogate model and so on. In addition, some end-to-end neural network models also show the potential of pure data-driven weather forecast. These models use spatial-temporal observation data as input and directly output the prediction results in terms of deterministic results or probabilities. Some of them perform well in short-term severe convective weather, precipitation, and long-term climate forecast. Existing works employ various artificial intelligence technology methods, mainly including large-scale calculation of neural network, feature analysis, interpretability, and customized loss function. However, there are still some challenges, the potential of artificial intelligence needs to be further explored. Some issues should be carefully considered, including weak interpretability, uncertainty analysis and the coupling with conventional numerical models, and how to use physical knowledge to guide the design of artificial intelligence model is also worth addressing. To deal with these challenges, some promising suggestions are proposed. Bayesian network and causal network will help to establish more comprehensive and profound feature engineering. Using Bayesian inference to generate distribution characteristics of current meteorological states may be an alternative to efficient and effective uncertainty quantification. The development of some standard workflow and framework will contribute to the coupling of conventional numerical model and artificial intelligence module. Successful artificial intelligence applications in weather forecast require deep cooperation between meteorological experts and computer experts who focus on artificial intelligence and high-performance computing.