Assessment and prediction of regional climate based on a multimodel ensemble machine learning method

• Published in: Climate dynamics Vol. 61; no. 9-10; pp. 4139 - 4158
• Main Authors: Fu, Yinghao, Zhuang, Haoran, Shen, Xiaojing, Li, Wangcheng
• Format: Journal Article
• Language: English
• Published: Berlin/Heidelberg Springer Berlin Heidelberg 01.11.2023; Springer Nature B.V
• Subjects: Algorithms; Climate change; Climate models; Climate prediction; Climatology; Earth and Environmental Science; Earth Sciences; Geophysics/Geodesy; Global climate; Global climate models; Learning algorithms; Machine learning; Maximum temperatures; Minimum temperatures; Modelling; Oceanography; Precipitation; Regional climates; Solar radiation; Stacking; Systematic errors; Model fusion; CMIP6; Machine learning; Spatial downscaling
• ISSN: 0930-7575; 1432-0894
• DOI: 10.1007/s00382-023-06787-7

Accurate modeling of climate change at local scales is critical for climate applications. This study proposes a regional downscaling model (stacking-MME) based on the fusion of multiple machine learning models (stacking). The performance of the model was evaluated for simulating precipitation, solar radiation, maximum temperature and minimum temperature and predicted three future possible changes in climate variables over time (near-term (2031–2040), medium-term (2051–2060), and long-term (2081–2090)). After determining the optimal GCM(Global climate model) based on rating metric calculations, the parametric and structural uncertainties in the GCM simulation of CMIP6 (Sixth International Coupling Model Comparison Project) were reduced. Furthermore, the performance of MME (multimodel ensembles) was enhanced by integrating three machine learning algorithms. The results show that among the nine machine learning models, the Light Gradient Boosting Machine, Gradient Boosting Regressor and Random Forest have the best performances. These three models are also considered for the development of stacking model fusion. The Stacking-MME model can reliably reduce the systematic error of GCMs and has the potential to better predict climate. In the SSP245 and SSP585 situations, precipitation will increase by 23.79% and 29.26% at the end of the twenty-first century, respectively. Maximum and minimum temperatures will increase by 1.48 and 2.89 °C and 1.22 and 2.36 °C, respectively.