Downscaling of Precipitation for Climate Change Projections Using Multiple Machine Learning Techniques: Case Study of Shenzhen City, China

To examine the characteristics of future precipitation under climate change is of great significance to urban water security. In this paper, multiple machine learning techniques, i.e., statistical downscaling model (SDSM), support vector machine (SVM), and multilayer perceptron (MLP), were used to downscale large-scale climatic variables simulated by the General Circulation Models (GCMs) to precipitation on a local scale. It was demonstrated in Shenzhen city, China, through multisite downscaling schemes based on projections from the Max Planck Institute Earth System Model (MPI-ESM1.2-HR), Meteorological Research Institute Earth System Model Version 2.0 (MRI-ESM2.0), and Beijing Climate Center Climate System Model (BCC-CSM2-MR). The obtained results showed that the downscaled precipitation would provide good monthly simulations against observations at 10 discrete stations. Regardless of superior performance of SVM and MLP over SDSM, the daily precipitation simulations should be further improved, and downscaling of heavy daily precipitations would be promoted by quantile mapping corrections. Due to the relatively poor simulation performance of BCC-CSM2-MR, the other two climate models were considered under the Shared Socioeconomic Pathways (SSP1-2.6, SSP2-4.5, and SSP5-8.5 scenarios) for ensemble precipitation projections for 2015-2100. Under the SSP1-2.6 scenario, the amounts of annual average precipitation for 10 stations were estimated to be higher relative to the historical period (2.7%-17%), and 9 out of 10 stations presented an increasing trend. However, downward trends also existed at three stations when it comes to scenarios SSP2-4.5 and SSP5-8.5. Moreover, a significantly positive trend was found to dominate the trend changes of annual extreme daily precipitation during 2015-2050, but the detected trends at stations were greatly dependent on the downscaling techniques and climate models. Besides, the increase in daily extreme precipitations for various return periods as well as statistically different precipitation characteristics for discrete stations would further shed light on urgent demands on urban resilient strategies for climate change adaptation.

Automated summary

This study used multiple machine learning techniques to downscale future precipitation and compared the performance of different downscaling and climate models. The downscaled precipitation showed good agreement with observations, and under the SSP1-2.6 scenario, there was an increase in precipitation, while a decreasing trend was observed under the SSP2-4.5 and SSP5-8.5 scenarios.