【作者】 郑世斌；

【导师】 李斌；

【作者基本信息】 郑州大学 ， 人工智能（专业学位）， 2025， 硕士

【摘要】 东亚地区作为全球气候变化的敏感区,近年来极端高温与降水事件频发,对区域防灾减灾和可持续发展提出了严峻挑战。由区域气候模式(RCM)驱动的CORDEX-EA-II实验为东亚地区提供了未来气候预测,但由于模式自身的不确定性,其输出结果与观测值间仍存在一定偏差,这就要求采用偏差订正技术对RCM模拟数据进行后处理。传统统计方法在处理气象数据的复杂非线性关系及高维时空依赖性时存在局限。近年来,深度学习方法在模式偏差订正领域展现出优势,但其在CORDEX-EA区域数据中的应用仍处于空白。本研究将深度学习应用于东亚地区RCM气温与降水模拟数据的偏差订正,并提出两种改进的深度学习方法,旨在进一步提升该地区气温和降水预测的可靠性。本文的主要工作内容如下:(1)提出一种基于U-Net的深度学习偏差订正方法CE-MS-Unet,用于订正CORDEX-EA-II气温模拟数据的偏差。该方法通过引入多尺度残差块和日历月数据嵌入模块,在增强模型对复杂地形下气温空间特征融合能力的同时考虑了季节性规律,改善了其气温订正效果。订正后结果对比显示,CE-MS-Unet在测试期内有效减少了气温模拟数据的偏差,区域平均MAE和RMSE值分别降低了20.3%和18.2%。订正后气温数据在偏度、峰度的空间分布和月平均气温的季节性变化等方面更接近观测,表明CE-MS-Unet能够较好地捕捉气温的空间和时间特征,有效改善了区域气候模式的气温模拟准确度。(2)为应对CORDEX-EA-II降水数据的强空间变异性和季节性波动等复杂特征,本文在CE-MS-Unet与Cycle-GAN的基础上,提出改进的条件多尺度循环一致性生成对抗网络c MS-Cycle-GAN。通过将简化的CE-MS-Unet作为生成器,并使用带孔空间金字塔池化块(ASPP)优化判别器的性能,增强了模型对复杂降水事件的空间上下文建模能力。订正后结果表明,c MS-Cycle-GAN在改进RCM降水模拟数据整体一致性方面,将区域平均MAE和RMSE值分别降低了23.4%和37.9%。相比传统统计方法LS和QDM,c MS-Cycle-GAN在调整降水空间分布与极端事件频率上优势明显。此外,通过消融实验验证了U-Net结构生成器、日历月数据条件生成模块和ASPP模块对模型整体性能的提升作用,证明了这些优化策略对于提升降水订正效果的价值。更多还原

【Abstract】 As a climate-sensitive region,East Asia has witnessed an increasing frequency of extreme temperature and precipitation events in recent years,posing significant challenges to regional disaster prevention and sustainable development.The CORDEX-EA-II experiment,driven by Regional Climate Models(RCMs),provides future climate projections for East Asia.However,inherent model uncertainties lead to systematic biases between RCM outputs and observational data,necessitating the application of bias correction techniques for post-processing.Traditional statistical methods exhibit limitations in handling the complex nonlinear relationships and high-dimensional spatiotemporal dependencies inherent in meteorological data.Although deep learning approaches have demonstrated promising capabilities in model bias correction,their application to CORDEX-EA regional data remains unexplored.This study pioneers the application of deep learning for bias correction of temperature and precipitation simulations from RCMs in East Asia,proposing two enhanced deep learning methodologies to improve the reliability of temperature and precipitation predictions in this region.The main contributions of this work are as follows:(1)We propose an enhanced deep learning bias correction method,CE-MS-Unet,for correcting temperature biases in CORDEX-EA-II simulations.This approach incorporates multi-scale residual blocks and a calendar month embedding module,enhancing the model’s capability to fuse spatial features of temperature under complex terrain while considering seasonal patterns,thereby improving temperature correction performance.Comparative results demonstrate that CE-MS-Unet effectively reduces temperature biases during the testing period,with regional mean MAE and RMSE values decreasing by 20.3%and 18.2%,respectively.The corrected temperature data show improved consistency with observations in terms of spatial distributions of skewness and kurtosis,as well as seasonal variations in monthly mean temperature,indicating that CE-MS-Unet effectively captures both spatial and temporal characteristics of temperature,significantly enhancing the accuracy of regional climate model simulations.(2)To address the complex characteristics of CORDEX-EA-II precipitation data,including strong spatial variability and seasonal fluctuations,we propose an improved conditional Multi-Scale Cycle-Consistent Generative Adversarial Network(c MS-Cycle-GAN)based on CE-MS-Unet and Cycle-GAN architectures.By employing a simplified CE-MS-Unet as the generator and enhancing the discriminator’s performance with Atrous Spatial Pyramid Pooling(ASPP)blocks,the model’s capability for spatial context modeling of complex precipitation events is significantly improved.The correction results demonstrate that c MS-Cycle-GAN reduces regional mean MAE and RMSE values by 23.4%and 37.9%,respectively,in improving the overall consistency of RCM precipitation simulations.Compared to traditional statistical methods(LS and QDM),c MS-Cycle-GAN shows superior performance in adjusting precipitation spatial distributions and extreme event frequencies.Furthermore,ablation studies confirm the contributions of the U-Net structured generator,calendar month conditional generation module,and ASPP blocks to the model’s overall performance,validating the effectiveness of these optimization strategies in enhancing precipitation correction.更多还原