深度學習神經網路推估氣候變遷降雨量之研究

Applying Deep Neuro Network to Estimate Future Rainfall under Climate Chamge

葉威里(Wei-Li Ye)

指導教授 ： 林旭信

中原大學/工學院/土木工程研究所/碩士(2020年)

https://doi.org/10.6840/cycu202000734

摘要

摘要
本研究蒐集聯合國氣候變遷專門委員會（Intergovernmental Panel on Climate Change，IPCC）第五次評估報告(AR5)全球氣候模式之ACCESS以及CSMK3模式不同濃度排放情境(rcp45、rcp85)資料，以深度學習類神經網路 (Deep Neural Network ,DNN)作為統計降尺度之架構，推估淡水、臺中、高雄、花蓮未來可能之月降雨。研究中蒐集中央氣象局淡水、臺中、高雄、花蓮1950/01~2005/12之月降雨量資料，分別以原始變數和核主成分分析(Kernel Principal Component Analysis, KPCA)所得之核主成分作為DNN降尺度模式之學習資料，利用歷史情境階段性優化DNN模式參數以建立降尺度模式推估案例區域之月降雨量，並使用多變量線性回歸作為基準模式比較其效能。
使用三分類分析(依照歷史雨量降冪排列之第30%、70%將預測雨量分為偏低、正常、偏高三個分類)、累積機率、相對機率等進行分析，經三分類分析顯示全年雨量中，淡水之雨量主要集中在正常範圍(93.7~201mm)，其他測站則分佈在正常及偏多範圍；四測站夏季之雨量則較為集中在正常區間。經累積機率圖顯示，預測雨量於全年以及冬季時，各區間雨量發生機率較平均，於夏季時，雨量之發生機率則較為集中。經相對機率分析顯示，淡水之冬夏季雨量均在100~200mm的發生機率較高，故乾濕季較不明顯，其餘測站冬夏季發生之雨量均在不同的區間內，故乾濕季較為明顯，如臺中之冬季雨量於0~100mm發生機率較高，而夏季之雨量在200~400mm發生機率較高。由於預測因子與歷史因子之趨勢較為類似，故rcp45與rcp85之模型差別較為有限。

關鍵字

氣候變遷 ； AR5 ； 核主成分分析 ； 機器學習 ； 深度學習

並列摘要

ABSTRACT
This study develops statistical downscaling model which based on Deep Neural Network (DNN) simulating the tendency of future rain fall. The input variables of DNN are used by Original variable and Kernel Principal Component Analysis (KPCA). Our stations are Tamsui, Taichung, Kaohsiung and Hualien. We used general circulation model data ( ACCESS, CSMK3 ) of Intergovernmental Panel on Climate Change (IPCC) fifth assessment report (AR5) as the variables of rainfall prediction model. We collect the monthly rainfall data from the Central Meteorological Bureau of Tamsui, Taichung, Kaohsiung, and Hualien from 1950/01 to 2005/12. We used multivariate linear regression as a base model to compare with the DNN model established in this study.
We used RCP4.5 and RCP8.5 scenarios of AR5 to predict rainfall, and we used three-category analysis (According to the 30% and 70% of the descending power of historical rainfall, the predicted rainfall is divided into three categories: low, normal, and high), mulative probability, relative probability to analyze. The three-category analysis shows that at the annual rainfall, the rainfall of Tamsui is mainly concentrated in the normal range(93.7~201mm), while other stations are distributed in the normal and excessive range; the rainfall of the four stations in summer is more concentrated in the normal range. The cumulative probability chart shows that the forecasted rainfall is more even in the whole year and in winter, Summer rainfall is more concentrated in the same interval. The relative probability analysis shows that the chance of rain of Tamsui in winter and summer is higher in 100~200mm, so the dry and wet seasons are less obvious. The rainfall in the winter and summer at other stations are in different intervals, so the dry and wet seasons are more Obviously. For example, the chance of rain of Taichung in winter is higher in 0~100mm and the rainfall in summer is more likely to occur in 200~400mm. Because the trend of forecast factors and historical factors are similar, RCP45 and RCP85 difference is more limited.

並列關鍵字

Climate change ； AR5 ； Kernel Principal Component Analysis ； machine learning ； Deep Neural Network

參考文獻

Bai, Y. et al. (2016) ‘Daily reservoir inflow forecasting using multiscale deep feature learning with hybrid models’, Journal of Hydrology, 532, pp. 193–206. doi: 10.1016/j.jhydrol.2015.11.011.

Bai, Y. et al. (2017) ‘A multi-pattern deep fusion model for short-term bus passenger flow forecasting’, Applied Soft Computing, 58, pp. 669–680. doi: 10.1016/j.asoc.2017.05.011.

Bengio, Y. (2012) ‘Practical recommendations for gradient-based training of deep architectures’, in Neural networks: Tricks of the trade. Springer, pp. 437–478.

Busseti, E. (2012) ‘Deep learning for time series modeling’, Technical report, Stanford University, pp. 1–5.

Chen, J. (2012) ‘Design of deep belief networks for short-term prediction of drought index using data in the Huaihe river basin’, Mathematical Problems in Engineering, 2012.