應用自主性演算法與適應性模糊推論系統評估未來降雨趨勢

Applying Group Method of Data Handling and Adaptive Network-Based Fuzzy Inference System to Evaluate the Tendency of Future Rainfall

洪哲縺(Che-Lien Hung)

指導教授 ： 林旭信

中原大學/工學院/土木工程研究所/碩士(2015年)

https://doi.org/10.6840/cycu201500920

摘要

本研究以適應性網路模糊推論系統(Adaptive Network-Based Fuzzy Inference System，ANFIS)與自組性演算法(Group Method of Data Handling，GMDH)發展ASDM(ANFIS Statistical Downscaling Model)與GSDM(GMDH Statistical Downscaling Model)統計降尺度模式模擬探討淡水、台中、高雄及花蓮未來降雨趨勢。利用主成份分析(Principal Component Analysis，PCA)對全球環流模式(General Circulation Model，GCM)四種模式歷史情境分析資料決定ASDM模式輸入變數；以GMDH啟發式自組架構(Heuristic Self-Organization)之特性，自動優選較佳之GSDM模式輸入變數。利用四種GCM及淡水、台中、高雄及花蓮測站歷史氣象紀錄，評估未來情境短期(2020年~2039年)、中期(2050年~2069年)及長期(2080年~2099年)之降雨趨勢。 　
　　模擬結果顯示，CGMR模式主成份分析結果選取主成份1至3變異比例達85.925%，CSMK3模式選取主成份1至4變異比例達84.544%，GFCM2.1模式選取主成份1至5變異比例達83.763%，MRCGCM模式選取主成份1至6變異比例達86.156%，有效地減少模式運算時間以及降低模式複雜度；整體而言，模式效能以GSDM模式優於ASDM模式，四個測站中高雄測站GSDM模式以CGMR模式歷史情境20C3M相關係數0.72為最高。利用已建立ASDM與GSDM模式代入未來情境A1B預測未來雨量趨勢，以高雄測站CGMR-GSDM結果為最佳，短、中及長期相關係數為0.68、0.61及0.57，平均降雨量變化率為-24.61%、-27.81%及-28.5%；冬夏兩季以淡水測站結果為最佳，夏季短、中及長期平均降雨量變化率為30.2%、26.71%及18.26%，冬季短、中、長期之平均降雨變化率為40.46%、54.99%及55.44%。

關鍵字

統計降尺度；主成份分析；全球環流模式；適應性網路模糊推論系統；自組性網路架構

並列摘要

This study develops ASDM, statistical downscaling model based on adaptive network fuzzy inference system (ANFIS), and GSDM, statistical downscaling model based on Group Method of Data Handling (GMDH) to simulate the tendency of future rainfall in Tamsui, Taichung, Kaohsiung, and Hualien, respectively. The input variables of ASDM are determined by using principal component analysis (PCA) for four historical scenario data of global circulation models (GCM); by applying the characteristics of heuristic self-organization of GMDH, the optimal input variables of GSDM are selected automatically as well. The future tendency of rainfall in short term (2020-2039), medium term (2050-2069) and long term (2080-2099) are assessed by applying four GCMs and historical data in Tamsui, Taichung, Kaohsiung, and Hualien station.
The simulated results show that the variation percentage of CGMR is 85.925% by selecting 3 principal components (PC); that of CSMK3 is 84.544% by selecting 4 PC; that of GFCM2.1 is 83.763% by selecting 5 PC; that of MRCGCM is 86.156% by selecting 6 PC. The computation time and complexity of the models can be reduced. For all the models, the performance of GSDM is better than ASDM. The correlation coefficient (CC) of GSDM, which is the best between the models, based on historical scenario 20C3M of CGMR is about 0.72 in Kaohsiung. It also show that CC of rainfall in Kaohsiung are 0.68, 0.61 and 0.57 and the percentage change of average are -24.61%, -27.81% and -28.5% in the short, medium and long-term, respectively. The better percentage changes of average rainfall in Tamsui are 30.2%, 26.71% and 18.26% in summer season and 40.46%, 54.99% and 55.44% in winter season.

並列關鍵字

Statistical downscaling ； Principal component analysis ； General circulation model ； Adaptive network fuzzy inference system ； Group method of data handling

參考文獻

高子昂，2013，「資料處理群集分析演算法應用於颱風暴潮偏差之預測---以麥寮潮位站為例」，碩士論文，國立成功大學。

胡鈞甯，2013，「非線性主成分分析結合神經網路之氣候變遷統計降尺度模式」，碩士論文，中原大學土木工程學系研究所，中壢。

莊家閔，2012，「氣候變遷統計降尺度不確定性分析之研究」，碩士論文，中原大學土木工程學系研究所，中壢。

甘秉玄，2009，「結合聚類分析與人工智慧於颱風時雨量即時預測」，碩士論文，中原大學土木工程學系研究所，中壢。

陳立偉，2008，「氣候變遷對水資源之衝擊評估-以牡丹水庫集水區為例」，碩士論文，中原大學土木工程學系研究所，中壢。