結合類神經網路及主成分分析於颱風時期大氣參數對降雨量推估之研究

A Study of Atmospheric Parameters for Rainfall Estimation during Typhoon Period Using Neural Networks and Principal Component Analysis

林子傑(Tzu-Chieh Lin)

指導教授 ： 張麗秋

淡江大學/工學院/水資源及環境工程學系碩士班/碩士(2022年)

https://doi.org/10.6846/TKU.2022.00703

摘要

台灣坐落在西北太平洋上，為熱帶氣旋與颱風侵襲的主要路徑，平均每年有四到五個颱風侵襲台；同時台灣山高地狹、地形陡峭、川短流急，使得颱風所帶來的豐沛雨量引發水庫排洪不及的危機，而準確的降雨預報可提高流量推估之準確性，有助於水庫的防洪操作策略之參考，可提前預放與調節水庫水位，預留足夠的防洪空間，此為值得探討且重要的議題。
本研究以石門水庫集水區最為研究區域，透過蒐集颱風時期集水區測站之歷史資料與ERA5大氣參數之網格資料，建置倒傳遞神經網路模式(BPNN)以預測未來1~3小時之集水區降雨量，並依照不同輸入項與降雨量之移動平均之結合可分三種模式，模式一(篩選大氣參數)、模式二(篩選參數之前十個主成分)與模式三(篩選參數之前五個主成分)，以分析輸入因子對BPNN模式預測結果之影響，並討論大氣參數與降雨量之關係。
根據結果顯示，以篩選參數作為輸入項之BPNN模式大致上能掌握降雨趨勢，說明本研究所篩選之大氣參數若颱風時期能取得即時觀測資料，能作為推估未來時雨量之參考依據；模式二與模式三之結果表現均優於模式一，可證明經由主成分分析保留重要特徵的降維方式，能提高模式之預測準確度及運算效率。

關鍵字

類神經網路 ； 倒傳遞類神經網路 ； 主成分分析 ； 大氣參數 ； 時雨量預報

並列摘要

Taiwan is located in the northwestern Pacific Ocean and is the main route of tropical cyclones and typhoons. On average, four to five typhoons strike Taiwan every year, bringing torrential rain and flooding. Due to the mountainous and steep terrain in Taiwan, it makes abundant rainfall a typhoon brings rapidly flow into rivers and leads to a catastrophe within hours if the flood discharge of the reservoir do not be well controlled. With an accurate rainfall forecast, we can improve the accuracy of inflow estimation, which is crucial for referencing flood control operation strategy of the reservoir.
In this study, we chose Shimen watershed as study area, collecting historical station data of Shimen watershed and ERA5 data during typhoon period so that a back-propagation neural network(BPNN) could be built to predict the rainfall within the watershed 1 to 3 hours ahead. On the basis of the combination of different input and the moving average of rainfall, we divided it into three models, Model 1 (screening atmospheric parameters) , Model 2 (ten principal components before the screening parameters) and Model 3 (five principal components before the screening parameters). Finally, the influence of each input factors were analyzed through the results of the BPNN model, and discuss the relationship between atmospheric parameters and rainfall.
The finding indicates that the BPNN model with the screening parameters as input can roughly grasp the rainfall trend, and that the screening atmospheric parameters can be used as a feature for estimating future rainfall if real-time observation data can be obtained during the typhoon period. The results of model 2 and 3 work better than model one, which proves that the dimensionality reduction method, which retains important features through principal component analysis, can indeed improve the prediction accuracy and computational efficiency of the model.

並列關鍵字

Artificial Neural Networks(ANN) ； Back Propagation Neural Network(BPNN) ； Principal Component Analysis(PCA) ； Atmospheric Parameters ； Hourly Rainfall Forecast

參考文獻

1. Chang, L. C., Chang, F. J., Yang, S. N., Tsai, F. H., Chang, T. H., Herricks, E. E. 2020. Self-organizing maps of typhoon tracks allow for flood forecasts up to two days in advance. Nature communications, 11(1), 1-13.

2. Chang, L. C., Shen, H. Y., Chang, F. J. 2014. Regional flood inundation nowcast using hybrid SOM and dynamic neural networks. Journal of Hydrology, 519, 476-489.

3. Chang, L. C., Shen, H. Y., Wang, Y. F., Huang, J. Y., Lin, Y. T. 2010. Clustering-based hybrid inundation model for forecasting flood inundation depths. Journal of hydrology, 385(1-4), 257-268.

4. Chang, L. C., Chang, F. J. 2001. Intelligent control for modelling of real‐time reservoir operation. Hydrological processes, 15(9), 1621-1634.

5. Chen, N., Xiong, C., Du, W., Wang, C., Lin, X., Chen, Z. 2019. An improved genetic algorithm coupling a back-propagation neural network model (IGA-BPNN) for water-level predictions. Water, 11(9), 1795.