計算集水區平均降雨與實際監測資料之分析比較－以翡翠水庫集水區為例

Comparative Analysis of Calculated Average Rainfall and Actual Monitoring Data in Watersheds: A Case Study of the Feitsui Reservoir Watershed

詹翔屹(Shing-Yi Jhan)；陳智誠(Chihcheng Chen)；鄭清江(Ching-Jiang Jeng)

流域平均雨量 ； 水文資料分析 ； 機器學習 ； 翡翠水庫 ； 洪峰流量 ； Watershed average rainfall ； Hydrological data analysis ； Machine learning in hydrology ； Feitsui Reservoir ； Flood peak

水保技術

第18卷第1、2期（2024 / 06 / 01）

8 - 21

繁體中文;英文

流域平均雨量是河川流量、水庫蓄水量、洪水頻率強度及歷線繪製等水文資料最重要依據之一，一般流域內會設置多個雨量站，而計算集水區平均雨量廣泛使用包含算術平均數法（Arithmetic Averaging Method）、徐昇式多邊形法（Thiessen Polygons Method）、高度平衡多邊形法（Height Balance Polygons Method）等數學方法作計算，本文利用翡翠水庫集水區雨量站實際監測數據，透過機器學習數據分析方法，選擇隨機森林迴歸特徵選取法（Random Forest Regressor）、置換特徵重要性選取法（Permutaion Feature Importance）與算術平均數法（Arithmetic Averaging Method）、高度平衡多邊形法（Height Balance Polygons Method）等方法作計算比較，計算結果不論是徐昇多邊形或是高度平衡多邊形或是算術平均數法都與實際監測數據計算之隨機森林迴歸特徵選取法及置換特徵重要性選取法權重有明顯差異。以碧湖測站為例，權重差異高達9%以上，權重佔比高的測站也不同，顯示純粹用數學方式計算與實際監測數據透過機器學習的數據分析結果，直接影響後續洪峰演算結果，建議後續蒐集各水庫集水區實際監測數據透過機器學習或深度學習的數據分析與演算，建立數值模型供集水區各項分析使用。

The average rainfall over a watershed is one of the most important bases for hydrological data such as river flow, reservoir storage, flood frequency intensity, and hydrograph plotting. Generally, multiple rainfall stations are set up within a watershed, and several mathematical methods are widely used to calculate the average rainfall of the catchment area, including the Arithmetic Averaging Method, the Thiessen Polygons Method, and the Height Balance Polygons Method. This paper utilizes actual monitoring data from the rainfall stations in the Feitsui Reservoir catchment area and employs data analysis methods through machine learning. Specifically, we compare the results of using the Random Forest Regressor, Permutation Feature Importance, Arithmetic Averaging Method, and Height Balance Polygons Method. The calculation results indicate significant differences in the weights derived from the Thiessen Polygons Method, Height Balance Polygons Method, and Arithmetic Averaging Method compared to those derived from the Random Forest Regressor and Permutation Feature Importance based on actual monitoring data. For instance, the weight difference for the Bihu Station is more than 9%, and the stations with high weight ratios also differ. This demonstrates that purely mathematical calculations and data analysis results through machine learning based on actual monitoring data directly affect subsequent flood peak calculations. Therefore, it is recommended to collect actual monitoring data from various reservoir catchment areas and use data analysis and calculation through machine learning or deep learning to establish numerical models for various analyses within the catchment areas.