運用類神經網路探討颱風路徑對集水區雨量空間分布影響之研究

Investigating the effect of typhoon track on rainfall spatial distribution in a watershed using artificial neural networks

黃克禮(Gary Wee)

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(2019年)

https://doi.org/10.6846/TKU.2019.00944

摘要

台灣坐落於西北太平洋颱風的主要路徑上，一年內台灣受到颱風影響數次，颱風除了帶來豐沛的水資源，也可能引發嚴重的災害。台灣的山地連綿起伏，造成洪水在颱風期間移動地更快，對水庫或河川下游地帶的衝擊大。水庫在台灣是非常有效的防洪設施，然而在防災期間，水庫操作面臨最大問題是颱風降雨預報的不準確性。台灣的颱風降雨預報之所以不容易預測準確，是因為預報員對颱風的降雨機制仍理解不足；除了颱風結構複雜，另外颱風也受到地形因素影響，其降雨機制不易被歸納或推測。
了解颱風的降雨情況，有助於水資源與防災系統的管理與規劃。目前已經發展的颱風降雨預報模式有幾種，包括了數值模式 (Numerical Weather Prediction, NWP)，劇烈天氣監測系統 (Quantitative Precipitation Estimation and Segregation Using Multiple Sensors, QPESUMS)，系集模式颱風定量降水預報(Ensemble Typhoon Quantitative Precipitation Forecast, ETQPF)。過去的學者多在研究颱風內部的降雨分布，或者颱風於大面積的降雨情形，像是颱風為整個台灣所帶來的降雨量，卻鮮少有研究在討論集水區的颱風降雨情形。另外，雖然已經有研究利用類神經網路 (Artificial Neural Networks, ANNs) 分析與預測一般的降雨，但是颱風降雨的分析與預報依舊多使用數值方法或統計方法。
本論文利用前饋式、非監督式、競爭式的類神經網路——自組特徵映射類神經網路 (Self-organizing map, SOM) 來分析颱風期間石門水庫集水區的降雨時空分布。研究發現路徑相似的颱風，其降雨的空間變化也會相似。因此，颱風降雨的空間分布與颱風路徑有高度相關性；倘若有兩場路徑不同之颱風中心位於同一個經緯度，其降雨空間分布不一定相同。
由於颱風路徑會影響集水區降雨之時空分布，相似路徑之颱風對同一集水區降雨之時空分布影響相似，所形成雨型也較為相似。因此，本研究使用特徵雨量組體圖 (Feature hydrographs of rainfall, FHR) 來描述路徑類似之颱風所帶來的雨型。每場颱風歷經時間不同、強度不同，造成降雨強度與總雨量亦不相同，故須進行正規化，即將歷經時間換成 0~1 之間的數值，將時雨量轉換成百分比，再將該神經元內所有正規化雨量歷程進行平均，即為特徵雨量組體圖。研究結果顯示，特徵雨量組體圖若呈中央集中型，且峰值較高與較寬，其颱風降雨的破壞性較高。反之，分布相對平均的特徵雨量組體圖，一般表示颱風遠離石門水庫集水區，颱風對水庫的影響較低。一旦中央氣象局發布颱風預報，本研究可根據預報的颱風路徑，利用特徵雨量組體圖來估算該場颱風的雨量組體圖，提供防災機構與水庫操作單位非常有利的資訊。
本研究分析在歷史颱風場次中ETQPF於石門水庫的降雨預報誤差；若颱風距離石門水庫較近，ETQPF的準確度較高。相反的，對於距離石門水庫較遠的颱風，其ETQPF的預報誤差高，這可能是因為颱風降雨機制的不確定性大，發生於石門水庫的降雨主要受到颱風外圍環流的影響，而非颱風主要結構。為了改善颱風降雨預報，本研究篩選出表現優良的系集預報成員，并利用加權平均法重新計算颱風預報降雨量。系集預報成員的權重是比較颱風的預報路徑與實際路徑之決定系數(R2)，倘若某成員的颱風路徑預報越準確，其權重值越高。本研究發現選擇可靠度高之系集預報成員，ETQPF之降雨預報改善率可達最高90%。

關鍵字

類神經網路 ； 自組特徵映射網路 ； 颱風路徑 ； 降雨空間分佈 ； 降雨預報

並列摘要

Typhoons hit Taiwan several times every year, which could cause serious flood disasters. Because mountainous terrains and steep landforms can rapidly accelerate the speed of flood during typhoon events, rivers cannot be a stable source of water supply. Reservoirs become the most effective floodwater storage facilities for alleviating flood damages in Taiwan. Forecasting typhoon rainfall is a long-standing and challenging issue due to the complexity of typhoon.
This study focused on typhoon rainfall including the spatial distribution of rainfall, the temporal distribution of rainfall (hyetographs) and the total rainfall. For the spatial distribution of rainfall, self-organizing map (SOM) is implemented to explore the spatial distribution of typhoon rainfall. The spatial distribution of rainfall has similar change processes for the similar tracks of the typhoons. Rainfall spatial distribution is highly related to the classification of the typhoon track; if two typhoons have different track, their rainfall spatial distributions would be probably different, even they are in the same location. The feature hyetographs of rainfall (FHR) in the same classification of typhoon tracks are constructed. For the classification of typhoon tracks with great impact on the Shihmen watershed, the rainfall hyetographs are central distributed with a higher and wider peak that reveals the destructivity of the typhoon rainfall. In contrast, the relatively flat distribution of rainfall hyetograph means the typhoon is far away from the watershed and less harmful. FHR can be used to estimate the rainfall hyetographs based on the forecast typhoon tracks, this is highly valuable to the emergency response agencies and reservoir operation units. The forecast error of ETQPF in the Shihmen watershed is investigated for improving the total rainfall forecast. If the locations of the typhoon are near the watershed, the ETQPF has higher accuracy. On the other hand, typhoons distancing from the watershed, usually have higher uncertainty in the factors of typhoon rainfall because of the effect of the typhoon outer circulation flow, instead of the typhoon main structure, resulting in lower accurate ETQPF forecast. This study proposes the weighted average method to recalculate the typhoon rainfall by selecting QPFs from well-performed ensemble members. The more reliable the ensemble members are, the higher the values their weights are. The weights of the ensemble members, based on the determination coefficient between the forecast typhoon tracks of ensemble members and the actual tracks, are given to recalculate the total rainfall. When highly reliable ensemble members are chosen, the improvement rate of rainfall prediction could be more than 90% compared to the ETQPF.

並列關鍵字

Artificial neural network (ANN) ； Self-organizing map (SOM) ； Typhoon Track ； Rainfall Spatial Distribution ； Rainfall Forecast

參考文獻

1. Burpee, R.W., and Black, M.L. (1989). Temporal and Spatial Variations of Rainfall Near the Centers of Two Tropical Cyclones. Monthly Weather Review 117, 2204-2218.

2. Central Weather Bureau (2016). Typhoon Knowledge. Retrieved from https://www.cwb.gov.tw/V7e/knowledge/encyclopedia/ty017.htm

3. Central Weather Bureau (2018). Meteorological Education and Training. Retrieved from https://metet.cwb.gov.tw/cwb/

4. Chang, C.P., Yeh, T.C., and Chen, J.M. (1993). Effects of Terrain on the Surface Structure of Typhoons over Taiwan. Monthly Weather Review 121, 734-752.

5. Chang, F.-J., Chang, K.-Y., and Chang, L.-C. (2008). Counterpropagation fuzzy-neural network for city flood control system. Journal of Hydrology 358, 24-34.