Investigate typhoon paths by self-organizing maps and predict long-term flow during typhoon periods

臺灣地處颱風路徑常經之地，每年夏秋兩季有遭受颱風侵襲與威脅之虞，不同颱風路徑造成各地災情況不同，如2004年艾利、2007柯羅莎、2009年莫拉克、2012蘇拉與2013蘇力颱風分別造成北、中、南不同地區之災情，所挾帶的豪大雨量及高降雨強度為水庫操作人員帶來極大的挑戰。以臺灣地區集水區小、坡陡流急之地文條件下，發展預測河川流量或水庫入流量模式多以數小時預報之短時距為主，可提供防災減洪參考之應變時間較短；若能在颱風來臨前，預測颱風全洪程流量歷線將有助於水庫在颱風來臨前之調節性放水策略與即時防洪操作之決策參考。 本研究以分析颱風路徑分類與水庫入流量型態之關係，提出預測颱風時期水庫全洪程入流量歷線之方法論；主要是應用自組特徵映射網路(SOM)進行颱風路徑分類，並依不同颱風路徑分類定義其流量特徵曲線，即可從集水區總降雨量預報資料，套配出全洪程流量歷線。本模式分為颱風路徑分類與水庫入流量套配兩階段步驟；在颱風路徑分類階段，將颱風路徑之經緯坐標轉換為網格向量，以SOM將颱風路徑網格向量進行分類，產生颱風路徑分類之拓樸圖，進而製作拓樸圖中各神經元之流量特徵曲線及統計資料；在流量預測階段，以測試颱風事件進行流量預測，將颱風路徑進行分類，進而以颱風被分配至某神經元的流量特徵曲線、逕流係數及氣象局預報的總降雨量，即可進行颱風全洪程流量預測。 結果顯示，路徑拓樸圖中可看出各神經元皆表現出有別於其他神經元的颱風路徑，而神經元與鄰近神經元間之拓樸關係，亦能些微分類出颱風路徑之差異，且各神經元之颱風路徑及其洪峰位置相當接近，而大部分神經元中所屬之流量特徵曲線相似度極高，僅少數神經元有較大差異；全洪程流量預測模式能夠提前掌握其流量趨勢，並準確預測洪峰時間，洪峰量及總入流量亦在接受範圍內，推估之流量歷程與實際流量歷程相關係數極高。Taiwan is often attacked by typhoons due to its geographic location in Monsoon Asia. This island suffers from severe typhoon threats during summer and autumn in each year, and different typhoon paths may cause different disasters in Taiwan. For example, the heavy rainfall coupled with high intensity induced by Typhoons AERE (2004), KROSA (2007), MOLAVE (2009), SAOLA (2012) and SOULIK (2013) brought disasters in northern, middle and southern Taiwan, which made great challenges to reservoir operators. Taiwan has small catchments, steep-sloped terrains and rapid river flow. Therefore the development of forecast models for river flow or reservoir inflow are mainly of short-term scales, e.g. hourly forecasting, for providing a shorter response time to disaster management and flood mitigation. If we can provide flow prediction for the whole typhoon period before a typhoon hits Taiwan, it will help to make the regulation strategy of reservoir discharge and provide the reference guide of real-time flood control operation. In this study, we analyze the relationship between the classification of typhoon paths and reservoir inflow patterns to propose the methodology of long-term inflow prediction during the whole typhoon period. The main idea is to first classify typhoon paths by using the self-organizing maps (SOM), then define the flow characteristic curves based on the classification results of typhoon paths, and consequently combine all the results with the total rainfall forecast of the catchment to obtain the desired long-term inflow prediction. The proposed model is divided into two stages: typhoon path classification; and reservoir inflow forecast. In the stage of typhoon path classification, the latitude and longitude coordinates of each typhoon path are converted into a grid vector, and the SOM is used to classify all the grid vectors to generate a topological map of typhoon paths. As a result, the flow characteristic curve and statistics of each neuron in the topological map can be obtained. In the stage of reservoir inflow forecast, we use the testing typhoon events to make inflow forecasts. The typhoon path of each testing event is classified into a neuron of the SOM. Then long-term inflow prediction can be made for the whole typhoon period based on the flow characteristic curve, the runoff coefficient and the total rainfall forecast (provided by the Central Bureau in Taiwan) of the neuron into which the testing typhoon event is classified. The results indicate that the performances of the neurons in the topological map of typhoon paths are distinct from each other, and the topological relationship between a neuron and its neighboring neurons also shows slight differences in typhoon paths. Besides, the typhoon paths and peak flows in each neuron are quite similar, and the flow characteristic curves of most of the neurons in the SOM shows high similarity. The proposed inflow prediction model can grasp the trend for the whole typhoon period and accurately predict the timing of peak flow before a typhoon hits Taiwan. Results indicate that peak flow and total forecasted inflow also fall within the acceptable range, and the predicted and actual flow hydrographs produce a high correlation coefficient