員山子分洪工程自2004年啟動應急分洪至2012年已有24場颱風豪雨事件之分洪，有效降低基隆河員山子分洪堰中下游河川水位，洪災狀況已明顯改善。 本文綜整歷史分洪事件降雨紀錄與水位變化資料，建立員山子攔河堰堰前水位達分洪預警水位(62.5公尺)到分洪警報水位(63.0公尺)時水位與雨量之迴歸模式，利用系集預報分析研判分洪預警水位前各小時可能引發分洪時之雨量警戒值，提供預報員山子分洪時機。 本研究利用8個系集預報模式進行分析，結果顯示當員山子攔河堰堰前水位達分洪預警水位時，火燒寮及三貂嶺雨量站前一小時及前二小時之雨量值超過3個模式分洪時雨量之門檻值，則預測員山子將啟動分洪。;There have been 24 flood diversion events (including the 2004 emergency flood diversion event) since Yuanshanzi flood diversion project completed in 2004 till 2012. The diversions successfully lower water level of Keelung River and prevent flooding caused by heavy rain happen in upstream. Flooding threats have been significantly reduced for the downstream area. In this paper, the history of flood records, including rainfall observations and water level data, is analyzed to establish a regression model on water level verses rainfall. Specifically, the analysis is focus on water levels between the warming water level (EL=62.5m) to flood diversion level (EL=63.0m) at the Yuanshanzi flood diversion site. Ensemble forecasting analysis is introduced to judge whether given rainfall observations in current time periods may lead to flood diversion in the coming hour. Thus, rainfall alert values are proposed to be indicators for predicting flood diversion taking place. In this study, eight ensemble forecasting models for analysis are proposed. Results show that while the 1-hour and 2-hour rainfall on Huoshaoliao and Shangdiaoling rainfall observations higher than rainfall alert values, if three out of eight models indicate diversion taking places, the prediction on diversion is suggested to be issued.