台灣早期住宅建設，並無強制要求將住宅內雨水與污水管線分流，導致汛期及豪大雨時，外水流入污水下水道系統，其管線設計已無法負荷現今極端氣候所帶來瞬間雨量，影響污水管網正常使用及後端污水處理廠操作處理，污水下水道系統形成疏洪渠道，污水處理廠淪為大型抽水站，不堪負荷甚至災損。 本研究為探討污水下水道管網高水量異常模式，利用案例中降雨量、地下水位之每小時相關公開數據，全年共8,760組數據。研究方法第一個步驟以裝設驗證設備，取得比對實際液位數據，然後進行資料前處理，將全年數據進行除錯及分類；第二個步驟即將水量異常資料庫應用監督式學習倒傳遞類神經網路程式，將資料庫之數據參數使機器自我學習建立降雨量、地下水位及綜合數據對管網液位異常預測三種最佳化模式；第三步驟為將訓練完成後所有模式加入未知的測試降雨量及地下水位資料，放入網路程式驗證，獲取管網液位預測值；使用已知的案例實際值與預測值進行準確率及標準偏差計算。 結果顯示本研究所提出之綜合數據對管網液位異常預測模式準確率達八成以上，如果再加入線性迴歸分析修正，可以得到更佳的預測結果。建立之污水收集管網高水量異常預測模式，進而可使相關單位及決策者，可在不必另行購買昂貴監測設備安裝於各區域下，提前預警調整操作參數及分散高洪峰水量等異常問題。 ;Taiwan’s early residential construction didn’t enforce separation of sewage pipe and storm water pipe, which is why additional water along with underground water run into the same underground piping system during heavy rain and high tide situation. Such piping system can no long handle the amount of water from recent extreme weather condition, piping network becomes flood channel, and sewage plant station become a big pump station, resulting overloading or even catastrophic disaster. This study mainly explores the abnormal water level pattern in the sewage piping network. Data in the case study is recorded every hour, 24 sets a day, which equals to 8760 sets per year. It’s divided into three steps: First step is to install verification device, compare data, debug errors and sorting information. Second step is to apply the abnormal water level from the database to a supervised back propagation neural network model learning，applying abnormal rain fall water level mode, abnormal underground water mode, and all round estimation mode, to neural network model learning. Last step is to apply all finished mode into training data, entering network verification, and use Linear loop analysis to compare the real figure with the estimated figure. Results were promising to 80% accuracy. By establishing abnormal sewage collecting network system, allowing relevant institutes or officers to warn and adjust control figures and divert flood without paying expensive equipment to be installed at different sectors.
