以機器學習模型預測再生水廠放流水中的氨氮含量

Abstract

在台灣，水短缺的問題日益嚴重，因此再生水作為永續水資源的使用量逐漸增加。然而，再生水廠的處理過程會產生濃排水，有機會造成水體污染。有效的水質監測和預測對於管理十分重要，其中氨氮是評估水質的一個關鍵指標。本研究利用機器學習技術建立預測再生水廠逆滲透濃排水中氨氮含量的模型，並比較淺層學習、深度學習以及自動化機器學習演算法，以尋找最有效的預測模型。為了可以提早預測濃排水水質，本研究以不同數據量及相關性組合，訓練模型預測當天和預測未來日子的氨氮含量。預測結果主要以R平方值進行模型效能評估，顯示長短期記憶網絡(LSTM)的表現整體超越其他演算法。以再生水的所有單元之水質預測濃排水，LSTM模型可達到0.96的R平方值。即使減少訓練數據中的再生水廠處理單元，只使用進流水水質和操作參數，LSTM仍可得出0.82的R平方值準確性。而於預測下一天的氨氮含量，LSTM模型的R平方值為0.64，顯示出尚可的預測準確性。本研究利用機器學習建立高準確性的濃排水氨氮預測模型，以期協助再生水廠未來的運行操作，提早預測之氨氮含量以提供再生水廠單位決策參考，同時此模型有助於降低附近河川受濃排水影響而引起的優養化之風險。

In Taiwan, the growing problem of water scarcity is promoting the increased use of reclaimed water as a sustainable resource. However, reclaimed water can produce waste products like concentrate, which may harm aquatic environments if discharged directly. It’s important to monitor and predict the quality of effluent effectively. This study employs machine learning techniques to develop a model that predicts ammonia nitrogen levels in the reverse osmosis concentrates from a water reclamation plant. Various shallow learning, deep learning and automated machine learning algorithms are tested to find the most effective algorithm for the prediction model, meanwhile exploring the optimal dataset combinations for both immediate and future effluent quality predictions. Results were evaluated based on R-squared values, indicating that the LSTM method outperforms other algorithms. With comprehensive treatment data, the LSTM model achieved a high accuracy with an R-squared value of 0.96. Even with reduced data, including only the influent and operational parameters, the LSTM model maintained an R-squared value of 0.82. For future predictions one day ahead, the LSTM model achieved an R-squared of 0.64, representing a fair accuracy. To summarize, this study utilizes machine learning algorithms to develop a highly accurate prediction model for ammonia nitrogen levels in concentrates of water reclamation plant. This model aims to assist water reclamation plant operators by providing early predictions to support decision making, thereby potentially reducing the risk of eutrophication in nearby rivers affected by effluent discharge.