根據世界衛生組織所述，水是生命不可缺少的元素之一。淨水場是為人們提供自來水的重要基礎建設。在淨水場的複雜程序中，混凝作為淨水流程的第一道程序，尤為重要。目前，多數淨水場的混凝程序需仰賴廠內工作人員之主觀判斷，並輔以杯瓶試驗（Jar test）進行加藥作業。然而，由於原水水質變異性大，若要即時針對不同的水質條件調整加藥量，將會提升人力負擔。近年來物聯網（Internet of Things，IoT）以及人工智慧（Artificial Intelligence，AI）技術發展快速且日益普及，若能將此技術引入混凝程序，得以更加自動和智能地處理。因此，本研究旨在應用AIoT（Artificial Intelligence of Things）實現自動化及最佳化的混凝程序，以達到節省能源、降低成本之目的。自動化程序係指運用AI及統計模型依據水質條件模擬操作人員對混凝劑劑量之判斷，並利用水量及理論公式計算攪拌轉速。而最佳化程序係指以水質條件、混凝劑劑量、轉速等訓練資料，通過AI及統計模型推估對應的濁度變化量，並考量飲用水水質標準及藥劑與能源成本後，求得最佳化操作參數。結果表明，自動化程序中長短期記憶模型之準確性較高，其RMSD（Root-Mean-Square Difference）為0.0013 ppm，而最佳化模型中多元非線性迴歸之準確性較高，其RMSD為1.57 NTU。整體而言，不論是自動化程序或最佳化程序皆可以依據水質的變化，準確地推估混凝相關之操作參數。;According to the World Health Organization, water is one of the indispensable elements of life. Water treatment plants are the infrastructure to provide people with clean water. Within the complex procedure in water treatment plants, coagulation is particularly important as the first procedure of the water treatment process. At present, most water treatment plants require a lot of manual judgement on coagulation process through regular jar test and sensor observations. However, to deal with the dynamically changing nature of water quality and water volume, the coagulation dosing procedure becomes a labor-intensive task. With the recent development of Artificial Intelligence of Things (AIoT), this procedure could be handled automatically and intelligently. Therefore, this research aims on applying the AIoT to achieve an optimized and intelligent coagulation dosing procedure for the purpose of saving energy and reducing costs. The proposed automation procedure refers to the use of Artificial Intelligence (AI) and statistical models to simulate human decisions on the coagulant dosage based on water quality, and the use of the water volume and theoretical formula to calculate the mixing speed. In addition, the proposed optimization procedure refers to estimating the turbidity change through AI and statistical models based on water quality, coagulant dosage, and mixing speed, which consequently helps estimate the lowest cost of coagulant dosage and energy consumption that satisfies operation standards. The results show that the Long Short-Term Memory achieves the highest accuracy of 0.0013 ppm Root-Mean-Square Difference (RMSD) among the examined models in the automation procedure, while the Multiple Non-Linear Regression achieves the highest accuracy of 1.57 NTU RMSD in the optimization procedure. In general, both procedures can accurately estimate the coagulant dosage and mixing speed settings according to the changes of water quality.
