Real-scale demonstration of digital twins-based aeration control policy optimization in partial nitritation/Anammox process: Policy iterative dynamic programming approach

被引：0

作者：

Heo, SungKu ^{[1
,2
]}

Oh, Taeseok ^{[3
]}

Woo, TaeYong ^{[1
]}

Kim, SangYoon ^{[1
]}

Choi, Yunkyu ^{[3
]}

Park, Minseok ^{[3
]}

Kim, Jeonghoon ^{[3
]}

Yoo, ChangKyoo ^{[1
]}

机构：

[1] Integrated Engineering, Dept. of Environmental Science and Engineering, College of Engineering, Kyung Hee University, 1732 Deogyeong-daero, Giheung-gu, Gyeonggi-do, Yongin-si,17104, Korea, Republic of

[2] Department of civil and environmental engineering, Imperial College London, London,SW7 2AZ, United Kingdom

[3] BKT Co. Ltd, 25 Yuseong-daero 1184 beon-gil, Yuseong-gu, Daejeon,34109, Korea, Republic of

来源：

Desalination | 2025年 / 593卷

基金：

新加坡国家研究基金会;

关键词：

Application programs - Chemical oxygen demand - Chemicals removal (water treatment) - Nitrogen removal - Reinforcement learning - Water aeration;

D O I：

10.1016/j.desal.2024.118235

中图分类号：

学科分类号：

摘要：

Partial nitritation (PN) and anaerobic ammonium oxidation (Anammox) process is a promising energy-efficient nitrogen removal method in wastewater sector. Recently, artificial intelligence (AI)-driven process operation techniques are widely researched. However, there is few research to demonstrate AI application into a full-scale wastewater treatment plant (WWTP) due to operational complexity of WWTP. This study conducts a real-scale demonstration of digital twin-based aeration control policy (DT-O2CTRL) to autonomously control the full-scale PN/A process under high nitrogen influent loads. For this, chemical oxygen demand (COD) and NH4-N in influent and reactors, were collected through the online sensors. Then, digital twin (DT) model of full-scale PN/A process was mathematically developed. Finally, policy iterative dynamic programming (PIDP), inspired from the reinforcement learning, was suggested as the core algorithm of AI-O2CTRL to maintain a NO2-N/NH4-N ratio (NNR) which is a critical operation factor in PN/A process. The results showed that the DT model showed an accuracy of >95 %. Based on the DT model, the AI-O2CTRL algorithm autonomously controls the NNR at the target value of 1.1 and reduces electricity consumption by 16.7 % when treating around 400 m3/d of enriching nitrogen loads. Finally, it can reduce the operational cost by 19,724.01$/year regardless of the influent load fluctuations. © 2024 Elsevier B.V.

引用

下载