LSTM-Based Neural Network Controllers as Drop-In Replacements for PI Controllers in a Wastewater Treatment Plant
Adil, Muhammad (Universitat Autònoma de Barcelona)
Vilanova, Ramon (Universitat Autònoma de Barcelona)

Data:	2025
Resum:	Wastewater Treatment Plants (WWTPs) rely on automatic control strategies to regulate pollutant concentrations and comply with environmental standards. Among them, Proportional Integral (PI) controllers are widely adopted for their simplicity and robustness, yet their effectiveness is limited by the nonlinear and time-varying dynamics of biological processes. In this work, Long Short-Term Memory (LSTM)-based Artificial Neural Network (ANN) PI controllers are proposed as data-driven replacements for conventional PIs in key WWTP feedback loops. Using the Benchmark Simulation Model No. 1 (BSM1), ANN controllers were trained to replicate the behavior of default nitrate and nitrite nitrogen ((Formula presented. )) and dissolved oxygen ((Formula presented. )) loops, under both time-agnostic and time-aware strategies with three- and four-input configurations. The four-input time-aware model delivered the best results, reproducing PI behavior with high accuracy (coefficient of determination, (Formula presented. )) and considerably reducing control errors. For instance, under storm influent conditions, the (Formula presented. ) controller reduced the Integral of Squared Error ((Formula presented. )) and Integral of Absolute Error ((Formula presented. )) by 84. 7% and 68. 4%, respectively, compared with the default PI. Beyond loop-level improvements, a Transfer Learning (TL) extension was explored: the trained (Formula presented. ) controller was directly applied to additional aerated reactors ((Formula presented. ) and (Formula presented. )) without retraining, replacing fixed aeration and demonstrating adaptability while reducing design effort. Plant-wide evaluation with the (Formula presented. ) loop and three dissolved oxygen loops ((Formula presented. ) - (Formula presented. )), all controlled by LSTM-based PI controllers, under storm influent conditions, showed further reductions in the Effluent Quality Index ((Formula presented. )) and the Overall Cost Index ((Formula presented. )) by 0. 84% and 1. 47%, respectively, highlighting simultaneous gains in effluent quality and operational economy. Additionally, the actuator and energy analyses showed that the LSTM-based controllers produced realistic and smooth control signals, maintained consistent energy use, and ensured stable overall operation, confirming the practical feasibility of the proposed approach.
Ajuts:	Ministerio de Ciencia, Innovación y Universidades PID2024-156522OB-C33 Generalitat de Catalunya 2021/SGR-00197 Agencia Estatal de Investigación PID2019-105434RB-C33
Drets:	Aquest document està subjecte a una llicència d'ús Creative Commons. Es permet la reproducció total o parcial, la distribució, la comunicació pública de l'obra i la creació d'obres derivades, fins i tot amb finalitats comercials, sempre i quan es reconegui l'autoria de l'obra original.
Llengua:	Anglès
Document:	Article ; recerca ; Versió publicada
Matèria:	Artificial Neural Networks (ANNs) ; Benchmark Simulation Model No. 1 (BSM1) ; Data-driven process control; Long Short-Term Memory (LSTM ; Operational Cost Index (OCI) ; Proportional-Integral (PI) control ; Transfer learning
Publicat a:	Applied Sciences, Vol. 15, Num. 22 (November 2025) , art. 12046, ISSN 1454-5101

DOI: 10.3390/app152212046

19 p, 924.5 KB

El registre apareix a les col·leccions:
Articles > Articles de recerca
Articles > Articles publicats