Comparative Analysis of Adam- and PSO-Optimized ANFIS Models for Intelligent Control of Wastewater Treatment Processes

The effective control of industrial wastewater treatment processes based on ion-exchange resins is challenging due to their nonlinear dynamics, time-varying operating conditions, and uncertainty in water quality parameters. In particular, variations in water hardness and total dissolved solids (TDS) strongly influence purification efficiency, necessitating adaptive and intelligent control strategies beyond conventional approaches. In this study, an adaptive neuro-fuzzy inference system (ANFIS)-based intelligent control model is developed to regulate the opening degree of a control valve governing the wastewater flow rate in an ion-exchange treatment process. A laboratory-scale experimental setup was designed and implemented, and a dataset of 300 experimental samples was collected under diverse operating conditions. Water hardness and TDS were selected as input variables, while the valve opening degree was defined as the output control variable. A clustering-based rule extraction method was employed to construct the ANFIS structure, and model parameters were optimized using the Adam optimizer and Particle Swarm Optimization (PSO). The performance of the ANFIS–Adam and ANFIS–PSO models was evaluated using regression and control performance metrics, including RMSE, MAE, R², settling time, and integral absolute error (IAE). The results indicate that both optimization algorithms significantly enhance ANFIS performance while exhibiting complementary strengths. The ANFIS–Adam model achieves faster convergence and improved dynamic response, reducing settling time by approximately 15–20%, making it suitable for real-time control applications. In contrast, the ANFIS–PSO model demonstrates superior robustness and global search capability, achieving up to 10–15% lower RMSE and improved steady-state accuracy. Both models attain high prediction accuracy (R² > 0.96), suggesting that the choice of optimizer should be guided by specific control objectives, such as real-time responsiveness or robust offline tuning.