(674h) Machine Learning Based Controller Design for a Wastewater Treatment Plant Benchmark

The Benchmark Simulation Model No. 1 (BSM1) serves as a fundamental tool for simulating the complex dynamics of wastewater treatment plants. In recent years, there has been a growing interest in enhancing the control strategies of this benchmark plant to improve its efficiency and environmental performance. This abstract presents a novel approach of controller design for BSM1, leveraging the power of machine learning.

Traditional industrial control methods, such as PID, often struggle to handle wastewater treatment processes due to its nonlinear and time-varying nature, as well as stringent constraints. Similarly, the model-based advanced process controller, such as model predictive control (MPC), suffers from uncertainties, disturbances, and model mismatch, rending it less suitable for implementation in such a complex process. In fact, the MPC was applied for highly simplified BSM1, but rarely for the full-size simulator. Moreover, the effectiveness of these control strategies is heavily reliant on optimized setpoints, further complicating the control process. In response to these challenges, we propose using machine learning to simultaneously optimize the controller parameters and setpoints of BSM1. The design features include 7 parameters for 2 anti-windup PID, and 2 setpoints. The response variable to be fitted in machine learning is the average water quality constraint violation across 15 distinct rain scenarios. Once the machine learning model is established, an optimization algorithm is developed to find a better controller design. This fitting-and-optimize procedure is repeated until no further improvement can be achieved.

While conventional machine learning methods, such as neural network, SVM, and gradient boosting have been employed to fit the performance data, they often fail to obtain satisfactory performance (R-squared) even with Bayesian optimization for hyperparameter tuning. To capture intricate patterns and adapt to changing conditions, we propose a classification-regression procedure to divide the design space into several subspaces for localized data fitting. In addition, a feature extraction and selection algorithm is introduced to enable more flexible function structure in the data fitting. Given such a complex machine learning model, the searching algorithm is executed to find the optimal design, which is further evaluated through simulation of 15 rain scenarios.

One key advantage of the proposed approach is its adaptability to varying influent conditions and unforeseen disturbances. The proposed machine learning framework can continuously learn and update its internal representation, allowing the control system to adjust in real time to changes in the wastewater characteristics. This adaptive capability is crucial for maintaining high treatment efficiency and complying with stringent environmental regulations. Furthermore, the resulting controller aims to address the challenge of minimizing energy consumption while meeting effluent quality standards. By optimizing the controller parameters and setpoint, the treatment plant can operate more efficiently, reducing both operational costs and water quality violation.

Simulation studies validate the superiority of our proposed machine learning framework over existing methodologies such as SVM, neural networks, and gradient boosting, in terms of both data fitting accuracy and controller performance improvement.