(59af) Using Artificial Neural Networks for Real-Time Tuning of PID Controllers

Many chemical processes can be operated closer to optimality using real-time tuning rather than traditional proportional-integral-derivative (PID) control [1-2]. Artificial Neural Networks (ANNs), a branch of machine learning, have been demonstrated to be capable of predicting PID parameters for second order mechanical systems and third order chemical systems across a wide range of system dynamics [3-4]. This implies that they could be utilized for real-time tuning by updating PID tuning parameters as process inputs and/or setpoints change over time. Processes which depend on renewable energy sources, which are often intermittent, could benefit from employing such dynamic control to account for changes in energy availability. This work utilizes an ANN to predict PID tuning parameters for a third-order chemical system with a dynamic setpoint.

A Python script was used to generate a database of PID controller parameters which produce acceptable responses for an input step-change. This was performed for 100,000 variations of the chemical system. These were used to train the ANN to predict the PID tuning parameters based on the system inputs and dynamics. The ANN-tuned PID controller was then tested on the same system with setpoints which vary across time to see how well the PID controller performed.

The ANN-tuned PID-controller was found to out-perform the Ziegler-Nichols tuning method and the Tyreus-Luyben tuning method. The ANN-generated parameters resulted in almost three times smaller overshoot and significantly shorter settling time in every case tested. It was also found that the ANN could control the chemical system effectively, so that the output closely follows the setpoint as it changes over time. Once the dynamics of a system have been modeled mathematically, a database of PID parameters can be generated and an ANN can be trained to predict real-time PID-controller parameters for a controller to regulate that system. Finally, the results indicate that ANNs allow for improved control in processes which employ intermittent energy or resource supplies; this would considerably enhance the performance of such processes.

References

[1] P. Vanrolleghem, L. Benedetti, J. Meirlaen, 2005, Modelling and real-time control of the integrated urban wastewater system, Environ. Modell. Softw., 20, 4

[2] J. Koo, D. Park, S. Ryu, G. H. Kim, Y. W. Lee, 2019, Design of a self-tuning adaptive model predictive controller using recursive model parameter estimation for real-time plasma variable control, Comput. Chem. Eng., 123

[3] T. Bestwick, K. V. Camarda, 2023, Artificial Neural Network-Based Real-Time PID Controller Tuning, Comput. Aided Chem. Eng. (in press)

[4] Y. S. Lee, D. W. Jang, 2021, Optimization of Neural Network Based Self-Tuning PID Controllers for Second Order Mechanical Systems, Appl. Sci., 11, 17