Adaptive Tuning of a Proportional Integral Controller Using a Transformer Network

The tuning of gains in Proportional-Integral (PI) controllers is a fundamental challenge in industrial automation, particularly when systems experience changes in their dynamics. This work proposes a solution based on artificial intelligence to address this challenge, combining classical control methods with deep neural networks such as Transformers leveraging attention mechanisms. A dataset is generated using the classical Pole Placement method, adjusting PI controller parameters in response to changes in the dynamics of a first-order system. This dataset serves as the foundation for training a deep neural network based on the Transformer architecture, specifically a TabTransformer, which learns to predict the appropriate PI controller gains (Kp and Ti) based on the system’s features and dynamics. Results suggest that the proposed approach shows promise compared to standard neural networks, obtaining R² values of 0.9866 for Kp and 0.9979 for Ti prediction.