Targeted Output Driven Dataset Generation Methodology for Training Neural Network Controllers in Specific n^th Order Systems

The performance of a neural network controller majorly depends on the dataset used for training. Traditionally, Proportional-Integral (PI) controller data is used to train a neural network. In this scenario, the performance of the neural controller is limited by the performance of the PI. Similarly, any other controller data, when used, restricts the neural controller’s performance within its limitations. In this paper, a targeted output-driven dataset that theoretically gives the best desirable response is derived and used for training the neural controllers. Therefore, enhanced performance could be achieved that is better than that of traditional PI controllers. Detailed theoretical derivations are provided for obtaining the proposed dataset generation methodology for any n<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">th</sup> order system without zeros. The viability of the proposed neural controller was simulated and trained in a MATLAB/Simulink environment, and its response to variation in reference input has been inferred. The performance comparisons are made with the PI controller data and presented.