Enhancing Neural Network Performance in Skin-Pass Mill Modeling through Generalized Extremal Optimization

Abstract

This study is justified by the need to precisely adjust controllers in skin-pass mills to ensure the quality of steel sheets, given the complexities of the rolling process. Traditional models fail to incorporate the historical wear of the mills, thus data-driven modeling is proposed as a viable approach to accurately represent these conditions and optimize production quality. The aim is to optimize the architectures of Artificial Neural Networks (ANNs) using the Generalized Extremal Optimization (GEO) algorithm, which stands out for its ability to adjust network architectures with a single free variable, simplifying configuration and enhancing computational efficiency. The methodology involved using the GEO algorithm to balance error minimization and computational cost, suitable for implementations in Programmable Logic Controllers (PLCs) with limited resources. As a significant result, the optimized ANN architecture achieved a Mean Relative Squared Error (MRSE) of 0.8418%, demonstrating an efficient solution given the application's constraints.