A rolling mill stiffness prediction model based on optimized multi-output support vector regression

Abstract The stiffness of a hot rolling mill is a paramount parameter for producing high-quality thin-strip steel. Roller misalignment, primarily caused by wear, accounts for approximately 60% of stiffness variation, challenging thickness control. This study proposes a novel hybrid model to predict mill stiffness under such misalignment conditions. First, a high-fidelity Finite Element Analysis (FEA) model was developed and validated against experimental data, showing a maximum relative error of 5.49%. Leveraging data generated from this FEA model, a Multi-output Support Vector Regression (MSVR) model was constructed. Its hyperparameters were optimized via Particle Swarm Optimization (PSO) to map the nonlinear relationship between wear conditions and mill stiffness. The proposed PSO-MSVR model demonstrated significantly higher prediction accuracy than standard MSVR and Backpropagation Neural Networks (BPNN). When tested with real mill data, PSO-MSVR achieved an average prediction error of less than 3%, proving its efficacy as a robust tool for stiffness prediction and maintenance guidance in industrial settings.