A Novel Data-Driven Rolling Force Predictive Model Based on PSO-BAS-HKSVR Approach

A Novel Data-Driven Rolling Force Predictive Model Based on PSO-BAS-HKSVR Approach

To remedy the defects of the single kernel function and PSO algorithm, a novel rolling force prediction model is proposed, combining particle swarm optimization (PSO) algorithm, beetle antennae search (BAS) algorithm and hybrid kernel function support vector regression (HKSVR), i.e., PSO-BAS-HKSVR model. Hybrid kernel function (HKF) is incorporated to reduce the defect of the single kernel function of support vector regression. In the meantime, PSO algorithm is improved and combined with BAS algorithm to optimize the HKSVR model parameters (C, g , d , ε , m ) Statistical indicators (R2 , RMSE, MAE and MAPE) are introduced to assess the comprehensive property of the model. The experimental data of the training and testing model originate from the actual production line of the steel plant. Rolling temperature, thickness reduction, initial strip thickness and width, front tension, back tension, roll diameter, and rolling speed are taken as the input variables. Under the identical experimental conditions, compared with the single SVR, PSO-SVR, PSO-HKSVR, BPNN, GRNN and RBF models, PSO-BAS-HKSVR model exhibits the highest prediction accuracy and the optimal generalization ability. As indicated from the results PSO-BAS-HKSVR method is suited for the rolling force prediction and the optimization of model parameters in the hot strip rolling process.