Prediction of head thickness of hot-rolled strip based on deep neural network

Abstract

Aiming at the problem of low precision of hot-rolled strip head thickness, a prediction model based on deep neural network (DNN) was presented. Considering the root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE) and Pearson correlation coefficient (R), the data of 2,005 coils collected from a hot rolling production line were used to establish the model. Hybrid particle swarm optimization and genetic algorithm (HPSO-GA) was used to optimize the prediction performance of DNN model. Parameters setting of DNN model and HPSO-GA, including learning rate, hidden neurons, selection of optimizer, population size and crossover probability were investigated to obtain the optimal model. Prediction performance comparisons of support vector regression (SVR), random forest (RF), DNN and HPSO-GA-DNN were presented. HPSO-GA-DNN had the smallest error and the highest R value, in which RMSE, MAE, MAPE and R value were 0.0197mm, 0.0151mm, 0.499% and 0.989, respectively. And 96.74% of the prediction data had an absolute error of less than 0.05 mm. The results showed that HPSO-GA-DNN had strong learning ability and good generalization performance, and could be well applied to hot rolling production.