High Precision Prediction of Hot Rolled Medium Plate Based on Stacking Model

In the production of hot rolled steel plates, thickness is a crucial quality parameter that significantly influences both the product's yield and the stability of the rolling process. Considering the characteristics of multi-variable, nonlinearity, and strong coupling in the plate rolling process, the prediction accuracy of existing theoretical models is not high. In this paper, an improved ensemble method based on particle swarm optimization (PSO) and Stacking is proposed. Firstly, the random forest (RF) algorithm is introduced for feature selection to enhance the rationality of input features. Secondly, various data-driven models for predicting steel plate thickness are developed, employing PSO to optimize the model parameters. The results indicate that the predictions from the data-driven model outperform existing theoretical models, and the optimized model demonstrates higher accuracy. To further enhance prediction accuracy, a Stacking model that integrates multiple machine learning algorithms is developed, and various Stacking strategies are proposed for comparison. The test results show that the Stacking model with RF as the meta learner and other models as the base learners performs the best. On the test dataset, the values for MAE, RMSE, and R² are 0.9429, 1.1176, and 0.9812, respectively. Compared with the best single model, these values are improved by 32%, 34%, and 8%, respectively. This enhancement meets the requirements for high precision and stable production.