A dog‐bone profile is developed by inhomogeneous deformation of a slab, whose parameters can be used to describe the metal flow rule during the width reduction process by a sizing press. The accuracy of a dog‐bone profile is essential to ensure the high width precision of the roughing process. However, the existing equipment cannot satisfy the requirements of online measurement. To solve this problem, herein, a prediction model for dog‐bone parameters based on the finite‐element simulation and a neural network is presented. The finite‐element simulation can reproduce the field production process, revealing variations in the dog‐bone profile parameters. In addition, based on the simulation data, a neural network model is constructed to predict the dog‐profile parameters. By adjusting the hyperparameters, the prediction accuracy of the constructed deep neural network model with two hidden layers is improved. For the dog‐bone profile parameters, the mean squared error is less than 3 mm, the mean absolute percentage error is less than 2%, the maximum absolute percentage error is less than 9%, and the root mean squared error is less than 4 mm. Thus, an accurate prediction of the dog‐bone profile parameters is achieved, which can guide subsequent high‐precision width control.

During the slab sizing press (SP) process, the pressing force corresponds to the slab profile, which guides the production schedule design and the final profile control. To accomplish the prediction of pressing force for SP, an improved ensemble method based on chaotic Harris hawks optimizer (CHHO) and stacking is proposed. A mechanistic knowledge is introduced for feature selection that enhances the rationality of input features. Subsequently, 11 machine learning models are compared and 5 of them are selected as candidate learners for the stacking method. Based on the candidates learners, 8 stacking strategies are constructed, which the stacked model with extratree regressor, gradient boosted decision trees, and kernel ridge regression (KRR) as base‐learners and KRR as meta‐learner performs the best. The R2, MAE, mean square error, mean squared log error, and mean absolute percentage error for the test dataset are 0.9912, 0.0856, 0.0167, 0.0005, and 2.00%, respectively, and 95% of the prediction errors are less than 0.15 MN. Then, sensitivity analysis and predictive analysis based on Shapley Additive Explanations are performed to demonstrate the good alignment of the proposed model with physical reality. Furthermore, to cope with the complexity and uncertainty of the production process, the proposed CHHO‐stacking model and the kernel density estimation method are integrated to model the prediction interval.