What question did this study set out to answer?

To develop an optimization framework for the hot rolling process of thick plates focusing on energy consumption and power difference.

January 16, 2026

A Mechanics‐Informed Modeling and Optimization Framework for Hot Rolling of Thick Plate Using Convolutional Neural Networks and Multiobjective Particle Swarm Optimization Algorithm

Key Points

To develop an optimization framework for the hot rolling process of thick plates focusing on energy consumption and power difference.
Utilized industrial big data from various steel grades for model training.
Developed convolutional neural networks (CNNs) and a backpropagation (BP) network model for predicting rolling power.
Optimized the rolling power model using multiobjective particle swarm optimization algorithm to achieve a Pareto front.
Compared different CNN architectures based on performance and stability.
Models trained on balanced datasets outperformed those trained on original data.
CNN models showed greater accuracy and stability compared to BP models.
A single-layer CNN with optimized width exceeded the performance of a three-layer LeNet-5 architecture.
The solutions on the Pareto front provide flexible guidance for optimizing production based on specific requirements.

Abstract

The optimization of the hot rolling process for thick plates critically depends on establishing an appropriate rolling schedule. This study presents a multiobjective optimization framework targeting energy consumption and power difference within the rolling schedule. Initially, industrial big data across various steel grades are balanced to enhance model generalizability. Subsequently, convolutional neural network (CNN) models of varying depths and a backpropagation (BP) network model are developed to predict rolling power, with the influence of convolutional layer width on CNN performance also investigated. The rolling power model is then optimized using the multiobjective particle swarm optimization algorithm, yielding a Pareto front. Comparative analyses demonstrate that models trained on the balanced dataset achieve superior predictive performance over those using the original data. Furthermore, the CNN models exhibit greater accuracy and stability than the BP model, while a single‐layer CNN with an optimized width outperforms a three‐layer LeNet‐5 architecture. The solutions on the Pareto front can be selectively implemented based on specific production requirements, offering flexible guidance for optimizing the thick plate hot rolling process.

Mark Helpful

Bookmark

Relay