Interval Construction and Optimization for Mechanical Property Forecasting with Improved Neural Networks

Abstract

Efficient and accurate predication of mechanical properties is the key to controlling the production process. In this paper, a novel Prediction Interval (PI) based method is proposed for forecasting strip steel properties. It specifically consists of a Lower Upper Bound Estimation (LUBE) technique for PI generation based on Particle Swarm Optimization (PSO) and a Coverage Width Symmetry-based Criterion (CWSC) for PI evaluation. To evaluate the proposed method, computational experiments are carried out on two numerical datasets and two real-world datasets from a strip steel production process. A comparison between the results obtained by this work and previous work shows that the proposed method is viable and achieves more advantages. Moreover, the PI constructed on the real-world datasets achieve better quality, demonstrating that the proposed method has good potential in real-world problems.