This study tackles two challenges in aluminum-magnesium (Al-Mg) alloy rolling – control delay from measurement hysteresis and strong nonlinearity with limited passband precision – by proposing a predictive control scheme that couples an optimized starting-point combinatorial discrete grey model (OSCDGM), with an improved PID neural network (IPIDNN).
An OSCDGM framework based on an Optimized Starting-Point Discrete Grey Model (OSDGM) uses change-point-driven updating with tail correction and Fourier-series residual compensation to achieve high-accuracy real-time prediction of inter-stand thickness. Predicted and measured values are fused to drive online learning of an IPIDNN controller, which augments a PID neural network through variable-speed integration, incomplete differentiation and tanh-based nonlinear mapping to enhance tracking, passband control and dynamic response.
Field experiments show that OSCDGM reduces prediction errors by 36–42% relative to Even Grey Model (EGM), Discrete Grey Model (DGM) and Support Vector Machine (SVM). IPIDNN shortens settling time by 79–88% compared with Back Propagation-Proportional-Integral-Derivative (BP-PID), PID Neural Network (PIDNN) and model predictive control. Under disturbances and noise, the standard deviation of control deviation is reduced by 68% versus conventional PID, with fluctuations confined within ±3s and thickness accuracy improved by about 27% in production.
Application to full-pass rolling of AZ31B alloy verifies engineering applicability and demonstrates markedly improved thickness consistency and stability, while highlighting the need to assess generalization to alloys with different plastic properties.
The study establishes an integrated architecture combining high-precision dynamic prediction with structurally enhanced neural-network-based control, providing a paradigm for intelligent control of high-precision metal rolling with significant time delays and nonlinearity.
