RF-KNN-Assisted Local Gaussian Process Regression for Heat Transfer Coefficient Prediction in Hot Strip Coiling...

Accurate prediction of the heat transfer coefficient is essential for improving the coiling temperature control in hot strip rolling, especially under frequent rolling condition changes. Conventional layer-based self-learning methods may lead to boundary discontinuities, insufficient sample support for new gauges, and limited information sharing among similar operating conditions. To address these limitations, this paper proposes a random-forest (RF) and K-nearest-neighbor (KNN)-assisted local Gaussian process regression framework for the heat transfer coefficient in hot strip rolling. In the proposed method, RF is first used to select key variables and guide similar-case retrieval. KNN is then employed to retrieve the historical strips most similar to the current strip and to construct a local sample space. Instead of directly using conventional distance-weighted averaging, Gaussian process regression (GPR) is established on the retrieved local samples to model the nonlinear relationship between the process variables and the heat transfer correction coefficient. The proposed method outperforms conventional KNN-based weighting methods in terms of all the evaluation metrics for both first coils and in-lot coils at different speeds. Industrial validation shows that the measured coiling temperature is controlled within ±20 °C over more than 96.5% of the coil length. The results demonstrate that the proposed framework improves the adaptability and online compensation capability of the controlled cooling temperature models.