Data-driven flatness presetting control ensemble method for skin pass rolling mill

The presetting values of the skin pass rolling mill determine the flatness accuracy. However, the traditional presetting model for skin pass rolling mills is not sufficiently accurate, resulting in severe flatness defects. To address this challenge, an ensemble presetting control method was proposed in this paper. In order to fully acquire the experience knowledge from historical production data, the density peak clustering (DPC) method was used to construct a tabular presetting model by finding clustering centers corresponding to serial numbers. To accommodate the unforeseeable fluctuations in incoming materials, the CatBoost model was employed to predict specific presetting values for the strip based on incoming material information. The main feature of this ensemble model is the integration of historical data commonalities and specific coil characteristics through the three-sigma principle to ensure appropriate presetting values. The tabular and tree models were also applied to ensure computational speed is sufficiently fast for practical application in the factory. The simulation results indicate that the Mean I-Unit (IU) value of the entire strip length was reduced by 15.560%, and the flatness quality was improved. The optimized presetting values were applied in actual production. The results indicate that the optimized strip head flatness was significantly improved, with the mean IU value of the first 50 m of the strip decreased by 22.99%, and the flatness of the entire length of the strip was also improved, with the mean IU value of the entire strip length reduced by 15.74% compared to the traditional presetting control model. Meanwhile, the time for various flatness control measures to reach the stable feedback stage was significantly reduced. The above research provides references for the presetting values optimization of the skin pass rolling mill.