Data-driven adaptive setting algorithm for coiling temperature model parameter

Abstract

To improve the coiling temperature control accuracy for change-over strip or the first coil of batch hot-rolling, data mining technology was adopted to infer the water cooling learning coefficient which is used in coiling temperature model preset for actual rolling strip from massive production data. Firstly, cooling feature parameters were recognized and defined respectively as absolute, relative, equal and tactical type. Then, the similar distance of each feature parameter between actual rolling strip and each historical rolled strip was calculated and summed. When the total similar distance of each rolled strip met the requirement, the produced strip was clustered as similar with actual rolling strip. Meanwhile, the weight value of the similar strip was calculated by considering its time effect. Secondly, based on the cooling information of the head and tail ends of each similar rolled strip, three object functions which are respectively composed of temperature predictive error and related penalty items such as a penalty deviated from regression learning coefficient and a penalty departed from the default learning coefficient were created and the corresponding constraints were also given. Gradient descent method was utilized to solve the quadratic programming problem. After three mathematical optimization calculations, a referenced learning coefficient and two parameters reflecting the relationship between the learning coefficient with rolling speed and target coiling temperature were obtained and then used to compute the learning coefficient needed in the cooling schedule calculation according to thread speed and target coiling temperature of the actual rolling strip. Application results show that the presented model’s adaptive parameter setting algorithm, based on the cooling data of 100,000 rolled strips can enhance the pre-setup ability of the coiling temperature model for strip head end. The adaptive setting ability of the learning coefficient will increase with the diversity of the strip cooling data stored in the memory and the number of similar strips retrieved.

Keywords

• cooling feature parameter
• clustering
• self-adaptation
• parameter estimation
• coiling temperature