基于即时学习的高炉炼铁过程数据驱动自适应预测控制

Data-driven just-in-time learning based adaptive predictive control for blast furnace ironmaking

中文关键词  高炉  数据驱动  即时学习  线性化  模型预测控制  工业数据异常

英文关键词  Blast furnaces  Data-driven  Just-in-time learning (JITL)  Linearization  Model predictive control  Industrial data exceptions

中文摘要

      针对高炉炼铁过程, 本文提出一种基于即时学习的高炉铁水质量自适应预测控制方法(JITL-APC). 该方法的特点是控制器通过k向量近邻(k-VNN)方法搜索数据库中的I/O数据信息, 对非线性系统进行局部建模, 并在此基础上计算控制律. 而且, 该方法中引入了工业异常数据处理机制, 利用JITL学习子集中的平均数据项, 对异常数据项进行填补或替换, 从而消除异常数据对控制系统的影响. 此外, 本文提出一种JITL模型保留策略(MRS), 避免由于数据库中相似数据样本不足导致的局部模型严重失配, 并通过实时收集I/O数据更新数据库, 使控制器自适应不同的工况条件, MRS还可以有效抑制噪声干扰的影响,从而提高控制系统的稳定性. 最后, 基于某大型钢铁厂2#高炉的数值仿真实验, 充分验证了该方法的有效性.

英文摘要

      In this paper, a data-driven adaptive predictive control method based on just-in-time learning (JITL-APC) is proposed for blast furnace ironmaking process. The feature of the proposed approach is that the controller uses the k-Vector Nearest Neighbor(k-VNN) strategy to search the I/O data information in the database to establish the local model for the nonlinear system, and then calculates the control law based on the local model. Moreover, we introduce an industrial data exception handling mechanism in this method to fill or replace the abnormal data items by using the average data items in the JITL learning subset to eliminate the influence of abnormal data on the control system. Meanwhile, a JITL model retention strategy (MRS) is proposed to avoid the serious mismatch of local models caused by insufficient data samples in the database. In addition, JITL-APC updates the database by collecting I/O data in real time, so that the controller can smoothly adapt to different working conditions. MRS can also effectively suppress the influence of noise interference and improve the stability of the control system. Finally, industrial experiments have been carried out on the 2# blast furnace in a larger Iron & Steel Group Co. of China, which proves the validity of this method.