Just-in-time Learning Based Adaptive Predictive Control Of Molten Iron Quality In Blast Furnace Ironmaking

Iron and steel making are one of the most fundamental industries in the modern world,and it is also the main industry of China’s national economy.The steel industry is vital to the country’s social and economic development.Ironmaking blast furnace(BF)is a key unit that consumes more than 70%of the energy in the whole steelmaking processes.The blast furnace ironmaking is still at the stage of manual operation.The blast furnace is prone to failure under manual operation,which results in huge economic losses due to its complicated internal environment.Therefore,promoting the automatic control of the blast furnace ironmaking process is the main goal to further achieve stable operation of the blast furnace.The molten iron quality can display the internal state of the blast furnace and the relevant physical and chemical information during the blast furnace ironmaking process.Through the molten iron quality,it is possible to understand the operating status of the blast furnace more comprehensively,and provide guidance for the automatic control of the blast furnace.Therefore,in order to achieve stable operation of the blast furnace ironmaking process and to produce qualified molten iron products,it is necessary to effectively monitor and control the molten iron quality.However,the operation of BF system is further complicated by the interacting effects of gas-solid,gas-liquid,solid-solid,and solid-liquid phases accompanied simultaneously with multiphase coupling and multiphysics field coexisting,making it difficult to establish an accurate model of the molten iron quality mechanism.In addition,due to the characteristics of slow time-varying,non-linear,multi-output coupling and other characteristics presented in the blast furnace ironmaking process,the traditional control method based on the offline global model has a poor control effect.Considering above problems,this paper carries out the research and application on just-in-time learning baesd adaptive predictive control of Molten Iron Quality Indices in Blast ironmaking process Furnace under the framework of local learning theory and conduct experiment on#2 BF of Liuzhou Iron and Steel Company in Guangxi province with the support of National Natural Science Foundations.Specific works are as follows:1.Aiming at the control problem of key process index[Si]in the blast furnace ironmaking process,a adaptive predictive control method for metal iron silicon content based on linear just-in-time learning(JITL-ARX-APC)is proposed.The method is characterized in that the controller searches the I/O data information in the database through the K-VNN method,performs local linear modeling on the nonlinear system,and calculates the control law based on the linear predictive control theory.In addition,an industrial anomaly data processing mechanism is proposed to use the average data items in the JITL learning subset to fill or replace the anomalous data items,thereby eliminating the impact of abnormal data on the control system.Finally,based on the numerical simulation experiment the effectiveness of the method is fully verified.2.In order to further improve the nonlinear fitting ability of the just-in-time learning local model,a JITL-RLSSVR based adaptive predictive control(JITL-RLSSVRAPC)method is proposed by combining the LSSVR nonlinear modeling method with the just-in-time learning algorithm.Furthermore,the nonlinear predictive controller is designed by using the global convergence and super linear convergence speed sequential quadratic programming(SQP)algorithm.Firstly,the similar learning subset is obtained from the database query,and the LSSVR model parameters are updated by using the data samples in the learning subset.Secondly,in order to prevent the model structure from being too complicated,an LSSVR model pruning strategy is adopted to delete the overall performance of the model.Meanwhile,minimize the redundant samples to minimize the promotion performance of the model.Finally,through the industrial test and comparison verification of large 2#blast furnace,the proposed method has good control performance for nonlinear controlled objects.Based on the proposed JIT-RLSSVR design predictive controller,the prediction model has higher prediction accuracy and better control performance.3.In view of the control problem of metal iron quality indexes,and in order to solve the difficulty of multi-output decoupling control,the adaptive predictive control method based on recursive multi-output LSSVR just-in-time learning inverse model is proposed.First,in order to add or delete modeling data samples through online recursion,the MLSSVR incremental learning formula and the MLSSVR decrement learning formula are derived based on the multi-output LSSVR(MLSSVR)model.Then,in order to speed up the model verification speed in the just-in-time learning process,the MLSSVR fast leave-one-out cross-validation formula(MLSSVR-FLOO)was further derived.Meanwhile,the recursive MLSSVR just-in-time learning strategy is used to establish the online prediction model and inverse system model of the process,and a pseudo-linear system is constructed to realize the decoupling control of the quality of multiple molten irons,while improving the model’s online learning ability.Finally,the control tests shows that the proposed control method has good adaptive and decoupling control performance.