Data And Model Driven Intelligent Optimization Of Complex Industrial Processes And Its Applications

Complex industrial processes have the problems of complicated reaction mechanism,great fluctuation in the ingredient of raw material,large time delay,strong inner coupling and so on.As a result,it is hard to build mathematical models of the optimization problems in the complex industrial processes.For example,the fused magnesium furnace is charged intermittently and the electrodes are buried in it.It is hard to build an optimization model for the energy consumption per ton when the average current is taken as the decision variable.In addition,complex industrial processes include many manufacturing facilities or processes and face the challenges of benefits,quality,environment,safety and so on.For example,the multi-furnace production mode is adopted by the magnesia enterprises,so the optimal decision-making for the average current of every furnace is related to multiple conflicted objectives,such as total output,total energy consumption and high-quality rate.The optimal decision-making for the average current of all furnaces is a high-dimensional multi-objective problem(MOP).As a result,we should study the intelligent optimization method using data and model driven and apply it to solve the optimal decision-making problem of the average current of the fused magnesium furnace.Supported by the National 973 Program "Comprehensive Production and Process Integrated Control System Overall Control Strategy and Operation Control Method(2009CB320601)"and National Natural Science Foundation Project "Overall Optimization Control Research Based on Data and Model for Complex Industrial Processes(61020106003)",we study the intelligent optimization methods using data and model driven based on the background of the fused magnesium furnaces.The main work has been summarized as follows:1)Surrogate models of the existing evolutionary algorithms(EAs)are modified,and three surrogate-assisted EAs(SAEAs)are proposed to solve optimization problems without mathematical models,high-dimensional optimization problems and high-dimensional MOPs.Benchmark problems are used to compare the proposed algorithms with the existing EAs,and the results verify the effectiveness of the proposed algorithms.The main work are summarized as follows:a)EAs usually solve optimization problems without mathematical models based on surrogate models built by historical data.A modified surrogate model combing a polynomial model and a Gaussian process(GP)model is proposed to improve the performance of SAEAs,and a multi-objective EA(MOEA)assisted by the modified surrogate model is proposed.In the proposed algorithm,the GP model is made use of to screen the solutions found by the MOEA and the low-order polynomial model is made use of to evaluate the promising solutions found by the GP.The alternation between the GP model and the polynomial model is repeated many times until the stopping criterion is reached.Nine benchmark problems are used to compare the proposed algorithm with a MOEA assisted by the polynomial model and a MOEA assisted by the GP model.The results show that the proposed algorithm obtains the best values of hypervolume indicator,which demonstrates that the proposed algorithm is better than the other two compared algorithms.b)A new surrogate,a heterogeneous ensemble model,is proposed for the issue that the uncertainty estimations of GPs are inaccurate on high-dimensional optimization problems.In the proposed heterogeneous ensemble model,feature selection and feature extraction are utilized to transform the decision variables,and the least square support vector machine and the radial basis function networks are adopted to build member models based on the original decision variables,a selected subset of the decision variables and a set of extracted decision variables.The heterogeneous ensemble model is a weighted aggregation of nine different member models.A MOEA assisted by the heterogeneous ensemble is proposed.Sixteen high-dimensional benchmark problems are used to compare the proposed algorithm with a MOEA assisted by the GP model.The results demonstrate that the proposed algorithm performs better than the GP-assisted MOEA on high-dimensional optimization problems and the proposed algorithm reduces the computational time.c)A new surrogate,an efficient dropout neural network(EDN),is proposed to solve the high-dimensional MOPs.The proposed EDN randomly ignores neurons in both its training time and test time,and as a result,EDN can estimate the fitness values and the confidence level of the estimated fitness.Compared with the conventional dropout neural networks for uncertainty estimations,the proposed EDN does not need to store and update a large number of neural networks during the optimization,thereby significantly reducing the computational cost.A MOEA assisted by EDN is proposed.Sixteen high-dimensional multi-objective benchmark problems are used to compare the proposed algorithm with a MOEA assisted by the GP model and a MOEA assisted by the heterogeneous ensemble model.The results demonstrate that the proposed algorithm is better than the other two compared algorithms on high-dimensional MOPs in both performance and computational time.2)An intelligent optimization method using data and model driven is proposed to solve the optimization problem of ECT of a fused magnesium furnace based on the modified S AEAs above.The simulation experiments are conducted by the real data,and the results demonstrate that the proposed intelligent optimization method performs better than the EAs assisted by modified surrogates.a)Based on the ECT static model for a production batch,an ECT dynamic model taking the average current as the input is built for different production batches.The prediction model for ECT is built based on system identification and stochastic configuration network.b)A new surrogate combining ECT prediction model and the GP model is proposed,and an intelligent optimization method using data and model driven is proposed based on the new surrogate.c)The simulation experiments are conducted by the real data to compare the proposed intelligent optimization method and the EAs assisted by modified surrogates.The results demonstrate that the proposed intelligent optimization method can find the optimal average current more quickly.