Model Reconstruction And Optimization Acceleration With Microstructural Quality Indices

Product quality issues have always been a bottleneck restricting the competitiveness of process enterprises in China.Based on existing equipment and process conditions,how to carry out fine modeling,design optimization,and control for product quality is of great significance.It could lead to a high-end product value chain.Compared with traditional macro quality indices,the molecular structure of polymers essentially determines the end-use properties of products.Therefore,the optimal control oriented by the micro quality indices becomes the key to solving the quality problems of products.However,models involving microstructural indices are often complex in structure,strong in nonlinearity,large in scale,and difficult to solve,which poses great challenges for process optimization with requirements on microstructural indices.This thesis focuses on resolving a series of key issues in process systems optimization guided by microstructural indices of products.It combines deterministic models and statistic models,and develops model reconstruction methods and solution acceleration methods for industrial optimization problems.The aim is to realize accurate control on product quality and production efficiency improvement for polymerization processes.The main research contents and contributions are summarized as follows:1.Parallel solution method for large-scale equation oriented models.The equation-oriented(EO)models are widely used for process simulation and optimization.But its solution procedure is a challenging and time-consuming task,which makes it difficult to apply in time-sensitive cases.Based on successive quadratic programming algorithm,a parallel function evaluation method for solution to large-scale EO models is proposed.This method utilizes both modern multi-core processor and GPU technology,both of which implement accelerations for numerical solution.Guidelines are given for selecting applicable platforms based on the model characteristics.2.Parallel-Monte-Carlo-simulation based optimization method.Monte Carlo simulation is a typical method to calculate microstructural quality indices.But this method is computationally expensive,unable to obtain derivatives,making it difficult to be optimized.Based on GPU parallel technology,this project proposes a systematic optimization method for Monte-Carlo-embedded models,which includes adaptive simulation algorithm and successive boundary shrinkage method.By adjusting the simulation accuracy to meet the optimization requirement,this systematic method can significantly speedup the optimization procedure.Besides,through successive sub-problem solutions,it can enhance the optimization convergence.3.Adaptive acceleration approach for Monte Carlo models.An adaptive re-sampling method is proposed to deal with the inefficiency issue of Monte Carlo simulation.Inspired by the Bootstrap method,this method combines the molecular simulation and the resampling with replacement together to realize the chain growth acceleration and the number of simulated chains reduction.Without essentially losing accuracy,this method achieves good accelerations for calculating microstructural indices on steady state.Both simulation and optimization cases based on microstructures are presented.4.Model reduction method using orthogonal collocation for dynamic population balance models.The complete population balance models are commonly used to describe microstructures.But for dynamic process,they consist of a huge set of algebraic differential equations,which are difficult to solve.This project exploits the specific structure of population balance equations,and proposes a model reconstruction method for dynamic molecular weight distribution(MWD).This method uses orthogonal collocation on finite elements in two dimensions to capture both the MWD’s dynamic feature in time and its distributive feature in chain length.The resulting system is a solvable nonlinear programming(NLP)problem,which provides an efficient way for simulation and optimization with dynamic MWD.5.Dynamic optimization for grade transition processes with MWD.Dynamic optimization for grade transition based on MWD is highly important,but challenging.On the basis of model reconstruction on dynamic MWD,this project proposes a simultaneous dynamic optimization method for grade transition.This method includes single-stage formulation and multistage formulation.As compact NLP problems,both formulations avoid stability determination with derivatives,and achieve direct minimization in transition time and off-spec production during grade transition process.