Intelligent Manufacturing Oriented Robust Optimization Methods And Algorithms For Production Scheduling

As a milestone of the development in the manufacturing industry,intelligent manufacturing has become the common goal and motivation for numerous enterprises.However,due to the existence of barriers between industries and between technical fields,the transformation towards intelligent manufacturing has brought great challenges to both engineering practice and theoretical research.In order to effectively promote the collaborative transformation of architecture,key technology and operation mode the integrated application of data,model and algorithm has become the key to enable transformation,which is also the only way to improve enterprises' abilities of perception,analysis,decision making and execution.According to the industry 4.0 reference architecture(RAMI4.0),Smart manufacturing system architecture and Smart manufacturing ecosystem(SME)proposed by Germany,China and the United States successively,enterprises have gradually focused on the manufacturing operation management(MOM)layer in the traditional enterprise pyramid.As a classic decision activity in MOM,the reliability,effectiveness and robustness of the pro-duction scheduling decisions can effectively reflect the intelligent level of manufacturing enterprises.This thesis aims to overcome the difficulty in applying mathematical models in production scheduling.Through a series of studies on data analysis,flexible robust optimization,engineering strategy,and global algorithms,the precision and practical performance of robust optimization models are improved for production scheduling.This thesis is organized as follows:At first,novel data driven uncertainty sets are defined based on probability density and probability distribution,which can precisely describe the multi source,multimodal,and correlated disturbances in the real world production process.Nonparametric methods are utilized to capture the probability distribution and correlations of uncertainties.A practical data analyzing method is proposed for estimating joint probability distribution of multi dimensional uncertainties.Convexification strategies are also developed for simplifying nonconvex uncertainty sets,which provide valid support for precise robust optimization of production scheduling problems.Then,induced by data driven uncertainty sets,novel robust formulations are proposed for dealing with independent left hand side uncertainty,matrix uncertainty,correlated uncertainty,and uncertainties with multi modal probability distributions.In the new formulations,uncertainties are independently controlled with separated confidence levels,which are defined as new adjustable parameters.We provide an estimation of the probability bounds on the violation of robust constraints,which is explicitly related to the chosen value of adjustable parameters.Thus,decision makers can customize robust scheduling models according to the expected performance of robustness and conservatism.To effectively apply the above data analyzing and modeling techniques,a practical strategy is studied,we firstly implement the proposed robust scheduling methods to an integrated production and utility system in an ethylene plant.Considering the volatility in energy consumption of furnaces,offline and online experiments are performed to analyze the perf'ormance and applicability of our robust scheduling approach.To maintain the power of our decision making strategy for a long term implementation,we design a data driven rolling-horizon robust scheduling strategy to actively integrate newly generated process data,adaptively adjust uncertainty sets,and dynamically update the online mathematical model.Another focus of this work lies on the global optimization algorithm of large scale mixed integer engineering models,which is the core engine for accelerating decision making in the process industry.A novel presolving method is developed for linear programs(LPs)and mixed integer linear programs(MILPs),which can simplify the original formulation to a smaller size with fewer constraints and variables.On the other hand,we propose optimality conditions for mixed-integer nonlinear programs(MINLP)and develop novel domain reduction algorithms through bound propagation on the subgradients of the objective and constraints.All these methods serve as preprocessing procedures in the branch and bound algorithm,which are designed for general cases and implemented in the global solver BARON.Computational improvements are observed through a wide range test on open source numerical and engineering models.At the end of this thesis,promising future interests in improving optimization theory and implementing global algorithms on engineering field are discussed.Insights to the transformation trends of industrial decision making system and reference architecture are concluded.