Research On The Fusion Method Of Integer Optimization And Data Analytics For Scheduling Problem Of Batch Annealing Operation

In this thesis,taking the batch annealing operation at cold rolling stage in steel plants as background,the scheduling problem for batch annealing operation is studied.The scheduling problem for batch annealing operation is to group the coils into batches,one of which is allocated to a suitable empty batch annealing furnace,so as to maximize the charging weight and minimize the annealing difference between coils.Different from the previous research,some data analytics methods are adopted to improve the integer optimization method,and a data analytics based branch-and-price algorithm is designed.Firstly,by analyzing the historical scheduling scenarios,a heuristic is designed for the scheduling problem,which can be used to effectively generate initial columns for branch-and-price algorithm.Then,a SVR method is combined with strong branching strategy to improve the efficiency of branching variable selection.Finally,Q-learning method and Long-Short term memory method(LSTM)are adopted to improve the efficiency of branching node selection,and hence speed up the convergence of the branch-and-price algorithm.The main contents of this thesis are as follows:(1)A scheduling method based on data analytics is proposed to solve the scheduling problem for the batch annealing operation.In the basis of the correlation analysis,the characteristics of coils including the annealing curve,width,thickness,and outer diameter are selected as the input feature variables,and the coil-coil matching degree and the priorities of coils are obtained by support vector regression method.In the basis of analysis results,a scheduling method is designed to quickly generate the scheduling scenarios for the batch annealing operation.The computational experiments demonstrated the effectiveness of the proposed method.(2)In order to speed up the branch-and-price algorithm for the scheduling problem for the batch annealing operation,a branch-and-price algorithm based on data analytics is proposed.Through correlation analysis and normalization processing,a set of key processing parameters are selected as static features,meanwhile some dual variables and iterative parameters are selected as dynamic features.Then,combining with strong branch strategy,the support vector regression method is adopted to predict the upper bound corresponding to different branching variables,so as to improve branching efficiency.Computational experiments show that the proposed branching variables selection strategy can effectively improve the efficiency of the algorithm.(3)In order to improve the efficiency of branching process,Q learning method and LSTM method are proposed to select the branching node.In the branching process,the LSTM is used to learn the branching process and hence predict the function value of the given branching node.By introducing idea of the Q learning method,the predicted function value of the child node can be seen as the return value of its father node.Then,the node with the largest cumulative return value is selected as the new branching node.The new branch node selection strategy is compared with the existing depth-first strategy,the comparison results show the effectiveness of the proposed improved strategy.