Research On Data-driven Soft Sensor And Intelligent Optimization Methods

Soft sensor can be used to make a real-time estimate of key variables which are difficult to measure directly in the process of chemical production, so as to provide guidance for production optimization and closed-loop control. With chemical process becoming more complicated, people are expecting more progress in soft sensor technology.And so it is, soft sensor is improving in the crossing with other discipline. The focus of soft sensor technology currently includes data processing, data-driven modeling and intelligent optimization methods. In order to make a real-time estimation of propylene concentration in rectification tower and optimize the process, this paper conducts research from following directions:(1) Because modeling sample quality is closely related with the model, from the perspective of data processing, this paper emphasizes on secondary variables selection methods and clustering algorithm. After summarizing and integrating advantages of these methods, a feature weighted affine propagation clustering method based on mean impact value is put forward, the validity of which will be verified by simulation experiment.(2) In order to take soft measurement of propylene concentration in rectification tower, the paper takes deep research on data-driven modeling methods by summarizing generalized regression neural network (GRNN) and least-square support vector machine (LSSVM), then combines improved clustering algorithm and LSSVM to put forward multi-model method based on LSSVM. This method proves to be more precise than GRNN and LSSVM methods after applying to soft measurement of propylene concentration.(3)After studying genetic algorithm and particle swarm optimization algorithm, this paper makes corresponding improvements and puts forward multi-population mixed culture algorithm(MMPCA).Benchmark performance tests of unconstrained and constrained optimization problem shows that it has higher search efficiency and better stability compared to improved genetic algorithm, improved particle swarm optimization, multi-population elite sharing genetic algorithm. In order to solve process optimization problem of energy-saving and quality control in rectification tower based on data-driven model and avoid the problem of penalty coefficient selection, using "dumb value" to transform constrained problem into unconstrained problem is a good choice which turns out to be effective combined with MMPCA.