The Electrolyte Components' Prediction Based On Support Vector Machine

The process of electrolysis usually contains of a series of complex physical and chemical reaction occur which lead to the interactions and strong nonlinearities among the process variables. Moreover, to get better control performance, detailed electrolyte information is also required. On the other hand, the thermo technical parameters in the electrolysis process such as temperature, pressure, flow rate and liquid level are easy to measure and control, While the most important parameters such as copper concentration and vitriol concentration which have an important infect in the quality of the products, are hard to measure online. Soft sensors are more attractive as they are cheap, flexible, easy to implement and have no measuring delay. The key idea of soft-sensing is to estimate the hard-to-measure parameters using easily available process variables.This dissertation is mainly focusing on the development of soft-sensing methods based on supporting vector machines. The detailed discussions have been expanded particularly form theory study of support vector machine, choosing kernel function and parameters. Some parameter selection methods for support vector regression model have been compared in this paper. A new parameter setting methods for building SVR model was proposed based on multi-object optimal algorithms. Simulation results show that the model constructed by the optimal parameters has superior learning accuracy and generalization performance.Aiming at the problems that a black-box model is difficult to interpret and usually over depend on the training data, a parallel hybrid model is presented by combining prior process knowledge and support vector machines. It is comprised of two parts, the mechanism model based on metal balance and cubage balance of the process, and a support vector machines mode serving as a compensation of ignored process factors and production variation. The experimentation shows that the hybrid model can provide reliable predictions and good performance with less training data.The question for study is mainly aimed at the structure of soft-sensing model, the building of electrolyte mechanism model, the approach of building soft-sensing model based on support vector machine and the selection of model parameters. In addition, the research is combined with the soft sensor application in an electrolytic copper smelting plant. From the application results to a real electrolyte system, we can demonstrate the validity of this research in the prediction of electrolyte components.