Research Of Aluminum Electrolysis Process Modeling And Control Based On Least Squares Support Vector Machine

Aluminum electrolysis process is a nonlinear, multivariable, large time-delaying and complex industrial object, some of whose important technological parameters are difficult to measure on line, so traditional controlling methods, such as PID control and adaptive control can not achieve ideal effect. However, least squares support vector machine (LS-SVM) is a novel general learning method emerging from machine learning field in recent years, which is based on structural risk minimization principle, and can better solve some practical problems like small sample, nonlinear, high dimension and local minimum, etc. Thus it has been successfully applied in many fields, like classification, function approximation and time series prediction. In this dissertation, a study of modeling and control based on LS-SVM is made first and then part of the research results is applied in aluminum electrolysis process and has achieved satisfactory results.In this paper, four main tasks have been performed.(1) Electrolytic temperature, alumina concentration and electrode distance are three very important but hard to measure parameters. To this problem, a PSO LS-SVM algorithm based on LS-SVM and particle swarm optimization (PSO) is proposed to establish the soft-sensing models of the three parameters. Considering how to choose optimal LS-SVM algorithm parameters, the algorithm defines arithmetic mean of quadratic sum of predictive errors as fitness function firstly. Then, it applies PSO technique to iterate and search in the feasible regions so as to decrease the fitness value. Finally, it returns the optimal LS-SVM algorithm parameters and corresponding model parameters. The simulation results show that the soft-sensing models based on PSO LS-SVM algorithm have both stronger learning ability and stronger generalization ability than that of the soft-sensing models based on neural network (NN), proving the validity and advantage of the proposed algorithm.(2) A study on predictive control based on LS-SVM is done. To the multi-input multi-output and constrained nonlinear system, a kind of two-level control structure is proposed, and a single-step predictive control algorithm based on LS-SVM called CHAOS MPC algorithm is deduced. Considering the constraint of control variable, the algorithm applies chaos optimization technique to search in the feasible region to obtain the optimal predictive control law in real time. The simulation results show that the control precision of the algorithm is higher than that of the single-step predictive control algorithm based on NN.Similarly, considering multi-input multi-output, constrained nonlinear and time-delaying system, a kind of two-level controlling structure is proposed, and a multi-step predictive control algorithm based on LS-SVM and chaos optimization is deduced, which is called CHAOS MPC₁algorithm. The algorithm applies chaos optimization technique to solve the optimal predictive controlling law in real time. The simulation results show that the controlling precision of the algorithm is higher than that of the multi-step predictive control algorithm based on NN.(3) After research on the stability of predictive control system based on LS-SVM is done, a double mode control algorithm is proposed, which can guarantee closed-loop system stability. First, a terminal constraint is appended at the back of traditional performance index, and corresponding stability theorem is deduced by Lyapunov method. Then, according to the stability theorem, a double mode control algorithm is put forward. The algorithm first applies predictive control to drive the state to the terminal constraint set, then it uses local linear control to reduce computation cost and to drive the state to the origin finally. The simulation results prove the validity and advantage of the proposed algorithm.(4) Part of the research results are applied in "Design and development of advanced control system of aluminum electrolysis process" horizontal topic, and satisfactory results are achieved. First, an alumina concentration control strategy is proposed, which applies PSO LS-SVM algorithm to establish predictive model of alumina concentration, and applies CHAOS MPC algorithm to implement predictive control. Then a cell resistance controlling strategy based on cell resistance filtering and expert experience is presented. Finally, a cell status analytical and maintenance expert system based on LS-SVM is raised, which integrates LS-SVM, fuzzy control and expert system. System operation results show that the above methods can enhance the current efficiency by2.1%, and reduce DC power consumption by338kW.h/t-Al. That is to say, compared with the original methods, the energy saving effect of the proposed methods is obvious.