Study On Multi-Model Optimization Control In Reheat Furnace Based On Population Search Strategy

A billet reheating furnace is a key apparatus on a steel rollingproduction line, and is also the apparatus that consumes the most energy.The level of automatic control of the burning process directly determinesthe performance (energy consumption, scaling loss, percentage of scrap,quantity and quality of finished product, etc.). A conventional PIDcontrol method is difficult to handle the frequent changes in the operatingconditions of the reheating process, and thus cannot provide the precisionand response speed required for temperature control in an actual plant.Various methods of solving the optimal control problem for the burningprocess have been reported; they involve the modeling and control of theprocess and optimization of the reference temperature. However, sincethe models used in those methods employ linear partial differentialequations to describe the energy balance during heat transfer and makemany hypotheses and approximations, they are not very precise and donot adequately describe the dynamics of the process. When theoptimization and control of furnace temperatures is based mainly on animprecise static model, it is difficult to improve the control performance.To solve the problems with current methods, this dissertation presents anoptimization and control method for the reheating furnace based on apopulation search strategy, and employs the latest research results onintelligent control theory. It solves the problems of the modeling andadaptive control of furnace temperatures, and the optimization of theset-point temperatures of the billet reheating process so as to minimizeenergy consumption and the difference between the technical parametersof the billet and the required specifications.In this study, first, the problem of modeling a reheating furnace wasinvestigated, and three recurrent-neural-network (RNN) models and aheat transfer model were constructed. In order to realize the optimizationcontrol of the burning process, a recurrent neural network (RNN) basedon a back-propagation (BP) algorithm was built to predict thetemperatures of three zones of the furnace; it takes the features of theprocess (nonlinearity, a large time delay, dynamic changes, etc.) into account. Furthermore, through an analysis of the convergence of thebackpropagation (BP) algorithm, it was proven that the BP algorithmdoes not guarantee a global optimal solution, and the conditions formaintaining convergence in the BP algorithm were specified. In order tooptimize the set-point temperatures of the furnace, a heat transfer modelwas built to predict the temperature distribution of the billet. On the otherhand, in order to implement real-time control and optimization of thefurnace temperature, two RNN models based on a radial-basis-function(RBF) were built that predict the temperatures of the furnace and billet: asequential learning algorithm provides rapid adjustment of the RNNparameters; a growing-and-pruning (GAP) algorithm provides structuraloptimization; and an extended Kalman filter (EKF) for RBF-based RNNnetworks provides high-precision learning. The integration of the heattransfer model and the GAP-EKF RNN model yielded a hybrid predictorof the billet temperatures. The validity of the method was demonstratedby verification tests using actual operational data.Second, a multi-model temperature control strategy was devisedusing the following methods: decoupling control, multi-variable Smithcompensation, double cross and bounding, and a self-adaptive method.These methods were combined into a decoupled self-adaptive PID controlmethod. Since this method is based on a precise linear model, it isdifficult to obtain good control performance when there are large changesin the operating conditions. To deal with this situation, an expert fuzzycontrol scheme that is not based on a mathematical model was devised byintegrating the fuzzy method and expert control; and the stability of thisscheme was examined. To improve the adaptability of the expert fuzzycontroller, a fuzzy-neural-network (FNN) controller was constructedbased on the equivalence of an NN and a fuzzy system. The stability ofan FNN control system for a nonlinear plant was analyzed; and inparticular, the stabilizing conditions for the whole system were derivedfor a second-order plant. Then, a self-adaptive FNN control system wasconstructed by combining rule optimization and an immune-clonal-evolution (ICE) algorithm; and the global convergence of the immunealgorithm was examined. Finally, decoupled self-adaptive PID controland self-adaptive FNN control were integrated into a multi-model temperature control strategy that provides high precision under stableoperating conditions and global stability under changing operatingconditions. Simulations and actual runs demonstrated its validity. In orderto further improve the control performance, a tracking control schemebased on valve opening degree for flux was devised to handle thetime-varying input-output characteristics of the air and gas valves. Basedon fuzzy rules and an expert system, a strategy for the self-optimizationof the air-gas ratio was developed. In addition, an NN model was used toimplement an intelligent adaptive-speed control method that maps thetotal air flux and pressure of the main pipeline to the blower speed; itreduces the electrical-energy consumption of the blower.Third, based on an analysis of the optimal set-point temperatures ofthe billet reheating process, the multi-objective optimization problem ofminimizing the energy consumption and the difference between thetechnical parameters of the billet and the required specifications wasconsidered. A hybrid particle-swarm optimization (HPSO) algorithm wasused to search for the optimal reference input for the three furnacetemperatures. It employs a chaotic mechanism to obtain the initialpopulation, and also the message-handling mechanism of an immunealgorithm and clonal selection to improve the search precision. Optimalset-point temperatures are obtained by solving this optimization problemusing the FNN control model, predictors of billet and furnacetemperatures, and the HPSO algorithm. Simulation results demonstratedthe effectiveness of the method. This optimization method eliminates theproblems of blindness and randomness in producing set-pointtemperatures.Fourth, a distributed control system (DCS) with two-layerarchitecture and a monitor program was built. It features high-levelinformation management, centralized monitoring, optimal control, andfield-level automation. The process models mentioned above,multi-model control methods, and the HPSO algorithm for thetemperature setpoint are all integrated in the DCS to yield a multi-modelcontrol system based on population search strategy for a regenerativereheating furnace that provides global multi-objective optimal control.This intelligent control method improves the overall automatic- control level of the reheating process; reduces the scale loss, energyconsumption, and pollution; improves the quality of the product; andincreases the information management capability. It not only reduces thework load of operators, but also provides significant economic and socialbenefits. The modeling, optimization and control methods developed inthis study are an instructive example that can be adapted to other complexindustrial processes.