Research On Application Of ESN For Identification And Control Of Chemical Process

Identification and control of chemical processes is research difficulty and hot spot in thefield of process control. Chemical processes usually have highly nonlinear and time-varyingcharacteristics, so it is difficult to develop precise mechanism model of the process. Neuralnetworks (NN) have been used widely as modeling and controlling method of nonlinearsystems for its good properties, especially the recurrent neural networks (RNN), which hasdynamic characteristics and takes obvious advantages. However, RNN needs to train all theweights in the network, leading to the complex training algorithm and difficultly applied inpractical engineering.Echo state networks (ESN) is a novel RNN which has strong nonlinear dynamicapproximation ability and good short-term memory capability because of its peculiar statereservoir (SR). Moreover, only the output weights have to be trained. Therefore, on the basisof study the classical ESN and its training algorithm, an ESN with leaky integrator (LI)neurons is introduced. For the highly nonlinear chemical processes, an ridge regressionlearning algorithm based on the LIESN is introduced. Then the LIESN is applied toidentification and control of the chemical process. The experimental results show theeffectiveness of the method in comparison to conventional methods.The main contents of this thesis includes the following aspects:(1) Mainly study the fundamental structure and learning algorithm of classical ESN.Further analyze the ESN network offline and recursive least squares (RLS) of online learningalgorithm. And basis on this, introduce LIESN with improved network structure and give theridge regression algorithm based on this network.(2) The LIESN network is applied in chemical process identification. Aiming atsteam-water heat exchanger process, the continuous stirring reactor (CSTR) process, acidbase neutralization process and ethane-ethylene distillation column process, identificationexperiment are carried out respectively. Under the same conditions, the results show that theLIESN can get higher accuracy of identification, a rapid and stable learning speed incomparison to ESN network, back propagation (BP) neural network and fuzzy neural network(FNN) et al methods.(3) To study the application in chemical process control based on the LIESN network.Mainly gives the direct inverse control and model reference control (MRC) strategies basedon the LIESN. Direct inverse control of LIESN is applied to CSTR process control. First ofall, the system model is identified by the LIESN, and the control law is designed by controllerin order to achieve effectiveness of control. Then, MRC of LIESN is applied to thesteam-water heat exchanger process control. The controller of the LIESN is identified in order to make sure that output of the closed-loop sysytem is able to track the output of the referencemodel. Under the same conditions, the experimental results show the LIESN network caneffectively reflect dynamic characteristic of the control object, get higher accuracy of controland has good performance of trajectory tracking control in comparison to radial basis function(RBF) neural network et al methods.