On-line Estimation of Process Variables in an Industrial SBR Plant Using Neural Networks

STEGMAYER, G.S.; MINARI, R.J.; GUGLIOTTA, L.M.; CHIOTTI, O.A.; VEGA, J.R.

Buenos Aires (Argentina)

Congreso; XXII Interamerican Congress of Chemical Engineering - V Argentinian Congress of Chemical Engineering; 2006

This theoretical work investigates the industrial production of Styrene-Butadiene Rubber (SBR) in a continuous reactor train operated under steady-state conditions, with the aim of developing a neural network (NN) model for estimating several process and polymer quality variables in each reactor of the train. The estimated variables are the monomer conversion (x), solids content (xsol), polymer production (G), particle diameter (dp), average copolymer composition (pS), average molecular weights (Mn and Mw) and average branching degrees (Bn3 and Bn4). To this effect, the reaction heat is assumed to be measured in all reactors. The NN models are trained and validated with a large set of simulation data generated through an available first-principles (FP) model. The investigated industrial plant (Petrobras Energía S.A., Pto. Gral. San Martín, Santa Fe, Argentina) produces SBR grade 1502 in a train of 7 identical continuous reactors, with a reaction volume of 17500 L. All the reagents (styrene, butadiene, water, initiator, chain transfer agent, and emulsifier) are continuously fed into the first reactor, and the produced synthetic latex is removed from the last reactor. The FP polymerization model by Gugliotta et al. was employed to simulate the continuous production of the SBR latex. The model was based on the fundamental physicochemical mechanisms involved in emulsion polymerization, and it was adjusted to industrial measurements of both 1502 and 1712 grades, taken from SS and transient operations. The simulation results used for training and validating the NN model were obtained by simultaneously changing two of the reagent feeds into the first reactor of the train. With respect to the base recipe, the maximum change of each reagent feed was 20%. In the simulated cases, the reaction heat was assumed to be measured in all reactors of the train. The NN model involves 16 input variables (the mass feed rates into the first reactor of all the reagents and the “measured” reaction heat rates in each reactor); and 63 output variables [x(r), xsol(r), pS(r), dp(r), G(r), Mn(r), Mw(r), Bn3(r), and Bn4(r), where r= 1, ...7 represents the reactor number]. The input variables have different sensitivity on the estimated variables, and for this reason the problem is divided and modeled with two independent neural networks. One NN model is used for estimating x(r), xsol(r), pS(r), dp(r), and G(r). The other NN model is employed for estimating Mn(r), Mw(r), Bn3(r) and Bn4(r). The proposed models are based on the multilayer perceptron (MLP) neural network, where each hidden neuron has the hyperbolic tangent as activation function and all variables are normalized in the [-1,+1] range before training, to improve model accuracy and speed up learning. The output neurons have a linear activation function which is a common choice in MLP models. During training, network parameters are optimized using the Levenberg-Marquardt algorithm, which assures good performance and speed in execution. The NN models perfectly learned the training data, and a good estimation was also obtained when data not included into the training set was given to the model. Based on the good accuracy obtained in the estimations, we foresee their use for on-line reactor control purposes, which highly depend on accurate estimates of the main process and polymer quality variables. The advantages of the NN models are its accuracy and execution speed, as opposite to traditional FP mathematical models, which are accurate but time-consuming and difficult to implement.