(302f) Modeling of a Three-Phase Industrial Batch Reactor Using a Hybrid First-Principles Neural-Network Model

The goal of this work is the development of an industrial reactor model that can be used for process analysis and scale-up effect identification in order to overcome capacity limitations. The reaction system is a large scale, complex three-phase (solid-liquid-vapor) mixture with equilibrium reactions. The reaction leads to the production of the main product (P) and to an unwanted co-product (CP). The process is operated through 2 different operational stages which increases the complexity. The nature of the information available to build up the model has two forms: information is present in form of advanced process understanding and in form of significant number of plant measurements. While most of the available process parameters are measured on-line (temperature, pressure, flows, masses), concentration measurements are taken only during plant trials. In order to combine the two knowledge sources a hybrid first principles-black box model is proposed. The first principles part of the hybrid model consists of the reaction kinetics model determined on laboratory scale, solubilization equations and component mass balances. In the developed first principles model all component formation and depletion rates are known but the rate of evacuation of CP. In order to calculate this rate, which has a time variant non-linear fashion, a black box model is used in the form of a neural network. With this neural network we are able to model the evacuation rate of CP. Since there are two process operation stages two black box models are used to model the evacuation rate. In the first stage of the process operation the accumulation level of CP in a receiver tank is measured. Using this information the rate of evacuation of CP can be included in the hybrid model in the form of an inverse model of the accumulation using a neural network (NN1). With this step we eliminated the unknown in the process model that describes operation stage 1. In the second stage the accumulation rate of CP is not available anymore; instead of this concentration measurements are taken and similarly to stage 1 the evacuation rate is an unknown parameter that needs to be identified. The calculation of evacuation rate of CP in process operation stage 2 is posed as a neural network optimization problem with the goal of minimizing the overall hybrid-model predicted concentration deviations by manipulating the weights and biases of a second neural network (NN2) subject to the constraints formulated in the given kinetic model, evacuation rate calculated from NN1 and mass balances. The optimization problem presents many local optima and as a consequence a global optimizer such as genetic algorithms was preferred. Subsequently the results were refined with a local optimizer. After the analysis of the modeling results it was found out that the small scale first principles model has to be adjusted for the large scale process operation. After this the hybrid model is re-optimized. In this work a hybrid first principles neural network model was proposed to model and analyze a three phase large scale industrial reaction system. During the modeling process the scale-up effects were identified and the serial hybrid models are proposed for the identification of first principles modeling errors.