Neural network and model-predictive control for continuous neutralization reactor operation

This paper outlines neural network non-linear models to predict moisture in real time as a virtual on line analyzer (VOA). The objective is to reduce the moisture variability in a continuous neutralization reactor by implementing a model-predictive control (MPQ to manipulate the water addition. The acid-base reaction takes place in right balance of raw materials. The moisture control is essential to the reaction yield and avoids downstream process constraints. The first modeling step was to define variables that have statistical correlation and high effect on the predictable one (moisture). Then, it was selected enough historical data that represents the plant operation in long term. Outliers like plant shutdowns, downtimes or non-usual events were removed from the database. The VOA model was built by training the digital control system neural block using those historical data. The MPC was implemented considering constraints and disturbances variables to establish the process control strategy. Constraints were configured to avoid damages in equipments. Disturbances were defined to cause feed forward action. The MPC receives the predictable moisture from VOA and anticipates the water addition control. This process is monitored via computer graphic displays. The project achieved a significant reduction in moisture variability and eliminated off-grade products.