Technology Division - Neural Network Propcess Modeling Optimization

Process Modeling and Optimization

This article describes a methodology for optimizing a process using a mathematical model. This methodology is also demonstrated using a Neural Network based model of a chemical plant.

Artificial Neural Networks - Overview
Chemical Process Optimization Demo

Process Modeling

A Process can be an industrial plant such as a water treatment, chemical, or a food processing plant. Or a process can be a component within a plant such as a turbine, a heat exchanger, a dryer, etc. .

Regardless of what a process is composed of, it is characterized by having inputs, and outputs.

The objective of process modeling is to build a mathematical system that behaves similar to an actual plant, i.e., the mathematical system is to produce similar outputs as an actual plant given the same inputs.

A common way to build a plant model is to provide the inputs of the actual plant to the model. Then compare the outputs of the actual plant to the output of the model. This results in a modeling error that is used to adapt the mathematical plant model in a direction that will minimize the modeling error. The mechanism of adapting the mathematical model is also referred to as model training.

After training the plant model for several cycles with different inputs, and when the modeling error is sufficiently small, we will have a mathematical plant model that behaves very similar to the actual plant.

Process Optimization

Now that we have a plant model that mimics the behavior of an actual plant, we can use the model to optimize the actual plant.

For example, let us consider chemical plant. The inputs to the plants are several reactants. You mix together the different reactants to produce a product of a certain property. You can produce the same product by combining infinitely different proportions of the reactants. The objective of process optimization is to find the optimum reactant proportions in order to produce the desired product with a most cost effective manner.

One way to find the the optimal proportions of the reactants is to perform a Monte Carlo simulation on the process model.

Vary the proportions of the reactants randomly and observe the output of the model. If the output of the model is similar to the desired product, then we have a possible solution.
We next compute the cost of the reactants that produced this product.
We remember (store) the reactant proportions that produce the least cost for producing our desired product.

We iterate steps 1, 2, and 3 thousand, maybe millions of time until we no longer are able to reduce the cost of the reactants. We then have a solution for our process optimization. We have the optimum proportions of reactants that produce our desired product the cheapest way possible.

Note that we are doing the steps 1, 2, and 3 iterations on the mathematical model, so we can afford to do thousand of iterations in a second.