Modeling and simulation of pulp and paper quality characteristics using neural networks.

Application: A neural network modeling module has been implemented within a commercial simulation software system.

Simulation of pulp and paper processes has become a very valuable tool to the pulp and paper industries. Whether used for predictive analysis of the consequences of process configuration changes or for process optimization by virtually changing combinations of process parameter settings, or simply for gaining better understanding of an existing process, simulation is now one of the most powerful tools for these purposes. But what can we simulate exactly? Most commercial simulators are limited to the static treatment of a subset of process variables that explicitly appear in mass and energy balance equations. Other more versatile software will also handle the time dimension (dynamic simulation) in their balance equations.

Numerical methods are then used for solving the resulting differential equations pertaining to certain types of process modules such as tanks and controllers. However, an important category of process variables often used as qualitative metrics in process performance evaluation and optimization is usually absent from these models, due to the fact that they do not lend themselves to treatment in the typical mass and energy balance equations of most simulators. This paper presents a potential solution to this problem and describes the development of a neural network-based module that can be independent of the classic heat and material balance of process variables upon which the model is based.

Neural computing

A neural network is generally composed of a layer of input cells, one or two layers of hidden layer cells (called hiddenA and hiddenB in our work), and a layer of output cells. Every cell on hiddenA is connected to the input cells through the use of synaptic weights. Weights also appear between hiddenB cells and hiddenA cells, and between output cells and hiddenB cells, as shown Fig. 1. A network with no hiddenB layer is only connected with weights between hiddenA cells and input cells and between outputs cells and hiddenA cells.

[FIGURE 1 OMITTED]

Learning an input-output mapping for a particular problem lies in the convergence of the weight vector toward an optimal set of values. This minimizes some error criteria between the desired input-output mapping and the one returned by the network. Weights update, which make up the convergence, is typically performed using gradient descent methods; this therefore implies that the network can get trapped in local minima regions of the error surface. For the same problem, different neural networks should therefore be trained with different parameters settings, and the network with the most satisfactory results should be chosen.

Conclusions

One of the major characteristics of a neural network modeling module within a commercial simulation software system is its ease of use. The user simply needs to provide training patterns in the form of a text file, configure the network using the developed learning interface, and import the icon representing the training results into its simulation to integrate the neural network computations with those of other simulated modules and variables. The simulator can then track these variables through process streams and equipment.

Laperriere works at Universite du Quebec a Trois-Rivieres, Pulp and Paper Research Center, CP. 500, Trois-Rivieres, QC G9A 5H7, Canada. Wasik works at Aurel Systems Inc., 7197 Ridgeview Dr., Burnaby, BC V5A 4S1, Canada. Address correspondence to Laperriere by email at [e-mail].