|
|
Prediction of Wood Pulp K# With Neural Networks Collaborators : University of Maine and S.D. Warren Company Contact: Mohamad Musavi Project Summary The technical objective of this proposal is to develop a neural network approach for prediction and characterization of continuous wood pulp digester delignification, K#, and to integrate the results in the operation of an industrial digester system. The overall objective is to create a university-industry linkage for transfer of emerging intelligent technologies and development of human resources in a vital industry to Maine. In the pulp making process, the degree to which the raw chips are pulped is estimated by measuring the K# which is related to the lignin content remaining in the pulp. Availability of an accurate K# during any time of digester operation is very critical in reducing variability and significant savings could be achieved through reductions in energy and raw material consumption. These savings would not only impact the pulp digester itself, but would cascade throughout the entire papermaking process to the Bleach Plant, and the Paper Machines themselves since all these areas are downstream of the pulp digester and rely on consistent quality pulp to run effectively. To assure quality control, K# is normally measured by hourly sampling and subsequent lab testing. Unfortunately, the process time delays between the resulting K# at the Digester blowline and the key process variables, such as upper and lower liquor heater temperatures, are between 1 and 5 hours. Even then, the feedback only comes once an hour. Also, the testing itself is subject to approximately ±5% error. As a result, adjustments are made very cautiously so as not to overreact to transient or erroneous information. This leads to significantly less optimal operation. Ideally, one wishes to predict the K# for any given set of process variables before, or as soon as, the raw materials are entered into the digester so that any corrective action can be taken immediately without having to wait a few hours later when the actual pulp is ready for laboratory measurement. Currently, process operators do their best to adjust process variables based on experience. A useful model of K# would allow operators to test several process moves off-line, see the predicted K# which results, and tailor their moves to optimize K# and minimize variability. Along with the prediction, one would also wish to characterize K# in terms of process variables so that a more efficient long-term control strategy can be planned. The current proposal will address both issues of prediction and characterization of K# using adaptive and parallel computational properties of artificial neural networks. In our preliminary investigations, a radial basis function (RBF) neural predictive model was implemented. This network was trained with 350 hours and tested with 325 hours of real digester operation. Excellent results were achieved during the preliminary testing. The neural network and computer expertise will be provided by the Department of Electrical & Computer Engineering and our industrial partner, S.D. Warren Company, Skowhegan, Maine, will provide the data, experience, and direction to make a practical tool out of this research to benefit the entire pulp and paper industry. |
|
[Overview]
- [Goals] - [Faculty] - [Students] - [Industrial
Partners] |