Like many globalized industries, the pulp and paper sector finds itself with an increasingly demanding clientele, who continually expect a better and cheaper product. An important design strategy being employed to address this objective is through an analysis of the vast quantity of process and product data accumulated in plant-wide data historians, in order to improve operations. Mill processes are multivariate, meaning that the interactions between the variables are as important as the variables themselves. Process relationships must therefore be modeled as a group, using an appropriate simulation technique like Multivariate Analysis (MVA), with suitable data pre-processing to account for process upsets and other disturbances. In a previous paper, using an Eastern Canadian newsprint mill as an industrial case study, we showed that it was possible to find statistically significant correlations between wood chip refiner operation, intermediate pulp quality, and final paper quality using data-driven models. This was true even though some important process parameters went unmeasured, process lags changed with time, and the operation of key equipment items changed gradually with use. The present study compares the use of different timescales and combinations of unit operations to determine which ones yield the best MVA simulations. Because plant operating data were used, and experimental design was not practical, it is possible that some of the correlations found could be attributable to coincidence. We therefore added and removed variables and time periods to explore the validity of the models. The best MVA models were obtained by using a shorter (1-hour) data timescale, although use of a weighted-average filter helped to bridge the gap between these faster readings and the slower paper quality trends.
Read full abstract