Continuous Data Analytics – Quality Parameter Prediction

The prediction of quality parameters using Partial Least Squares (PLS) models is often the preferred technique when dealing with collinear data where multiple process input changes impact similar process output measurements. Predictions based on this linear technique can be used to extrapolate outside the data range used in model development. Process interactions may be accounted for using principal component analysis (PCA). Through the use of PCA the variables associated with a process may be represented by a few principal components. This allows the dimensions of the data set to be reduced and also eliminates the effects of the interaction between variables. As address in Advanced Control Foundation – Tools, Techniques, and Applications, through the application of Hotelling’s T2 statistic and the Q statistic to the reduced data set, it is possible to detect abnormal operation resulting from both measured and unmeasured/un-modeled disturbances. However, since the underlying technology is based on the deviation of process measurements from their mean value this must be taken into account during model development and in the on-line application of these technologies.

When applying continuous data analytics the application must account for changes in the mean value of the measurements that occur with changes in plant operation such as the production rate. For example, a swing boiler in the power house area of a plant must constantly respond to changes in throughput demand set by the plant master control as shown below.

Changes in the mean value of measurements may be accounted for by assigning specific states to a predefined process operating condition. The mean measurement values for the example boiler application are show below where boiler demand is selected as the state parameter broken into five operating regions.

The value of the state parameter and knowledge of how the measurement mean values change with the state parameter (based on analysis of plant operating data) can be used to automatically modify measurement mean values used in PCA and PLS analysis.