Empowering Process Control: JMP for Predictive Models for Optimization (2025-US-PO-2496)

In oil refining, a key optimization technique is to provide continuous online control to quality variables with infrequent sampling. It requires an empirical predictive model of the quality variable based on online process conditions. This model is then used as an inferred variable, which can be utilized in a continuous control algorithm.

There are several pitfalls where models with strong fit statistics fail to accurately predict the quality variable when deployed online. Many of the causes of model failure are outside of the engineer’s control, such as abnormal process conditions and input signal failures. A predictive model can be built that minimizes the impact of process externalities, but often this can hinder model accuracy. It is therefore the engineer’s challenge to construct predictive models that balance accuracy with reliability.

This paper explores several steps in the model development process that use JMP to build predictive models suited for online process control. By using a custom add-in, it's possible to aggregate raw process data into a format that filters process variability. The aggregated data is screened using such tools as Distribution and Screen Outliers to check for outliers and operating ranges. Model inputs are transformed with column formulas to further reduce process variability. Finally, the model is constructed using Generalized Regression in Fit Model to compare regression algorithms and minimize inputs, enhancing reliability. The end result is a model that provides enough accuracy to continuously optimize the process to a quality constraint, while remaining robust against process/instrumentation disturbances.