In chemical plant optimisation, different strategies can be applied, and each one tends to lead to different types of improvement because the underlying basis is different.

One approach I often use is correlation analysis, supported by heat tables, to study how each plant variable moves against the others. I usually start with several months of plant data, clean the raw dataset from NaN values, negative readings, and other abnormal points, and then build a correlation matrix. From there, I focus on variable pairs showing a correlation coefficient above 0.8. Statistically, this may not always be considered extremely strong, but in real plant operation it is often enough to signal that something important is happening.

This type of approach is one of the foundations of Advanced Process Control (APC). It can also provide valuable hints on how to improve plant performance by identifying which variable should be manipulated to obtain the desired result.

I have mentioned before the case of a “mysterious” dryer performance issue that turned out to be correlated with an increase in purge flow. At first sight, this made little sense. But once the correlation was explored more directly, it became clear that the dryer was being affected by poor product quality, which in turn was linked to catalyst quantity and correlated with the purge flow.

At the same time, correlations are often elusive. In many situations, calculations alone do not provide enough insight. Statistical relationships may be too weak, or too indirect, to explain plant behaviour properly. In those cases, performance needs to be understood through direct process analysis rather than statistical evidence alone.

This is one of the ideas behind the Glass Plant: making plant performance more visible, readable, and understandable. In some cases, this may also need to be complemented by literature research, especially when the problem has already been studied elsewhere.