What’s The Value of Inferential Models?

Inferential models are process engineering calculations, analyzing operating conditions of a unit to come up with an estimate of product quality.  Achieving a reliable inference model is easier said than done.  The modeler must understand the unit in question, and even then, model development takes engineering time, as well as lab support.  And that is only the beginning.  Once an inferential model becomes available, it would be necessary to monitor model predictions against lab data, applying occasional model maintenance required to keep the inferential fidelity high. 

Is this effort justified?  Can inferential model benefits be quantified?  APC makes money by pushing throughput or valuable product yields against constraints.

Income = Throughput * { ∑ Product yields * product price – feed price}

Example

Consider a typical CDU summer economics, maximization of throughput against condenser constraint.  The CDU fractionator receives hot vapor from the heater, and makes use of several cooling circuits: reflux and pumparounds, for gradual cooling and condensation of the products.  On hot days the cooling circuits are fully loaded.  Upon further throughput increase, condenser load and hence overhead drum temperature would rise.  LPG, normally nearly fully condensed, would remain partly in the vapor phase, loading up the overhead compressor.  Fractionator pressure is controlled by spilling back compressed vapor, and when vapor load goes up, at some point the spill-back valve is fully closed and pressure starts floating. 

APC action

Product demand is high in summer and APC would first, drive up all cooling circuits, and second, manipulate the throughput, keeping pressure control spill-back valve almost fully shut.  Throughput is maximized, going up or down with weather conditions.  It is no longer constant. 

Throughout that time, products properties must be kept within specifications and product yields optimized.  How would APC guarantee that?  Lab samples are taken at 6 am, results come back about 10 am, but by then the sun is out and unit conditions have changed.  The throughput maximization trend below illustrates what APC can achieve.  Throughput (blue line) can indeed be maximized, subject to weather and other constraints.  When inferential models are available (green line), APC can also maintain product qualities at target, optimizing both throughput and yields.  However, when inferential models are not available, APC (or the operator) must resort to giveaway operation, shown by the brown line.  While throughput is maximized, product yields cannot be optimized. 

CDU benefit estimate

What is the monetary difference between the green versus brown lines of this trend?  For crude units optimal versus approximate naptha/kerosene/diesel splitting is in the order of millions of dollars annually.  That is the value of a handful of good inference models, perhaps a million dollars per inference. At the end of the day APC is as good as the inferential models supporting it.