Soft sensors

Analysers and Inferentials

For most processes, there is a mix of analyser and laboratory based product property feedback being used to direct the operation.  For the successful implementation of a multivariable controller, minute to minute updates of the physical product properties are required, so that the multivariable controller can correct and control the product properties as soon as the process conditions change.  It is possible to estimate the physical property based on secondary process measurements; so-called inferred physical property estimators or “inferentials”.

When optimising a process, to maximize production of the most valuable products, the operator must have clear indication of the product qualities, either via on-stream analyzers or inferential models.  Without such knowledge the operator would be forced into a conservative operation where the most valuable streams are not fully maximized.

Continuous analyser measurement is the optimal solution, but online analysers are expensive to install and maintain.  The product properties and qualities can be estimated with inferentials, biased irregularly with lab samples.  The inferentials are only an estimation of the product quality, based on available measurements on the unit, like temperatures, flows and pressures, and can never have the accuracy of analysers.  It is therefore a compromise between capital cost (for installing and maintaining analysers) and yield benefits (as the unit has to be operated more conservatively with less accurate quality measurement).

GCC and GDS Rigorous Model Based Inferentials

AMT, together with Petrocontrol, offer rigorous model based inferentials, based on process models developed for multi-draw fractionators (GCC) and simple distillation columns (GDS).  The rigorous models give improved property predictions and generally lead to better acceptance of the complete multivariable control package – operators see the combined effect of inferentials and multivariable control.

The GCC Rigorous model estimates the TBP curve from unit conditions and then estimates the following product qualities:

  • ASTMX%
  • Flash points
  • Freeze and cloud points

The models include the mass and energy balances on the column and have a high degree of complexity and are therefore more time-consuming and expensive to develop.

The GDS calculation uses a semi rigorous model of the tower sections and then estimates the top and bottom product purities as a function of :

  • Temperatures, especially tray temperatures
  • Pressure
  • Liquid / Vapor flows
  • Tray efficiency

Empirical Inferentials with LAB feedback

AMT also offers inferentials, based on empirical correlations to LAB results using all available data such as vapor/liquid flows, pressures and temperatures in the columns.

To achieve best results for the empirical inferential, the data and LAB samples should ideally be taken under highly controlled conditions during the step testing period or a separate inferentials testing period:

  • With sufficient independent variation or movement on the different inputs to the inferentials to prevent cross-correlation
  • Ensure accurate recording of sample times
  • Prevent any “feedback” leading to causality issues

The long term sample frequency with empirical LAB based inferentials also needs to remain higher (at least 1 lab sample per day) than with rigorous model based inferentials (typically 1 lab sample per week).