PETR 153

While multivariate regression methods (PCR, PLS) have been successfully applied for analysis of petroleum feeds and products, these techniques are poorly suited to the prediction of crude assay data. The hundreds of models that would be needed to predict the assay information would be difficult to maintain, and predictions would still need to be interpolated to match distillation cut endpoints. Many of the properties that would need to be predicted are nonlinear functions of composition. Most significantly, it is necessary to predict properties of distillation cuts from analyses of the whole crude.

Crude Assay databases include software for calculating properties of crude blends. Virtual Assay exploits this known technology by analyzing unknown crudes as blends of assayed reference crudes. FT-IR spectra of an unknown crude are corrected for extraneous signals and scaling variation, concatenated with inspection data, and then analyzed as blends of reference crudes using an iterative non-negative least squares algorithm. Properties for the Virtual Blend are calculated using the crude assay blending algorithms. Virtual Assay predictions compare favorably with the precision of the wet assay measurements. The technology is sufficiently robust for use in refinery laboratories and on-line analysis.