Terminology for Assay Simulator System

Assay Prediction - Process of predicting assay cut properties using the property prediction functions. It may include the correction that is computed when using a Base Assay.

Assay Simulator - The name given to the Excel application that uses the CCPP models to predict new assay results. It converts WC properties to a C6+ basis, allows selection of the Base assay, tunes the model to match the Base and performs the prediction of the new assay. It also allows graphical representation of the properties, accuracy information and tuning factors.

Assay Updating - Process of predicting assay cut properties using the prediction functions and a Base Assay (to reduce model offset errors).

Base Assay - An assay that generally represents the crude in question, but which may not be entirely up to date or complete.

CCPP Models - The name given to the Excel Functions and other programs that make up the Crude Cut Property Prediction models. The system includes the property models as well as routines to perform raw property predictions, error checking, tuning and predictions using tuning factors.

Family of Crudes - A group of crudes that act in a predictable manner. Usually they will come from the same region of the world, but not all crudes from a particular location will necessarily be of the same family. Family "membership" is determined by how accurately the cut properties of a given crude are predicted when the crude is included with other family members. A family of all the worlds crudes is delivered with the CCPP and Assay Simulator applications

World Model - This is a group of property models based on a large group of assays from around the world. It includes about 7000 fractionated cuts from assays. These models use separate K factors for distillate and resid cuts.

Prediction Functions - Excel Functions, implemented in VBA, that are produced using the Neural Net software, and which predict assay cut properties from whole crude properties. Sometime two functions are created for a given property; one for the distillate cuts and one for the resids. These are integrated so that a single query can be made using initial and final cutpoints. A set of models is prepared for each Family of crudes. An end point greater than 1250F usually branches to the Resid function. These functions can also return prediction statistics and perform "Tuning" functions.

Whole Crude (WC) Properties - Properties of the sampled raw, uncut crude, such as specific gravity, sulfur, light ends, metals and Conradson carbon residue. These properties can be determined without the expensive and time-consuming fractionation process used in a conventional assay process. They can be correlated to the actual cut properties, but sometimes require transformations to make them work well.

Transformations - Changes to variables to make them correlate better. For instance, whole crude specific gravity and cut yields must be converted to a C6+ basis by mathematically removing the light ends. Viscosities are routinely converted to a Log(Log(Vis+1.5)), and cold properties may be converted to an absolute temperature basis. Metals, ConCarbon and some other properties are correlated using the Log(Value). All of these transformations are take place behind the scenes so the user does not see them.

Property Function Program - This is the software developed and used by HPI to generate the Excel property functions. The program collects the data with the desired properties, eliminates undesirable data, makes transformations of variables, activates and controls the Neural Net software and generates the Excel VBA functions that are used by the Assay Prediction Program.

Accuracy Predictions - Most predicted properties have accuracy predictions calculated after the Neural Net software has been run. The prediction software generates up to 5 sets of accuracy predictions for different parts of the TBP temperature range. Predictions available to the user are generated by interpolating along these accuracy values. Many property accuracy values are generated on a relative basis (Error/Actual), but the end user programs convert these to absolute accuracy predictions.

Model Tuning - This is the process of determining the small corrections needed to make the model exactly match the Base assay data. In most cases it is adjustments to the whole crude K value, but can also be offset factors. For models using the "Absolute" method of error prediction the adjustment factor is simply actual-predicted value, whereas for properties with "Relative" errors (like viscosity) it is computed as: (Actual-Predicted)/Actual. The tuning can be controlled by the user to constrain the amount of K correction, but the default process controls the K to practical values. Whatever remaining error is left after the K correction is taken as an offset error.

Model Offset Error - This is the difference between the average predicted value and the average actual value. It represents an imperfection in modeling. This error can be eliminated by using the Base assay to determine this offset, and then applying it to the new assay. The offset is determined by predicting the Base crude cut properties using just the whole crude properties. The difference between the predicted and actual Base cut properties is assumed to be an offset that is constant for that crude, and can then be applied to the new prediction.

Error Ratio - The ratio of: Absolute Error / Average Value. This is often a more useful measure of the error than the absolute value, where the values range from very small to quite large (0.001 to 5%) as with sulfur.

R^2 - This is the most common measure of how good well the correlation works. It is effectively the fraction of the error in the data that is predicted by the model. Thus a value of 1.0 is a perfect model. Mathematically it the ratio of:

(1-Standard Error of the Estimate / Standard Deviation of the target property),

assuming any offsets are removed by using a Base assay.

R^2 values are determined for up to six boiling point segments for each CCPP Model. This effectively eliminates the boiling point from the prediction errors, so it represents the effect of the Whole Crude properties. Tables and plots of the R^2 is available for all the errors.