Process model predictions

When using Base + Delta models, the models typically use one or more driver properties to calculate the properties of the unit outputs. The properties calculated by the model are predictions.

Example: A hydrotreater model is driven by the sulfur content of the input feed. The model predicts the sulfur content of the output streams of the unit, along with yields, density and other properties of the hydrotreated output, and properties of the hydrogen sulfide stream of the reactor.

The Predictions page in the Model Structure tab shows the predictions configured in the Base + Delta model associated with the selected process unit.

Note: To modify your predictions, click on the Process Model tab and edit the associated model.
Predictions can only be edited if the process model is editable. For Base + Delta models, creating a new prediction means entering a new base for the property prediction, and a new delta for every driver property used by the model.

Note: If a process model has been imported from a password-protected file, model details are unavailable.

The Solution column is the outgoing value for the prediction, calculated as a result of running the Base + Delta values. Where the process unit has multiple operational modes, the value is the predicted value for the currently selected mode. Where the Base + Delta model has multiple bases, the solution is the interpolated value between the active bases using the mode solution value for the Base + Delta interpolation variable.

You can disable a prediction by clearing the check box in the Active column for the corresponding row. When you disable a prediction in the Predictions page, the corresponding row in the tables under Base Delta Structure is disabled as well.

Prediction types

Several different types of prediction can be added to process models:

• Yield Predictions: the yield of a particular output stream.

• Property Predictions: a property of one of the unit's output streams.

• Utility per Feed Rate Predictions: a utility produced or consumed by the unit which depends on the rate of feed to the unit.

• Utility Base Load Predictions: a base load of utility produced or consumed by the unit per unit time.

• Operating Parameter Predictions: an operating parameter of the model.

For advice on creating the structure of predictions, see Prediction Structure.

Predictions using blend rules

You can associate a blend rule to predictions that involve properties:

1. Make sure the blend rule is included in the supply chain model (see: Blending).

2. Filter the property list and select the property that should be predicted.

3. Click on the Blend Rule option button.

4. Select the blend rule to associate with the property from the list of available blend rules.

   

When a blend rule is associated, the result of the prediction is expected to be in the units of measure of the blend rule. It will be transformed back to the real property value using the inverse of the blend rule before being displayed in the tool or used for product blending. See: Blend Indexed Property Predictions.

Predictions in blend index space are often associated with driver predictions also in blend index space.

Example: If the flash point Index blend rule is associated with the flash point property, it is expected that the Base + Delta model predicts flash point values in the index value (not as a temperature). This indexed value is then transformed back to the real flash point value (in a temperature unit) using the inverse of the blend rule to be used for product blending.

Warning: Blend rules and properties
If a blend rule is associated with the supply chain model, and this property is used in its normal unit of measure, then whenever the property is referenced within Base + Delta models, the original property value is replaced by its blend rule equivalent.
For example, a pour point index formula might be associated with the global supply chain model. If a process model predicted pour point (in temperature), then during optimization the pour point Base and Delta values would be replaced by terms derived from these using the blend rule. Depending on the blend rule used, this may introduce significant non-linearity into the optimization problem, resulting in slower behavior.
To avoid this problem, you can predict the property value in linear space (that is, in the units of measure of the blend rule) rather than in its real unit of measure. This allows the linearized index value to be used in the optimization formulation, resulting in better performance of the optimizer.
For example, rather than predicting pour point, the Base + Delta model could predict pour point index. As a result, all the predicted values would already be linear, and not require a special formulation within the optimization problem. At some point during the optimization, it may be necessary to convert from index space to the real unit of measure, but this can be done at the point it is needed, rather than at many places within the optimization. As the conversion from index to real space has to happen fewer times, there is less non-linearity within the optimization model and the problem responds faster.
This configuration is especially useful for properties which are 'passed through' the model, that is, properties which are the same in the feed as the product. As these do not change, they can be easily and understandably represented as indexed properties.

Yield predictions

The Yield Prediction Editor is used to edit the yield output properties associated with process model output streams, including swing cuts. Yields are always relative to the primary feed on the process model. This is the lower left-hand input to the unit.

Property predictions

The Property Prediction Editor is used to edit the properties associated with process model output streams, including swing cuts.

For each set of predictions you add for output streams, you can add corresponding drivers for the input streams. See the table below for details.

You can add predictions for multiple properties without closing the Property Editor. Click Add to add a set of predictions (and optionally drivers), then change the property to add another set. When you are done, click Close to leave the Property Editor.

To add a property prediction to a process model stream:

5. Right-click the predictions list and select Add Prediction > Property from the context menu.

6. Select the required output streams to associate the property with from the Outputs table.

7. Optionally, select the input streams for which to create a matching driver from the Inputs table.

8. Select the required Property.

9. Select the required unit of measure from the UoM menu, or alternatively select the Blend Rule radio button and choose a blend rule.

10. If the property is adjusted relative to the stream yield, select the Property × Yield check box and choose the yield basis and unit of measure. Select the Auto Calculation check box if you want to enter the property value in the Base Delta Structure table, instead of the value multiplied by the yield. AVEVA Unified Supply Chain will perform this calculation automatically when simulating or optimizing the case.

To edit a property prediction:

11. Right-click the required property for the required stream and select Edit Prediction from the context menu.

12. Make the required changes in the Property Prediction Editor and click OK.

Note: For a quick way to enable the Auto Calculation option for multiple Property × Yield predictions, enable the Property × Yield Conversion optional feature.
When this feature is enabled, the Convert Prop × Yld button is added to the Home ribbon tab. Click this button to enable the Auto Calculation option for all the Property × Yield predictions in the selected process model, and to rescale all the delta values for the relevant predictions. The process model must be editable for the button to be enabled.

To remove a property prediction, right-click the required property for the required stream and select Delete Prediction from the context menu.

Tip: You should include density @ 15°C on all output streams. This property is used for weight/volume conversions and so allows flow to be reviewed in both weight and volume terms.

Utility per feed rate predictions

The Utility per Feed Rate Prediction Editor is used to edit the utilities that are produced or consumed by the process unit. Utilities are produced or consumed relative to the amount of material being processed by the unit, that is, the more material the unit handles, the greater the amount of utility produced or consumed. When entering the utility unit of measure the unit of measure is defined as utility unit of measure per amount of process unit feed. For example BTU/100t means BTU units per 100 tonnes of feed.

Utility base load predictions

The Utility Base Load Prediction Editor is used to edit the utilities that are produced or consumed by the process unit per unit time, and do not depend on the rate of feed to the unit. That is, the base load does not change whether the unit is processing 1 tonne or 100 tonnes of feed. Most process units have a base load utility consumption, along with a per feed utility consumption.

Note: For a prediction of type Utility Base Load, in the Base Delta Structure table you should enter nonzero coefficients only in the Base and Offset columns. Nonzero coefficients in driver columns do not make sense in this case and would lead to unexpected results. A utility base load, by definition, is a constant value that does not depend on the rate of feed or other properties.
Add a prediction of type Utility per Feed Rate to model utility consumption as a function of unit feed.

Operating parameter predictions

The Operating Parameter Prediction Editor is used to predict values for operating parameters. These parameters must already be present for the model before they can be added as model predictions (see: Configure the Operating Parameters for your Process Units).

Prediction structure

Sign convention

By convention, increases in a yield prediction have a positive sign in Plan. This applies to both the base yield and also delta values (contrast this with some other systems where a negative yield indicates an increase in production).

Negative yields

Ensure that yields for base delta models can never have negative yield predictions.Plan enforces yields from process models to be positive. Where a Base + Delta model can predict a negative yield, this automatically implies a constraint on the behavior of the Base + Delta model, which can have unintended consequences during optimization. Where a process model has the ability to adjust its operating parameters and/or input drivers to prevent the negative yield then the optimizer will do so, but this will cause an implicit constraint on the model behavior. Where there is little freedom in the process model, it may not be possible to have a non-negative yield prediction, and the consequence may be the unit will shut down and refuse to operate.

To spot behavior associated with negative yield predictions look for:

• a process unit that refuses to run, even with a minimum flow and the associated infeasibility breaker removed

• a base in a multi-base model that refuses to operate. You may see errors regarding non-contiguous bases being active

• an active process unit that is normally in mass balance suddenly going out of mass balance.

To prevent negative yield predictions:

• ensure the yield predictions are always positive over all typical operating parameter and driver input ranges. Use the Process Model Simulator and check the model performance using typical input driver values

• ensure for weight based process models that output yields sum to 1, and that offsets and deltas sum to 0.

Blend indexed property predictions

After optimization, you may see the warning "Process unit <unit> Process model <model>, <property> requires non-linear conversion to Blend Rule".

For example, "Process Unit Visbreaker Process model Visbreaker, Viscosity @ 40°C requires non-linear conversion to Blend Rule".

This warning is generated when Plan has had to convert an entire Base + Delta model row into its blend index equivalent using the blend index formula. By default, where a property has an associated blend index, Plan will always use the indexed property value during optimization. If the property is entered in real engineering space, then the resultant calculations will always occur in linear blend index space with the conversion between engineering space and blend index space happening automatically and transparently for users (See: Blend Rules, Blend Indices in Process Models).

In some situations where a blend indexed property prediction is predicted in engineering space, the resultant converted equation may perform poorly. In this instance, converting the structure to linear blend index space may result in better behavior of the process model. To convert a prediction to blend index space, associate the blend rule directly with the prediction and driver. (See: Property Driver Editor, Predictions using Blend Rules).

Pass-through properties

Where properties 'pass-through' a process model (and do not change), you can leave these values in their original engineering unit of measure. For these properties, Plan will automatically convert this structure to its indexed property form.

Example: The freeze point property may pass-through a kerosene hydrotreater (so that the freeze point of the product is equal to the freeze point of the feed). Within Plan, you would create a driver and a prediction for freeze point that were linearly related (so the Base was 0 in both cases, and the Delta was 1).

Input yield predictions

Some process units have multiple feeds where the consumption of secondary feeds is relative to the flow and quality of the main feed.

Example: A hydrotreater consumes hydrogen proportional to the flow rate and sulfur content of the feed, while an alkylation unit consumes isobutane proportional to the flow and olefin content of the feed.

These secondary feeds (hydrogen and isobutane in the example above) are predicted using Input Yields. Input Yields are not added automatically (even when a process model is created in a unit with two or more inputs) and must be added manually. Setting an input pipe flow relative to another input pipe automatically turns the prediction into an input yield. Input yields can only be relative to the primary feed to the unit, that is, the feed at the bottom left of the process unit with an index of 0.

Losses

Many process units will lose some of their feed during processing. Material may be consumed in the reaction, breakdown, or deposit during the process. Where material is lost during a process, you can model it as an explicit loss stream within the unit. To do this, add another product named loss and attribute the loss of material to this product. The total sum of the product outputs should still be 100% (by weight).

However, directly measuring and attributing loss to individual process units may be difficult. Accurately measuring input and output flow of all unit products may be difficult, and day-to-day changes in unit operations (which are difficult to capture) may affect the actual loss. Instead, it may be simpler to treat loss across the whole plant as a single loss on a particular unit. That is, rather than modeling loss on individual process units, have a single material loss and allocate this to a single unit. See: Utilities and Loss for more information.