Jan-2014

Optimising distillation column product quality

Process simulation and random sampling were used to optimise product targets for a propylene splitter unit

JOSE BIRD and DARRYL SEILLIER
Valero Energy Corporation

Article Summary

In an attempt to ensure that product specifications are always met in distillation unit operations, refinery operations will often run a process unit at higher reboiler duties than required, resulting in excessive energy use and sub- optimal product yields. Process units are run in this less than efficient fashion to provide enough of a cushion for process variability resulting from changes in process conditions, such as changing feed compositions, feed rate and feed temperature. Therefore, accounting for process variability can be crucial when optimising distillation unit operations. This article shows how to determine the optimum operating targets when product specification constraints need to be met in the presence of process variability.

We present a methodology that uses a process simulator along with Monte Carlo simulation to account for changing process conditions. By using a process simulator, the performance of the unit can be evaluated outside the current operating range. To illustrate the application of the methodology used, the optimisation of the operations of a propylene-propane splitter distillation unit was chosen.1 Figure 1 provides a schematic of a typical propylene-propane splitter distillation unit. A mixture of approximately 75% propylene and 25% propane enters the unit. Reboiler steam provides the energy required to separate the feed into a side draw propylene product and a bottoms propane product. Chemical grade propylene specifications were assumed to be a minimum 93% propylene purity and a maximum 15 ppm methyl acetylene (MA) concentration. A concern that arises when operating the propylene splitter close to the propylene purity specification is the corresponding increase in the MA concentration of the propylene product, which might result in off-spec product. If this happens, the product is downgraded to refinery grade propylene. Accurately predicting the performance of the distillation unit at lower propylene purities needs to consider changes in feed composition, feed temperature and feed flow rate as operations is facing this challenge. The use of a process simulator with cases generated with Monte Carlo random sampling allowed for proper modelling of the distillation process mechanisms in the presence of process variability. 

The following steps summarise the methodology used in this study:
• Process simulation model calibrated to current unit operations
• Probability distributions generated for input process variables
• Monte Carlo simulation used to generate process simulation cases
• Process simulations conducted for all cases generated.

Steps 2-4 are repeated for the different propylene purity scenarios considered. The process simulation results are then used to generate joint probability density functions of propylene purity percentage and MA concentration. The resulting joint probability distributions are used to assess the ability of the process to meet product specifications at different operating targets and to determine the associated optimum unit operating targets.

A detailed description of the analysis methodology is provided below, followed by results of analysis and conclusions.

Analysis methodology
To examine the effect of process variability on the ability of the process to meet product specifications, Monte Carlo simulation was used to generate a set of process simulation cases that captured existing process variability. The following model inputs were considered to be primary sources of process variability: feed temperature, feed rate, feed propylene composition, feed MA concentration and propylene product percent purity. These inputs were assumed to follow independent normal distributions. Historical data with the APC system turned on were used to build the probability distributions of these inputs.

Figure 2 shows histograms and the calculated normal density functions for each of the model inputs. The mean and standard deviation for each input random variable were calculated and used to model the corresponding distributions. The propylene purity value generated for each of the cases was then used to estimate the propane concentration in the propylene product. The propane in the propylene product was set as one of the process specifications in the simulation. The concentration of the remaining trace components was assumed to be a constant when computing the propane concentration in the propylene product.

Figure 3 highlights the analysis methodology. First, 200 independent random samples for each of the model inputs is generated for the propylene purity target scenario considered. The propylene purity target was assumed to be equal to the mean of the propylene purity distribution. To generate the different propylene purity target scenarios, the mean of the propylene purity was shifted while the standard deviation was kept the same. Two hundred cases for each propylene purity target scenario were generated, with each case representing a different combination of the model inputs based on the 200 independent random samples generated. The SAS Analytics procedure Proc SimNormal2 was used to generate the random samples. The SAS Analytics scalable software environment was selected for this project due to its extensive statistical and charting capabilities.

The 200 cases were then evaluated using a process simulator. The Petro-SIM 4.1 process simulation software was used, and the model inputs entered using an Excel spreadsheet interface within Petro-SIM software. The results from the simulations were then used to estimate the incremental profit for each propylene purity scenario against the base case. The calculated incremental profit figures were then used to generate a profit response surface.3,4,5 The profit response surface was generated using two SAS Analytics procedures: Proc G3Grid and Proc G3D.2 The simulation results were also used to generate contour maps representing the joint probability distribution of propylene purity and MA concentration in the propylene product. The contour maps of the joint probability distributions were generated using the SAS Analytics procedure Proc KDE2, which uses kernel density estimation.

The process simulator was configured to use three process specifications: condenser temperature, propane weight percent in the propylene product, and propylene weight percent in the bottoms propane product. The propane weight percent in the propylene product was calculated from the difference after subtracting the propylene purity weight percent and the concentration of trace components from 100%. The weight percent propylene in the propane product was varied from 1-4% and was assumed to be a constant for each propylene purity scenario. To reduce the number of cases required to estimate the incremental profit response surface, the experimental central composite design configuration shown in Figure 4 was used.3,4 This configuration allows for consideration of second order terms in the construction of the profit response surface. The different scenarios were defined by the combination of average propylene purity and the propylene concentration in the propane.

Base model description
To build the propylene-propane splitter model, data historian information including flow rates, compositions, pressures and temperatures were gathered for feed and product streams. Distillation column data including number of trays as well as pressures and temperatures for condenser and reboiler were obtained. Component material balances using feed and product flow rates and compositions were performed to test the validity of the data. Propylene and propane product compositions and flow rates were used to reconstitute the feed stream.