Development of New Guidelines for Manufacturing
- Conference: AIChE Spring Meeting and Global Congress on Process Safety
- Year: 2018
- Proceeding: 2018 Spring Meeting and 14th Global Congress on Process Safety
- Group: Global Congress on Process Safety
The solutions have significantly reduced the fluctuation of the cracking severity and increased ease of operation. By reducing the variance of cracking severity by as much as 60%, the project improved the consistency of the furnace's yield slate, which in turn contributed to the stability of the downstream recovery section and potentially reduced operating costs.
The project used a neural network-based inferential model to predict process properties and a rule-based expert system to ensure the reliability of analyzer results. A comprehensive logic combined the analyzer feedback and predictive properties to determine the process value for severity controller (Fig. 1).
The solution was provided by the East China University of Science and Technology, which has successfully completed regular control for cracking furnaces and key columns in five ethylene plants owned by Sinopec.1 It covered all 11 liquid furnaces dealing with four kinds of feedstocks.
There are 14 cracking furnaces. Eleven of them deal with liquid feedstock, including light naphtha and heavy naphtha, hydrotreated vacuum gas oil, and atmospheric gas oil. Three of them handle gas feedstock, such as ethane and LPG.
Real time feed compositional data are unavailable except for the density.
The Honeywell TPS DCS system provides all the typical operating data of the furnace and transfer line exchanger, including feed rate, steam flow rate, coil inlet/outlet temperature (CIT/COT), and coil outlet pressure in the furnace, and the temperature and pressure at the TLE outlet.
The cracked-gas components including H2, CH4, C2H4, C2H6, C3H6, and C3H8 are measured by six on line analyzers, each shared by two liquid furnaces. For one furnace, the time delay of the cracked-gas component value is around 15 min.
Regular controls like the COT control, COT balancing control, and throughput control are in good service.
Cracking-severity control has been installed in one furnace but deteriorated into disuse partially due to the high maintenance cost of analyzer and frequent feed composition changes. The project failure in one furnace made the managers reluctant to install the severity-control project for the second time. Then all the effluent analyzers became less important and received little attention for a long time.
After discussions and evaluations of potential benefits and challenges, plant management decided to install severity control.Although COT has been used to express severity for decades since the plant was installed, it cannot completely replace cracking severity. As the feed composition changes or gradual coking proceeds, cracking-severity changes even if the COT is a controlled invariant. On the other hand, unlike the effluent analyzer, it is difficult to detect coil outlet thermometer problems because of the drift or mounting deficiency.
The primary objective was to stabilize the effluent composition from a given feedstock without adversely affecting the run length. The effluent composition is determined exclusively by severity and selectivity. Since the steam-to-hydrocarbon ratio is normally fixed in plant operation, which means the selectivity is almost fixed, the effluent composition can only be affected by the severity. By reducing the standard deviation of the cracking severity within a defined limit, management expected changes to the downstream recovery section to be minimized.
Another objective was to make full use of the on line analyzer by integrating it into severity control. The improvement in analyzer performance, in turn, would benefit not only this quality control project, but also the severity optimization project in the future.
The inferential models were designed to provide real-time and steady prediction for the severity. Since the on line gas chromatographs are normally shared by two furnaces and the analytical system is subjected to frequent faults, the inferential model would enable the severity control to hold the severity during analyzer sample intervals and fault recovery periods. Besides, the steady estimation by the inferential model can cross check the analyzer results.
A total of 23 inferential models were developed for all 11 liquid furnaces, each of which can deal with one to three kinds of feedstock. One model deals with one feedstock of one furnace.
The inferential models were developed based on the well established neural network tool. The performance of the NN model depends on the process data collection and processing, in-depth process knowledge, and skilled manipulation of NN.
Data collection for each furnace began from the project's preliminary design phase. Normally, each model requires more than 4 months of historical data at sampling frequency of 1 min. The collected data were checked by both project engineers and plant engineers, and those data points during nonrepresentative operation or instrument failure were marked.
In the data cleaning stage, the process data and analyzer samples at different times were synchronized first. Since the dead time was not always the same for every analyzer, the synchronization was done manually. Then the marked data and other bad data can be removed instead of before synchronization, so as not to deteriorate the time sequence. The gross error and random error were detected and removed by statistical and filtering methods.
The inputs for the inferential model include the COT, feed flow rate, steam-hydrocarbon ratio, and feed density, etc. They were chosen based on variable availability and reliability. Although the COP has great impact on the effluent composition and severity, it was not chosen as the input because the measurement are unreliable, possibly resulting from low instrument accuracy and the high temperature environment at the measuring point.
The model output can be propylene-ethylene ratio (PER) and methane-propylene ratio (MPR). Plant personnel always like the PER because it contains the ethylene component that is their major concern. Plant data trending, however, showed that PER was insensitive to the input variables for heavy feedstock, while MPR showed reasonable sensitivity. Therefore, PER was chosen as the severity for LNAP and HNAP and MPR for HVGO and AGO.
The weights and hidden nodes number were determined by in-house software. Evaluating the model performance considered both the maximum error and mean square error. Testing the model considered tracking directionality more important than accuracy.
Once a model with satisfactory performance was developed, the model information was added into the DCS program with control language for on line tracking analysis.
If the on line prediction performance was acceptable, then the model can be configured to the model bias module. The model was biased by analyzer results, the reliability of which depends on the analytical system maintenance and fault diagnosis system.
At the earlier stage of this project, we set up a special maintenance group to smooth away key problems and optimize the process and parameters in the analytical system, especially for the samplers and pretreatment. Following are the key improvements:
â¢ Rectification of faulty equipment and improper installation.
â¢ Optimization of the sample flow and temperature controller's set point in the sampler.
â¢ Preliminary pretreatment added to remove the oil and water carried with samples.
â¢ Rules made for preventive maintenance of the analytical system.
â¢ DCS data points established for corresponding compositions.
This work laid a solid foundation for the severity-control project. Occasional analysis errors may still occur, however, and because they are part of the closed control loop, it is vital to ensure that only correct analysis results be used for model correction.
A rule-based analytical diagnosis expert system for each furnace effluent analysis was developed. The rules include operation and maintenance experience, characteristics of the analytical instrument, and pyrolysis mechanism.
Conducting spatial and temporal correlation analysis on the operating conditions and the compositional results collected identified the analytical fault and its possible reason. The basic principle was to ensure analysis reliability, even at the cost of ignoring a good analysis.
Once a fault is identified, a fault flag and warning will be generated, which disappear only when the fault has been repaired or operators have confirmed. The warnings and corresponding indicators displayed on the DCS screen will shorten the restoration time.
The fault flag will also change the model's correction logic. Normally the severity value is a combination of the inferred value and a moving-window error between inference and analysis. When the failure flag is triggered, the moving-window error will be replaced by a cumulated average error before the analytical fault. Therefore, the severity controller can still work properly during the analytical fault recovery period, earning more on line time.
The estimated quantifiable benefit is $2.6 million/year and the payback period is less than 6 weeks. Weekly payback for the severity control was calculated as $ returned/week/$ invested.
Benefits are achieved by reduction in severity variance, increase in desirable products, and longevity of controllers.
We could not readily quantify a few additional benefits such as high-quality inferential control variables for severity (Fig. 2).
Another intangible benefit is that the consistent cracked product would stabilize the downstream separation section, allowing it to run closer to constraints. Moreover, it provides a foundation for the severity optimization project in future.