Fouling monitoring and prediction of heat exchangers via simulations powered by AI-driven models
Hydrocracker units are the most valuable conversion facilities in refineries. The outcome is the conversion of a variety of feedstocks to a range of products—the units that produce those products can be found at various points throughout a refinery.
Hydrocracker units are the most valuable conversion facilities in refineries. The outcome is the conversion of a variety of feedstocks to a range of products—the units that produce those products can be found at various points throughout a refinery. The main objective of these plants is to transform heavy products into valuable lighter products (including saturated hydrocarbons), depending on the reaction conditions. Major products from hydrocracking include jet fuel, diesel, gasoline fractions and liquefied petroleum gas (LPG).
The hydrocracking process is a catalytic cracking process assisted by the presence of an elevated partial pressure of hydrogen, a high-temperature environment and special selective catalysts. The products produced under high pressure and temperature are separated in the fractionator section—the quantity of removed unconverted oil from the bottom of the fractionator also indicates the unit’s conversion performance. Since these production and separation processes are energy-intensive in terms of fuel gas, steam, electricity and cooling water, any optimization that can be applied is beneficial in terms of reducing operating costs.
In addition to product profitability, refineries’ energy consumption is becoming another key point for future productivity. New investments concentrate on improving heat transfer resources with additional heat exchangers that utilize other fuel/steam sources. In this respect, hydrocracker units are key conversion units with high profit margins. Recently, a heat integration project was implemented in the one-stage, once-through hydrocracker unit. Here, the preflash system is integrated to reduce the duty of the fractionator furnace. Then, the stripper bottom stream is taken to the preflash drum where it is separated into lighter fractions that enter the column to be sent onto the kerosene draw tray of the fractionator column.
After this heat integration project, it is predicted that the load of the fractionator furnace will be reduced. Two centrifugal pumps were added to pressurize the preflash bottom stream. Subsequently, to increase the inlet temperature of the fractionator furnace, the preflash bottom stream was heated with the diesel pump around the stream and a newly allocated plate-fin type heat exchanger was implemented. After the heat exchanger, this stream was separated to the furnace branch valves and sent to the fractionator furnace. In FIG. 1, the equipment and plate-fin heat exchanger added with heat integration are shown.
FIG. 1. The equipment and plate-fin heat exchanger added with heat integration.
As shown in FIG. 1, a plate-fin heat exchanger was added to heat the preflash bottom stream with the diesel pump around the flow to increase the inlet temperature of the fractionator furnace. With this configuration, the furnace inlet was preheated and the fuel consumption of the furnace was reduced. Monitoring this plate-fin heat exchanger is critical as it is the first plate-fin heat exchanger used in hydrocarbon service for the unit and its fouling directly causes a load and fuel increase in the fractionator furnace.
AI-driven modeling
Non-TEMA (non-Tubular Exchangers Manufacturers Association) type heat exchangers—such as plate-fin, welded plate and frame exchangers—are commonly used in refining and petrochemical process units and power generation plants because they offer numerous superiority properties (very low fouling tendency, lower plot area, higher heat transfer area, etc.). However, process engineers face the challenge of simulating these type of heat exchangers rigorously and robustly via traditional simulation programs. Using artificial intelligence (AI)-driven machine-learning can help overcome these challenges.
By modeling the fouling in the plate-fin heat exchanger with machine-learning algorithms, any fuel consumption increase and inefficiencies in the furnaces will be determined in advance through a decision-making mechanism, and unit turnaround and cleaning periods can be safely organized in advance. Since hydrocracker units are energy-dense processes and affect the Energy Intensity Index (EII)—a petroleum refinery energy efficiency metric that compares actual energy consumption for a refinery with the “standard” energy consumption for a refinery of similar size and configuration—throughout the refinery, it is very important to simulate and predict fouling rates of the plate-fin heat exchanger. However, this type of heat exchanger usually cannot be modeled via classical simulation tools (at least not easily). In this case, first-principle modeling tools and AI-driven algorithms were combined to model the plate-fin type heat exchanger (non-TEMA). The integration of AI-driven algorithms into the simulation models (known as hybrid models) helps not only simulation but also the prediction of fouling rates in future operating conditions.
For the heat exchanger simulation, three different stages were completed during the modeling steps:
- Definition and collection of data
- Creation of AI-driven model
- Model deployment.
During the definition and collection of data, 25,000 hourly point datasets were used for eight different dependent and independent variables. Noises and bad data were cleaned, and outlier detection was applied on the rest of the datasets (~13% of noises were cleaned). The results were achieved with very high accuracy (R2 > 0.95) and predictability (Q2 > 0.95), as shown in FIG. 2. Lasso CV was used to create the machine-learning models.
FIG. 2. Predicted data vs. plant data for both cold and hot side outlet temperatures.
When the model results were compared with plant data, the hot side and cold side outlet temperature values were estimated with a ±3.3°C deviation (< 1.3%) and with a ±2.8°C deviation (< 1.1%), respectively. Model outputs are shown in FIGS. 3 and 4 for cold side outlet temperature and hot side outlet temperature, respectively.
FIG. 3. Model outputs vs. plant data for cold side outlet temperature.
FIG. 4. Model outputs vs. plant data for hot side outlet temperature.
Fouling of this heat exchanger increases the fuel consumption rate of the fired heater and negatively affects the energy efficiency of the hydrocracker unit. Therefore, information about the fouling rate of the exchanger must be gathered. For this purpose, any potential savings were analyzed by conducting fouling prediction using the hybrid model to determine whether or not to clean the heat exchanger.
The temperature loss caused from the fouling rate was calculated at 1°C–3°C, which equates to 100 kW–200 kW duty loss due to fouling, during 3 yr of operation (FIG. 5). The next 6 mos were also predicted for the cold side outlet temperature—the model showed that the outlet temperature would decrease ~2°C if the same operating conditions continued.
FIG. 5. Temperature loss on the cold side of the heat exchanger caused by fouling.
When evaluated with current refinery product margins, the daily profit of a one-stage, once-through hydrocracker unit is compared against the return of heat exchanger cleaning, the cost of unit downtime due to heat exchanger cleaning is 27 times higher.
Takeaways
AI-driven models provide not only rigorous modeling of equipment but also predictions that can be used for maintaining optimum and efficient refinery operations. The accuracy of the digital twins used was improved by leveraging a wealth of operating data and the power of machine-learning. This proves to be an easier and faster way to maintain updated conditions of simulation models.
This study has shown that process engineers can create value from plant data when AI-driven models are used. With the help of such models, it is possible to predict when it is necessary to stop the unit and clean the heat exchanger. HP
The Authors
Özsağiroğlu, E. - Tüpraş, Şişli-İstanbul, Turkey
Erhan Özsağiroğlu is a Simulation Superintendent at Tüpraş headquarters within the Process Optimization and Digitalization department. He previously worked as a Senior Process Engineer at an engineering company for 5 yr and holds a PhD in chemical engineering from Istanbul Technical University. The author can be reached at Erhan.Ozsagiroglu@tupras.com.tr.
Mert, D. K. - Tüpraş, Körfez-Kocaeli, Turkey
Diyar Kiliç Mert is a Process Superintendent for the CCR/ISO/Benfree units at the Tüpraş Izmit Refinery within the Process department. She previously worked as a Process Engineer in hydrocracker and hydrogen units for 6 yr and holds a BSc degree in chemical engineering from Istanbul Technical University. The author can be reached at diyar.kilic@tupras.com.tr.
Related Articles
From the Archive