During the operation of MOL’s Danube Refinery, nearly one hundred thousand data points are recorded every minute. The algorithms running in the Microsoft Azure Machine Learning cloud service help in gaining relevant business insights from these.
The Danube Refinery, located in Százhalombatta, began operations in 1965. With a processing capacity of 8.1 metric tons of crude oil per year, it is one of the largest refineries in the Central and Eastern European region. Microsoft’s advanced analytics solutions were used in the plant to examine how production processes could be optimized, the efficiency of the plant increased, and the lifetime of the processing equipment extended.
The Business Problem
In an oil refinery, end products—fuels (diesel, gasoline, kerosene, LPG), aromatics, base oils and paraffines—are produced as the result of very complex processes. The aim is the most efficient use of raw materials, while decreasing the quantity of the less valuable residual petroleum coke in the course of refining. For this purpose, MOL’s Danube Refinery operates a delayed coking plant, where valuable products are made by further processing residual oils resulting from previous refining processes.
The composition of the raw materials used is a decisive factor in the process of delayed coking. It has posed a major challenge for the professional staff of MOL’s Danube Refinery to determine what composition of raw materials to use in the coker in order to avoid steam eruptions during the coking process. The reason for this is that the coke released with such eruptions may enter into further stages of the refining process, and when this happens, the process must be stopped until the coke is cleaned from the equipment. In addition, steam eruptions may also damage the structure of the coker, not to mention the additional damage it may cause in the other equipment and parts of the multi-story building (e.g. stairs, elevators).
A wealth of data is generated in the course of the production process: the delayed coker plant alone is equipped with more than a thousand sensors. MOL’s team was looking for a solution capable of transforming this large volume of sensor data into actionable business information. “By analyzing the available data, we try to find the correct composition of raw materials, as well as the length of time that these materials need to spend in the delayed coker for the process to be optimal,” says Tibor Komróczki, head of the process information and automation team at MOL. “The aim is to maximize the value of the end product while minimizing the occurrence and impact of steam eruptions.”
The Solution: Microsoft Azure Machine Learning
After a careful review of the market, the choice fell on the Machine Learning solution available in the Microsoft Azure cloud service. The pilot project implemented with the involvement of Microsoft Hungary’s Services division aimed to examine how modern data analysis tools can support the optimization of a process in the oil industry, and whether they are capable of supplying information with real value.
Azure Machine Learning fit smoothly with the proof-of-concept project, as it is available from Microsoft’s cloud service with pay-as-you-go pricing. Therefore, no upfront IT investments were needed from MOL, and the organization was able to quickly gain experiences with the advanced analytic tools with minimal risk. The cloud-based machine learning algorithms were conditioned by the team of MOL’s Danube Refinery with several years’ worth of historical data.
According to Komróczki, the most attractive feature of Azure Machine Learning is that it can also be used by non-IT professionals. Even though MOL’s group of experts also includes IT specialists, the process information and automation team mostly consists of electrical and chemical engineers, as well as oil industry professionals. Azure Machine Learning is also accessible to them, as it does not require software development skills or programming experience.
“Machine learning technologies require significant knowledge and preliminary learning; the main advantage of Microsoft in this field is the user-friendly interface that makes Machine Learning technologies widely accessible,” says Dániel Percze, head of BI and Big Data at the IT department of MOL Group. “Using Azure Machine Learning, a tool that is one of the easiest to learn and use on the market, we can start leveraging advanced analytics with low risk and investment, and therefore, the project was also of strategic importance from an IT point of view.”
Advanced Analytics, Not Only for Data Scientists
“Advanced analytic tools are often based on open-source technologies, such as Python or R, which can be used freely,” adds Károly Ott, BI and Big Data expert at the IT department of MOL Group. “These script languages work efficiently in the hands of data scientists and experts; however, business users expect clear and easy-to-use GUIs, and this is exactly what Azure Machine Learning offers.”
A further argument for Microsoft’s solution is that it provides users with an open tool, where they can get to know and build upon the model, in contrast with solutions based on a “black box” principle, in which case the algorithms cannot be studied or modified.
On the basis of the initial experiences, Microsoft Azure Machine Learning passed the tests at MOL’s Danube Refinery with flying colors. The machine learning algorithms have confirmed several hypotheses on optimal raw material composition and coking duration—Ott thinks that this is proof that the method works. With Machine Learning, the team of MOL’s Danube Refinery was also able to identify which parameters have a perceptible influence on the process.
“Understanding and becoming skilled with Microsoft Azure Machine Learning allows us to apply these advanced analysis methods to other business problems as well,” Komróczki explains. MOL’s Danube Refinery uses OSIsoft’s PI System, a popular process information system in the manufacturing industry. “On the basis of international experiences, we saw that Microsoft’s solution can be easily aligned with OSIsoft PI, and that the two systems work well with each other,” Komróczki continues. MOL aims to implement a comprehensive data analysis system which can perform further tests based on the information derived from the OSIsoft PI system, and uncover continuous statistical insights within the refinery systems.
Algorithms Preserve Experiences
MOL’s Danube Refinery generates a large amount of data in the course of its operation. According to Ott, these data offer actionable information on practically all production processes. “The stored data contains the knowledge and experience our colleagues have accumulated over many years. With the help of Azure Machine Learning, we can collect and organize all this in order to preserve what we have learned.”
“Our plans for the next phase of the project include further training users on machine learning algorithms and Azure Machine Learning in particular, so that they could gain experience with model design and data analysis,” Percze explains. “We would like to provide business users with advanced analytic tools that can help them drill down into refinery data, evaluate the results, and use them to optimize production processes.”