At the facilities for storage of oil products are often in the production process there is a mix of various types of petroleum products. However, if the mixed oil products are incompatible, then there are qualitative losses, namely, the active form of the total sediment. The result of incompatibility:
The International Maritime Organization (IMO) Committee for the Protection of the Marine Environment (MEPC) has approved a ban on the transport of marine fuels with a sulfur content of more than 0.5%, which will enter into force on March 1, 2020. Determining the composition of the fuel and the presence of sediment in it is a very expensive process(the most commonly used method is chromatography, the cost is about 1 million rubles per sample). Using historical data containing information about the parameters of fuel and various mixtures, and the functionality of the AutoML framework FEDOT, it will be possible to model mixtures with the lowest sediment content.
The disadvantages of this approach:
_Solution: Since the final goals are to predict the amount of sediment, as well as the classification of fuel mixtures into dangerous and non-dangerous, from the point of view of precipitation, this task can be simultaneously interpreted as both regression task and classification task. The formulation of this problem in the form of a machine learning task is shown in Table 1.
Table 1. Example of a data format for machine learning problem statements.
|First sample for the mixture||Second sample for the mixture||Type of Machine learning task||Physical interpretation of target variable||Value of target variable|
|Number sample No. 13||Number sample No. 37||Regression||Percentage of sediment in the mixture||0.17|
|Number sample No. 13||Number sample No. 37||Classification||The class of mixture density sludge||1 class|
For our training data set, we used information about the physical and mechanical properties of the fuels in the mixture, information about their storage and transportation conditions, and information obtained during laboratory tests. In total, we managed to form seven features for each fuel mixture. The format of the training data set for the regression task is shown in Table 2.
Table 2. Example of a training data set for machine learning regression task.
|Number of mixture||Feature 1||Feature 2||Feature 3||Feature 4||Feature 5||Feature 6||Feature 7|
As can be seen from Table 3, the use of the Fedot the framework allowed us to obtain higher values of quality metrics for both the regression problem and the classification problem. Thus, we can say that the use of the AutoML approach, using the Fedot framework, for this task, has a pronounced economic effect and expediency in terms of modeling sedimentation processes.
Table 3. Results of applying the Fedot framework, and the baseline XGB model to the generated data set.
|Type of ML model||ROC-AUC||F1-score||MSE||R2|
|Baseline XGBoost model||0.737||0.714||0.215||0.723|