Применение мультимодального трансформера для прогнозирования выходных параметров насыщенных углеводородных соединений из состава тяжелой нефти в присутствии катализаторов

Abstract

Heavy oil fields are a promising energy source in the future due to the depletion of natural sources of light oil. However, extraction, transportation and refining of heavy oil is significantly more complicated than light oil - difficulties arise at almost all technological stages. One of such stages is laboratory analytics of heavy oil and selection of the most optimal catalyst for extraction of required fractions from crude oil sample. Different catalysts are actively used in petrochemical laboratories, but special attention is paid to those of them, the basis of which is metal. In this study, catalysts based on six different metals namely zinc, nickel, copper, manganese, lead and sodium were analyzed. In order to analyze the yield ratios of the required components from the crude heavy oil composition, it is necessary to test different types of catalysts sequentially on a base sample. The yield of different hydrocarbons on a small volume of oil can be reliably estimated by chromatographic study, which takes about 68 minutes for both the base oil sample and the different catalysts. Since testing 6 different catalysts would require almost 7 hours of chromatographic analysis, a rational solution would be to apply data mining techniques to this task. A multimodal transformer model was proposed to solve this problem. It takes as input two modalities: a chromatogram of a sample of pure crude oil presented as graphical data and accompanying tabular data, which are also generated by the chromatograph and consist of text and numbers. At the output, the model produces predictive tabular data that formalize the redistributed group composition of the oil and describe both the names of the newly produced hydrocarbons and their two qualitative characteristics: time and yield area. Obtaining the prediction makes it possible to significantly reduce the time, hardware and human resources required to select the right type of catalyst in petrochemical laboratories. In the process of the study, it was found that training of the intellectual model on the data of one field allows to perform further similar forecast with acceptable accuracy for the data of another heavy oil field. The magnitude of the prediction error of the intelligent model satisfies the requirements set by the petrochemical laboratory for practical application of the multimodal transformer.