Determination of Chemical Mixture Viscosity by Means of Machine Learning Methods

Mathematical modeling of petrochemical, oil refining and oil treatment processes is an integral part of the systems of pre-project analysis of newly created, operability of operating plants, as well as systems of automated control of technological processes. Chemical-technological processes are multi-level complex systems, the hierarchy of which varies depending on the process in progress. Each of the hierarchy levels is represented by a system of equations of varying complexity, which is further reduced to a system of nonlinear equations including transcendental functions. The solution of such systems requires considerable time even when using modern computational methods, which is unacceptable, especially when used in advanced process control systems. In this paper we propose a method of calculating the viscosity of a broad fraction of hydrocarbons based on the use of machine learning models, namely neural networks. In order to accelerate training and obtain the result, the decomposition of the initial problem into a set of tasks, less parametric, which facilitates the process of training the model. Thanks to this decomposition, for a number of regions of the feature space, there is no need to calculate viscosity. To solve the problem of calculating the viscosity of a chemical mixture, a program complex was created to ensure the interaction of trained neural networks.