Feature Selection Methods for Deep Learning Models of Soft Sensors in Oil Refining

The development of automated control systems results into industrial plants accumulating large amounts of data on the continuous state of technological processes. Multiple physical sensors record the system states at any given time, hence being crucially responsible for controlling the system and maintaining its parameters within hard limits. At the same time, irregularly conducted laboratory measures make up a significant part of the qualitative indicators of such processes, especially in the petrochemical industry. Mathematical models that generalize laboratory measured indicators to match the frequency of physical sensors are called soft sensors. On practice, soft sensors for laboratory data are represented by linear or last-recorded-value models. We investigate the task of analytically obtaining chemical indicators of the technological process in real time based on the values of physical sensors; the study is conducted on a real-world data set. Several problems are covered, including high dimension of the physical inputs compared to the laboratory data volume; scarcity of the laboratory data collected on a daily basis. Authors propose feature selection methods based on PLS regression (hierarchical clustering), Bayes trees, utilize existing graph neural network, as well as compare developed methods with existing popular approaches. Each of the proposed feature selection methods has been adapted to take into account the expert opinion of the industrial plant engineers. Authors investigate developed approaches alongside neural network methods for predicting time series including graph neural networks, fully connected and recurrent networks. The obtained experimental results show the advantage of using proposed feature selection based on PLS and Bayes in ensemble with simple recurrent networks or graph neural networks with preliminary interpolation. Separately, it is worth noting the ambiguity of assessing the developed models quality; authors propose a combined approach that takes into account the adequacy of the model, its correlation with the true laboratory values and averaged errors.