Abstract: Petroleum is one of the primary energy sources in modern society. As the leading process in the refining industry, atmospheric distillation requires real-time monitoring of its process to ensure efficiency and safety. Gasoline final point refers to the temperature at which gasoline components begin to transition from liquid to gas during the crude oil distillation process, and it is a key indicator for measuring the quality of the final product. This paper introduces and evaluates the performance of the Gaussian error linear unit (GELU), proposing the use of GELU as the activation function to replace the ReLU in the temporal convolutional network (TCN). Additionally, it modifies the residual structure to construct the Gaussian-TCN model. Samples were collected on the atmospheric tower top gasoline final point and its influencing factors at a refinery. The partial least squares (PLS) was used to reduce the dimensionality of the high-dimensional independent variable data, completing the selection of auxiliary variables for the gasoline final point. Predicting the gasoline final point at the top of the atmospheric tower using the constructed Gaussian TCN soft sensing model. Simulation results verified that the proposed model has a significantly better fit and prediction accuracy compared to traditional TCN prediction models, providing valuable insights for the high-efficiency development of the refining industry.

Key words: Gaussian error linear element, temporal convolutional network, soft sensing, gasoline final point