In the gasoline refining process, maintaining the gasoline Research Octane Number (RON) is the focus of gasoline cleaning. However, due to the nonlinearity and strong coupling between the control variables in the petroleum refining process, it is difficult to measure the octane number in the product gasoline. Considering the importance of different variables to the octane number, a gasoline octane number prediction method based on adaptive variable weighting is proposed to predict the octane number. In this method, a novel variable weighting module is used to capture the correlation between the variables to obtain the variable weights, and the importance of the main variables is enhanced by the adaptive variable weighting method, and the effects of other secondary variables are suppressed. Then, considering the impact of gasoline desulfurization on the octane number, the weighted variables are input to the octane number prediction module, and the model outputs the prediction results of octane number and sulfur content. Finally, model validation is performed based on industrial data. The results showed that, compared with the neural network prediction method without variable weighting module, the neural network prediction method based on random forest algorithm and the prediction method based on variable-wise weighted stacked autoencoder, the prediction method of gasoline octane number based on adaptive variable weight has higher prediction accuracy, and it can be used to optimize the operating conditions of the gasoline refining process.