As a trademark of gasoline, octane number is the most important indicator reflecting the combustion performance of gasoline. In view of the flaw of effective utilization of heavy oil resources, with octane as an evaluation index, linear interpolation and Laida criterion are used to preprocess the data set, meanwhile principal component analysis and hierarchical clustering are adopted to obtain the main variables for dimension reduction. An octane number loss prediction model based on BP neural network has been established for the prediction of the octane number and its indicators. Finally, the genetic algorithm is used to iteratively obtain the optimal parameter set, with the change trajectory displayed visually. Based on the experimental results, with the number of iterations of the prediction model being 2 048, the loss of the model is maintained at 9.982 6×10-6 . The sulfur content selected from the data set is no more than 5 μg/g and the main operating variables of the samples are unchanged within the value range. According to the constraints of the value range and step size of the main operating variables, the optimal operating conditions corresponding to the main variables are found to meet an octane loss of more than 30%. The research results show that the adopted mathematical model and optimization algorithm can be used for the prediction and optimization of the octane number loss problem.