Abstract: A multi-model adaptive integrated modeling method was proposed to evaluate the quality of finished gasoline blend in order to avoid the batch effect. Firstly, the batch effect caused by the difference of origin attributes of main materials and the limitations of PRFCM clustering algorithm in the application of multi-attribute data analysis are considered. By focusing only on the nonlinear relationship between the key attribute variables in the formulation and replacing the Euclidean distance with the Mahalanobis distance, an improved IPRFCM algorithm is obtained. Based on this algorithm, the historical data are divided into batches in the offline stage, and the NGBoost algorithm is selected to build batch sub-models. The batch membership degree of the current formula is calculated in the online stage, and the IPRFCM-NGBoost integrated model is obtained by fusing each submodel as the weight coefficient, which realizes the dynamic adaptive prediction of the quality of the blended formula. The experimental results of industrial data show that IPRFCM has clearer batch division results, a more accurate fusion coefficient, and lower calculation cost for the multi-attribute problem in the quality prediction of blended formula. Compared with traditional machine learning or deep learning models, IPRFCM-NGBoost has better prediction accuracy and generalization ability, especially when formulation sample sizes are limited.

Key words: tank batch blended formulation, natural gradient boosting, adaptive weighting, integrated model, quality prediction