罐式批次成品汽油调和配方集成建模方法

考虑罐式批次成品汽油调和过程中罐底余油对产品质量指标的影响，本文提出了一种基于改进多核模糊C均值（multi-kernel fuzzy C-means，MKFCM）与极端梯度提升树（extreme gradient boosting，XGBoost）集成的成品汽油调和的通用配方建模方法。该方法首先考虑各调和组分的差异性，提出一种自适应核参数计算方法对MKFCM改进，并将其用于罐底油聚类分析，旨在最大程度划分出性质相近的罐底油批次类型；在此基础上择优选用XGBoost算法，以各批次罐底余油的组分比例和产品预期质量指标作为输入，建立各批次子配方模型；在配方生成时，基于改进MKFCM求得当前罐底余油的隶属度向量，并以此为权值对子配方模型进行加权融合，最终以多模型集成的方式得到了成品汽油调和的通用配方。经使用某企业实际工业数据进行实验分析，结果表明，较单一模型或未改进MKFCM的集成模型，基于改进MKFCM-XGBoost的多模型集成配方，预测精度和泛化能力均更优，更适合罐式批次成品汽油调和过程。

Integrated modelling method for tank-batch finished gasoline blending formulations

During the blending process of tank batches of finished gasoline,an integrated modeling method for finished gasoline blending general formula was proposed by considering the influence of the remaining oil at the bottom of the tank on product quality indicators, which based on an improved multi-kernel fuzzy C-means (MKFCM) and an extreme gradient boosting tree (XGBoost). Firstly, considering the differences of the blending components, an adaptive kernel parameter calculation method was presented to improve MKFCM, and then it was used for tank bottom oil cluster analysis in order to classify tank bottom oils with similar properties to the greatest extent batch-type. Secondly, XGBoost algorithm was selected on this basis. The sub-formulation model of each batch was established by taking the component proportion of residual oil at the bottom of each batch of tank and the expected quality index of the product as the input. When generating the formula, the membership vector of the current tank bottom residual oil was obtained based on the improved MKFCM and the above membership vectors were used to weight and fuse the sub formula model. Finally, the blended general formula of finished gasoline was obtained by multi-model integration. The experimental analysis by using the actual industrial data of an enterprise showed that compared with a single model or an integrated model without the improved MKFCM, the multi-model integrated formula based on the improved MKFCM-XGBoost possessed better prediction accuracy and generalization ability, and it was more suitable for tank type process of blending batches of finished gasoline.