Study Of Soft Measurement Methods For 95% Point Of Light Diesel Fuel In An Atmospheric Reduced Pressure Distillation Unit

As a non-renewable resource,it is necessary to maintain the quality and quantity of petroleum products.In an atmospheric reduced pressure distillation unit,light diesel fuel can be obtained from the second lateral line of the atmospheric distillation tower and the 95%point is an important indicator of its quality.In view of the current imperfect technology,it is difficult to have a mature and low-cost instrument for real-time online measurement of the 95%point,this topic starts from the direction of soft measurement and mainly does the following research work.1.Using the box plot method to eliminate outliers,the advantages of the box plot method is that the shape of the data is not affected by outliers,and the data distribution does not have to obey a normal distribution.In order to reduce the difficulty of modelling,a layer of discrete wavelet decomposition is proposed for each dimensional input data.The method allows for adequate time-frequency analysis of the original signal,highlighting local features,mining data information and making it easier to capture the mathematical relationships between inputs and outputs.2.To address the problem that the extreme learning machine is sensitive to random initialization weights and biases,the Gaussian process regression is used to replace the hidden output matrix to find the generalized inverse,and an extreme learning machine for gaussian processes regression based on inverse multivariate quadratic kernel is proposed（iGPRELM）.Compared with the extreme learning machine model and the extreme learning machine for gaussian process regression based on gaussian kernel function（g GPRELM）,the iGPRELM model has more stable prediction results and can effectively overcome the overfitting problem of g GPRELM with improved generalisation performance.3.To address the problems that stochastic configuration networks are prone to overfitting and the generation of more redundant nodes in the implicit layer,a stochastic configuration network based on L₂ regularization and multiple population genetic algorithm（MPGA-RSC）is proposed.The L₂regularization is used to reduce the output weights and smooth the output fitting curve.Multiple population genetic algorithm is introduced to maximise the contribution of the current node to the training effect,relying on the unique supervisory mechanism of the stochastic configuration network and the one-by-one configuration of the hidden layer nodes.Compared with stochastic configuration network,and stochastic configuration network improved by L₂regularisation,MPGA-RSC has good approximation property and generalisation performance at a smaller number of hidden layer nodes.The models iGPRELM and MPGA-RSC were compared and analysed.MPGA-RSC had the largest regression fit coefficient（0.7615）and the lowest mean relative error（0.46%）for the 95%point prediction,which was better than iGPRELM.MPGA-RSC was finally selected for the soft measurement modelling of the 95%point for light diesel fuel.