Soft sensor modeling for fraction yield of crude oil based on ensemble deep learning

Abstract

Soft sensor modeling of crude oil physicochemical properties, especially fraction yield, is very important for refinery. Traditional method of properties evaluation in the laboratory is time-consuming and the near-infrared spectrum technology has shortcoming of low accuracy and strict application conditions. To solve these problems, the nuclear magnetic resonance (NMR) is introduced to properties evaluation of crude oil for refining process, where soft sensor modeling between the 1H NMR spectral data and the fraction yield plays an important role. This paper presents an ensemble deep learning based soft sensor modeling approach for fraction yield evaluation of crude oil. In this approach, a pre-processing method of spectral data is proposed by changing data structure, increasing data redundancy and generating virtual samples. Then, the model structure and its implementation are presented, which consists of a deep learning network, a near neighbor learning part and a random vector functional-link (RVFL) network based ensemble strategy. To avoid over-fitting, the training algorithm using k-fold cross-validation of ensemble deep learning model is also given. The experimental results show that the proposed model can learn the 1H NMR spectral data features well and improve accuracy for crude oil fraction yield evaluation.