Prediction Of Gel Point And Viscosity Of Mixed Crude Oil Based On Machine Learning

In recent years,the volume of crude oil transported by pipeline in China has been increasing year by year,and the scale of pipeline network has been constantly improving.The flow guarantee technology for pipeline operation safety has always been the main research topic in oil and gas storage and transportation field.Gel point and viscosity are two key parameters to measure the fluidity of crude oil.Accurate prediction of gel point and viscosity provides important reference for the flow assurance technology of piped mixed crude oil.At present,the empirical methods for calculating the gel point and viscosity of mixed crude oil can be divided into two types of models according to whether they depend on the equal ratio parameters of two components or not.However,these two models usually can not give consideration to accuracy and convenience,and most of them do not have universal significance,so they need to be optimized for actual oil products,which brings some inconvenience to engineering application.Thanks to the rapid development of big data and artificial intelligence,machine learning methods show obvious advantages in predicting complex systems,providing a new idea for predicting gel point and viscosity of mixed crude oil.Data is the foundation of machine learning.The data sources used in previous machine learning studies are too broad and not targeted to the mixing problem of pipeline crude oil,so the prediction results obtained lack practical engineering significance.At the same time,the research on data preprocessing is not deep enough.In terms of model,the commonly used neural network also has some shortcomings,such as complex structure and parameter setting and unsatisfactory prediction accuracy,so it has not been widely used in engineering.Aiming at the above problems,this paper follows the basic ideas of machine learning modeling at present stage and establishes a prediction method consisting of data pretreatment,model training,parameter optimization and performance evaluation.(1)The historical working condition data of A crude oil pipeline in China are selected,and the original data are integrated based on the analysis of the crude oil pipeline’s external transport mode and experimental data characteristics.Then,data cleaning,stratified sampling and min-max standardization methods are introduced to reduce the error caused by data during model training.Finally,Pearson correlation analysis was used to screen the input characteristics of the model to improve the training efficiency.(2)Considering the characteristics of high feature dimension and nonlinear correlation of the original data,the cutting-edge machine learning algorithms in recent years are investigated.Finally,four machine learning technologies,SVR,MLP,Xgboost and LightGBM with different principles,were used to establish the prediction models of gel point and viscosity of mixed crude oil respectively.And the Bayesian hyperparameter optimization method was introduced to realize the automatic parameter tuning of the algorithm and improve the prediction accuracy.(3)Traditional machine learning methods have been criticized for their poor interpretability.In this paper,SHAP value method and Optuna framework were introduced to analyze the influence of input characteristics and hyperparameters on the predictive performance of the model,which enhanced the interpretability of the model and helped researchers to understand,use and diagnose the model.The traditional empirical model and new machine learning model are evaluated.The results show that among the four machine learning models,the Bayesian Xgboost model has the highest accuracy in the test set,with an average relative error of 8.5%,slightly lower than the Cragoe-C-B fractional correction model(8.2%),which has the highest accuracy in the viscosity empirical model.The average absolute error of the model is 0.38℃,which is higher than that of Liu Tianyou’s model I(0.81℃).Therefore,it is recommended to use the Bayesian Xgboost model when the parameters of the equal ratio of two components of crude oil cannot be known.Compared with the neural network,Xgboost has obvious advantages in prediction accuracy(74% higher gel point and 58% higher viscosity).Compared with the previous manual parameter tuning process,Bayesian automatic hyperparameter optimization can more conveniently and effectively improve the prediction accuracy of the model(25% increase for Xgboost viscosity model,43% increase for gel point model).The maximum gel point prediction error of Xgboost is 1.23 ℃,which is still lower than the industry-specified error standard,indicating that Xgboost has strong stability.The new method can accurately predict the gel point and viscosity of pipeline mixed crude oil,and also has certain reference significance for intelligent flow assurance technology of crude oil pipeline.