A Casing Damage Prediction Method Based on Principal Component Analysis and Gradient Boosting Decision Tree Algorithm

Abstract

The examination and prevention of casing damage is an important work in oil and gas field development projects. Casing damage will directly influence the production of oil and gas, the effect of water injection and the life cycle of oil, gas and water well. With the number of casing damage wells increasing year by year in each oil field, inspections and precautions of casing damage status have become more and more important during development of the filed. With the development of big data technologies, we can predict the casing damage based on historical and real-time data, and optimize the maintenance intervals of casing. In this paper, we proposed a casing damage prediction method based on principal component analysis (PCA) and gradient boosting decision tree (GBDT) algorithm. Building a three-nodes Spark big data platform, we tested our proposed method on a dataset in an oil-field in mid-east China. Firstly, based on data analysis and expertise, selected 10 parameters which affect the casing damage most, including casing outside diameter, wall thickness, perforation density etc. Secondly, using PCA to reduce the dimension of parameters. Thirdly, building the algorithm model of casing damage risk assessment by GBDT, training and optimizing the model parameters. Fourthly, using the proposed model for casing damage prediction. Using precision and AUC(Area Under ROC Curve) to evaluate the proposed method, and compared with traditional methods, including decision tree, logistic regression and Naïve Bayes, the experiment results show that compared with traditional methods, the proposed method gain 86.3% precision and AUC is 1, both higher than traditional methods, which means it has better prediction precision and performance. The proposed method also can apply to predict the probability of a normal well becoming a casing damage well, and finally can provide the decision support for on-site operations. The proposed method using data-driven idea to provide the decision support for oilfield operations, and lay the foundation for the construction of intelligence oilfield.