Prediction of Casing Damage in Unconsolidated Sandstone Reservoirs Using Machine Learning Algorithms

Abstract

Despite numerous studies in the subject matter, there is no mature casing damage prediction method based on historical casing damage data in the oil and water well production stage for unconsolidated sandstone reservoirs. We use two popular algorithms to establish a prediction model for the sand-sand casing damage area, eXtreme Gradient Boosted Trees (XGBoost) and Light Gradient Boosted Trees (LGBM). According to data analysis and casing damage mechanism, we selected 19 casing damage factors for oil wells and 18 for water injection wells. Geological, reservoir, completion and historical production/operation data for 653 production layers and 212 injection layers in Gangxi Oilfield are collected to form dataset. Among them, the casings of 91 production layers and 22 injection layers were damaged. The dataset is split into 80% training and 20% holdout datasets. A training dataset is split into 10-fold cross validation. Two machine learning algorithms are evaluated predicting casing damage and their performance is compared. For production wells, the prediction accuracy of LGBM model is higher, up to 95.4%. For injection wells, the prediction accuracy of LGBM model is higher, up to 100%. Therefore, we can use more accurate model to predict casing damage in unconsolidated sandstone reservoirs, and determine main controllable factors of casing damage in risk wells, so as to provide technical guidance for technicians to take preventive measures.