This study revisits the field observations of sand production of gas wells in the Adriatic Sea to develop comprehensive correlations of key parameters and identify the most critical factors influencing the onset of sand production. The primary objective is to leverage profound data analysis and machine learning (ML) techniques to boost predictive consistency and provide actionable insights for sand-free production practices.
The research employs a two-pronged approach. First, a detailed data analysis is conducted, featuring vivid cross plots to illustrate relationships between cohesion strength and various parameters such as depth, interval transit time, cohesion strength, original static reservoir pressure, effective overburden stress, depleted pressure, total drawdown pressure, perforation interval, shot per foot, and produced gas and water rates. Second, the study applies ML models, including Gradient Boosting (GB), XGBoost (XGB), Random Forest (RF), CatBoost (CATB), and Support Vector Machine (SVM), to categorize data and identify the most critical features impacting sand production onset. SMOTE is utilized to rectify class imbalance, and GridSearchCV optimizes model parameters. Additionally, LOOCV is employed for rigorous model evaluation. SHAP analysis further interprets the model results.
The cross plots reveal significant relationships between cohesion strength and other essential metrics, providing thorough knowledge of their interactions. In the ML classification phase, the GB model achieves a perfect accuracy of 1.0, significantly outperforming XGB (0.86), RF (0.71), and CATB and SVM (both 0.57). The GB, XGB, and RF models identify critical features such as interval transit time, cohesion strength, and water production as major influencers on the onset of produced sand. SHAP analysis further elucidates the contribution of these features, offering interpretable insights into their importance. The findings suggest that ML models, particularly GB, can effectively predict sand dislodgement, thus aiding in developing sand production preventive strategies.
This paper introduces an alternative application of ML techniques to foresee sand production onset, an essential challenge in oil and gas well management. The study provides new insights and enhances predictive accuracy by integrating extensive data categorization with advanced classification models. The use of LOOCV ensures rigorous model evaluation, and SHAP analysis for model interpretation offers a transparent understanding of feature importance, making the findings highly valuable for sand control approaches. This research contributes additive information to the existing literature, emphasizing the potential of data-driven approaches in improving operational decisions in long-term production.