Interpretable knowledge-guided framework for modeling reservoir water-sensitivity damage based on Light Gradient Boosting Machine using Bayesian optimization and hybrid feature mining

Abstract

Reservoir water sensitivity damage significantly contributes to production declines in low-permeability oil and gas fields. An accurate and rapid assessment of water sensitivity is essential for effective mitigation or prevention strategies. Facing the intricate challenge of predicting high-dimensional water sensitivity damage, this study leverages trends in intelligent drilling and completion technologies. It adopts a Knowledge-guided, Bayesian Optimization-enhanced Light Gradient Boosting Machine (KBO-LightGBM) for modeling, augmented by Multiple Imputation by Chained Equations and Synthetic Minority Over-sampling Technique (MICE-SMOTE) to address missing and unbalanced data issues in oil and gas fields, thereby enhancing the scientific efficacy of data processing. The framework's precision and practicality were confirmed using data from 270 natural core samples across 15 oil fields. Findings include a correlation coefficient of 0.9679 on the test set, a root mean square error of 3.4797, and a mean absolute percentage error of 4.0936%. Interpretability analysis identified formation water mineralization, burial depth, feldspar content, and initial porosity as the predominant factors affecting water sensitivity. This research distinguishes itself with a broader dataset, covering 15 parameters of formation fluids and rock components. Weighting factor αDK and scale βDK were designed to integrate empirical correlations into LightGBM's loss function, theoretically mitigating model overfitting. Hence, the intelligent framework proposed herein accurately predicts reservoir water sensitivity damage, aiding in the development of reservoir damage control strategies.