Search

Accurately predicting well productivity is crucial for optimizing gas production and maximizing recovery from tight gas reservoirs. Machine learning (ML) techniques have been applied to build predictive models for the well productivity, but their high complexity and low interpretability can hinder their practical application. This study proposes using interpretable ML solutions, SHapley Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME), to provide explicit explanations of the ML prediction model. The study uses data from the Eastern Sulige tight gas field in the Ordos Basin, China, containing various geological and engineering factors. The results show that the gradient boosting decision tree model exhibits superior predictive performance compared to other ML models. The global interpretation using SHAP provides insights into the overall impact of these factors, while the local interpretation using SHAP and LIME offers individualized explanations of well productivity predictions. These results can facilitate improvements in well operations and field development planning, providing a better understanding of the underlying physical processes and supporting more informed and effective decision-making. Ultimately, this study demonstrates the potential of interpretable ML solutions to address the challenges of forecasting well productivity in tight gas reservoirs and enable more efficient and sustainable gas production. [ABSTRACT FROM AUTHOR]

The rapid economic growth of the Pearl River Delta, one of China's most economically developed areas, has led to a sharp increase in the demand for water resources, highlighting the importance of groundwater as a potential water source. However, the distribution of groundwater resources in this area is uneven, and excessive pollution emissions and seawater intrusion pose serious threats to groundwater quality. To address this challenge, this study employed a variety of machine learning and knowledge-driven models for groundwater potential mapping and vulnerability assessment. In terms of groundwater potential prediction, five machine learning algorithms were utilized, with the Light Gradient Boosting Machine (LightGBM) performing exceptionally well. The areas characterized by high and extremely high groundwater potential are predominantly concentrated in the central plains of the study area. In contrast, the groundwater potential is lower in the mountainous areas to the east, north, and south. Factors such as vadose zone, geomorphology, and land surface temperature were found to have the most significant impact on groundwater potential. For groundwater vulnerability assessment, a hybrid knowledge-driven framework was applied. The analysis revealed that the GALDIT model effectively assessed the risk of seawater intrusion. Near the Pearl River Estuary, the risk of seawater intrusion is notably high, while areas with low seawater intrusion risk are predominantly situated inland. Due to excessive pollution discharge, the DRASTIC model struggled to predict nitrate pollution. Additionally, the eXtreme Gradient Boosting (XGBoost) was used to analyze the impact of human activities on groundwater quality, with gross domestic product (GDP) and population density identified as key factors. Areas with a higher risk of nitrate pollution in groundwater are primarily located in industrial or densely populated areas where pollution sources are highly concentrated. This study proposed four sustainable groundwater management strategies, including groundwater extraction planning and pollution control. These strategies provide essential guidance for ensuring responsible and sustainable management of groundwater resources. The study underscores the critical role of comprehensive approaches in groundwater resource management. Despite certain limitations in factors such as factor selection, data accuracy, and the construction of knowledge-driven models, this study still offers valuable insights and experiences for sustainable urban development and water resource decision-making in coastal areas. [Display omitted] • LightGBM emerges as the optimal model for groundwater potential mapping. • Identifying key factors influencing groundwater potential. • Assessing groundwater vulnerability using a hybrid knowledge-driven framework. • Determining the impact of human activities on groundwater quality using XGBoost. [ABSTRACT FROM AUTHOR]