Surveillance Data Acquisition Planning Maximizing Expected Value Using Machine Learning

Abstract A robust reservoir surveillance strategy is essential for optimizing Well performance, scheduling interventions, and monitoring reservoir health. This study aims to enhance reservoir monitoring by leveraging data-driven methodologies. It evaluates the forecasting accuracy of key reservoir parameters, investigates clustering algorithms for identifying diagnostic patterns in well data, and standardizes monitoring planning by categorizing wells based on their diagnostic needs. The goal is to develop an efficient and scalable surveillance strategies for improved reservoir performance understanding and management. The methodology involved a comprehensive data preparation phase, including data collection, dataset preparation, and an overview of field and well performance. This was followed by building models for forecasting surveillance parameters, clustering analysis for well diagnostics, and generating a reservoir surveillance plan. The forecasting process utilized time series analysis, customized feature engineering, and spatio-temporal modeling. The principal component analysis (PCA) was applied for dimensionality reduction followed by and model training, testing, and evaluation to ensure accurate predictions. Additionally, a value of information (VOI) methodology was applied through spatio-temporal modeling to analyze surveillance scenarios comparing minimal versus comprehensive data acquisition strategies. The study achieved good predictive accuracy for crucial reservoir parameters, such as oil production rates, water saturation levels, cumulative production, and well activity status, through various machine learning algorithms enhanced with feature engineering, dimensionality reduction, and validation techniques. These methods provided robust forecasts of future well performance. The findings underscored the importance of understanding reservoir dynamics before model development, highlighting the impact of reservoir heterogeneity, well positioning, and water injection practices on well performance. Clustering analysis effectively streamlined well performance analysis at the cluster level, aiding the standardization of reservoir surveillance plans by categorizing wells based on their historical performance. This approach provided a more robust surveillance plan. Spatio-temporal predictions emphasized the impact of reduced data on forecast accuracy, underscoring the importance of comprehensive data collection to minimize uncertainties. By addressing these areas, future research can further improve the application of machine learning in reservoir surveillance planning, contributing to more efficient and sustainable production strategies. Future work will focus on refining these models, incorporating more robust time-series analysis, and leveraging real-time data to enhance reservoir management strategies. The study showcases valuable application of data-driven and machine learning approach for Well Performance analysis in driving future surveillance requirements. The clustering analysis offers a novel approach to categorizing wells based on their historical performance and providing more robust surveillance plan. Integrating various machine learning methodologies enhances reservoir monitoring and management strategies, providing a scalable solution for continuous surveillance and sustainable production.