Data-Driven Reservoir Performance Forecasting: Leveraging Machine Learning for Complex Reservoirs

Recent advancements in data collection, storage, and processing have facilitated the widespread use of data-driven models for forecasting reservoir and well performance, which is essential for evaluating hydrocarbon assets, making sound economic choices, and optimizing reservoir management. Data analytics workflows play a crucial role by supporting descriptive analysis, identifying variable relationships through diagnostic insights, utilizing predictive models through machine learning, and using prescriptive analytics to guide strategic decision-making for oil and gas reservoirs. This study presents an application of an integrated workflow (Artun et al. 2025) that leverages data analytics and machine learning to improve reservoir management and characterization, particularly in complex, highly fractured, and faulted reservoirs. The workflow begins with data collection and analysis, focusing on the spatial estimation of reservoir properties from well logs. A performance prediction model, based on artificial neural networks (ANNs), was developed using a data set from 19 wells. This model utilized 33 input parameters, that consisted raw well logs, estimated reservoir properties and operational/geographical features, to predict six key performance indicators related to production. The ANN model, comprising three hidden layers with a resilient backpropagation learning algorithm, demonstrated satisfactory accuracy with high R² values for both training and testing sets. The distance to the water-oil contact, active production days, porosity, depth, and permeability were identified as the most correlated variables for well performance. The importance of input parameters in the developed forecasting model, including well log and operational data, aligned with exploratory data analysis findings, confirming the model's reliability. The model was successfully applied to estimate the performance through a spatial grid, with the help of spatially estimated reservoir properties. By combining these with operational inputs, the forecasting model identified potential high-production zones in the reservoir for 2 years of production. The model particularly suggested the central and northeastern regions, and accurately predicted the performance of in-fill wells drilled in recent years.