Neural network-based production forecasting : dynamic modeling of reservoir parameters using the capacitance-resistance model



Journal Title

Journal ISSN

Volume Title



Reservoir characterization has historically relied on the Capacitance-Resistance Models (CRM) to provide insights into reservoir properties and dynamics. This thesis introduces a novel approach to reservoir characterization by proposing a dynamic model that simulates the productivity index as a function of time. Through rigorous experimentation and analysis, the robustness of the proposed model is demonstrated across diverse scenarios. This includes testing across diverse scenarios, ranging from synthetic data generated using CMG to real-life datasets from North Sea fields. Our dynamic model stands out for its predictive power, enabling the forecasting of reservoir behavior and production rates many days into the future. Its proficiency in capturing the temporal evolution of productivity indices provides a nuanced understanding of reservoir dynamics, allowing for proactive management and optimization of production strategies. Notably, the model reliably back-calculates preset parameters with high accuracy - even in the presence of noise - underscoring its reliability in real-world applications. This work opens up more possibilities for future research in refining neural network architectures for reservoir modeling, exploring the potential of transfer learning, and incorporating more complex dynamic reservoir properties. The implications of this research extend far beyond reservoir engineering. The underlying concept of dynamically parameterizing differential equations has universal applicability across various systems and forms of CRM, aiding in research in fields where system dynamics are governed by similar mathematical frameworks.


LCSH Subject Headings