Abstract
The use of machine learning techniques for production forecasting in the oil and gas industry has become increasingly prominent due to advancements in data conditioning and computational power. Traditional methods, such as decline curves, have long been the dominant approach for predicting production rates in single-phase wells. However, their inability to account for varying operational scenarios and field uncertainties presents significant limitations. This paper presents the application of a machine learning-based framework, specifically combination of Neural Network models, to predict the future performance of producing wells in several secondary drive reservoirs.
The study involved training and testing of multiple machine learning models using real production data from a cluster of wells in water and gas injection reservoir. A subset of well data was reserved for testing, excluded from the training set, to evaluate the model's capacity to learn the fundamental patterns in the time series data and to accurately model the non-linear relationships for prediction. The field data from selected reservoirs, characterized by secondary drive mechanisms, were used for predictive analysis on both individual wells and clusters of wells.
This case study demonstrates the superior performance of the machine learning model compared to conventional decline curves, particularly in scenarios with fluctuating operating conditions. The proposed technique yielded highly accurate predictions and demonstrated robust learning capabilities. This work also highlights limitations in selecting a single nonlinear equation for all reservoir drives and well types. Furthermore, this study highlights the critical role of model-optimization techniques in influencing prediction accuracy on production forecast. The inclusion of numerical simulation data, alongside historical data, enabled the model to identify relationships among key parameters, enhancing its predictive power.
This work provides empirical evidence that machine learning models, when properly optimized, offer a robust and reliable method for production forecasting, addressing complexities that traditional statistical methods often fail to capture. Numerical simulation models are not sufficient in reservoirs with secondary drive mechanism and frequently changing surface conditions. However, it is also important to utilize the key results from various modes (empirical, material balance, finite element and data science) for uncertainty management in production forecasting. The procedures and approaches developed in this study can be extended to other reservoir types, with appropriate adjustments to neural network structures and optimization techniques based on specific operational conditions. The paper outlines specific cases where ML models can significantly reduce prediction errors, serving as a guide for petroleum engineers to identify and adjust key parameters that enhance the realistic predictability of ML models. Ultimately, achieving forecasting accuracy through artificial intelligence lies in the right balance of functional domain expertise and the computational power to optimize and apply these models effectively.