Machine Learning-Based Time Series Models as Alternatives to Short-Term Traditional Decline Curve Models for Hydrocarbon Reserve Estimation

Abstract

Abstract This study evaluates the effectiveness of machine learning-based time series models as alternatives to short-term traditional decline curve models for estimating hydrocarbon reserves. To accurately estimate the hydrocarbons that can be economically recovered from a field, area, or region, the predicted quantities should closely match the actual observed quantities within the same period. In this study, two models were compared based on their Root Mean Square Deviation (RMSE) to solve the decline curve technique of reserve estimation – the traditional exponential model and the time series ML-based Recurrent Neural Network's Long Short-Term Memory (LSTM) model. The study results showed that the LSTM model outperformed the traditional exponential model, with an RMSE of 80.12 compared to 107.41 for reservoir K3, 30.24 to 141.52 for reservoir VII, and 80.56 to 169.81 for reservoir K5. These RMSE values indicate that the LSTM model had a better fit to observed data and thus had better goodness. Therefore, LSTMs serve as improved alternatives to short-term traditional decline curve models.