Data-Driven Approach in Reservoir Production Forecast: Comparison of Machine Learning and Deep Learning

Abstract The declining production of hydrocarbons and the increasing production of unwanted fluids pose significant challenges to the economic sustainability of the oil and gas industry. Accurate forecasting of oil, gas, and water production is crucial for optimized reservoir management, cost-effective decision-making, and strategic production planning. Traditional forecasting methods—such as volumetric analysis, material balance, decline curve analysis (DCA), and reservoir simulation often rely on simplifying assumptions or computationally intensive workflows, limiting their effectiveness in dynamic reservoir environments. This study investigates the potential of machine learning (ML) and deep learning (DL) models to enhance production forecasting accuracy. Four models were evaluated: Decision Tree (DT), Random Forest (RF), Recurrent Neural Networks (RNN), and Long Short-Term Memory (LSTM). Regression-based models (DT and RF) were compared with time-series models (RNN and LSTM) using operational parameters such as wellhead pressure, downhole pressure, choke size, and tubing pressure as inputs, with oil, gas, and water production as outputs. Model performance was assessed using Mean Absolute Error (MAE), Mean Squared Error (MSE), and the coefficient of determination (R²). Results showed that all models achieved R² above 0.85, with LSTM outperforming others (R² = 0.95), demonstrating its superior ability to capture long-term temporal dependencies in production data. This study underscores the effectiveness of data-driven approaches in enhancing production forecasting, reducing costs, and improving reservoir management strategies.