Multivariate Time Series Forecasting of Oil Production Based on Ensemble Deep Learning and Genetic Algorithm

Forecasting oil production is a substantial task in the petroleum industry as it helps decision-makers optimize storage and distribution operations and plan resources more efficiently. However, traditional methods for forecasting oil production, such as Numerical Reservoir Simulation (NRS), can be challenging due to the substantial effort involved and the high uncertainty associated with the various types of data used. Alternative methods, such as analytical methods and Decline Curve Analysis (DCA), fail to accurately reflect the physics of the actual system or account for dynamic changes in oil production operations and conditions. Therefore, more efficient methods are needed. In this study, an ensemble deep learning model composed of a Temporal Convolutional Network (TCN) and Long Short-Term Memory (LSTM) has been proposed, and its hyperparameters were optimized with Genetic Algorithm (GA). The workflow of this study involved extensive preprocessing to ensure the quality and relevance of the input data. As a result, only the interaction terms of average choke size, on stream hours, and time, in addition to gas volume, were utilized in the model development. To verify the robustness of the proposed model, its predictive performance was compared with four other models: LSTM, TCN, GRU, and RNN, using a testing set. The GA-TCN-LSTM model proposed in this study demonstrated promising results, reducing residual variance and outperforming the reference models with an RMSE of 199.39, wMAPE of 5.13, MAE of 117.11, and  of 0.93. Moreover, the proposed model was established using only three consistently available variables with oil production. These input features covered various operating conditions, making the proposed model applicable to most conventional oil fields.