<![CDATA[Background: The accurate prediction of infill well production is one of the major bottlenecks for hydrocarbon reservoir development. Traditional reservoir simulation tools are computationally expensive, taking weeks to months per scenario. Objective: This paper presents the development of a Deep Neural Network (DNN) model for prediction with hyperparameter optimization using the Tree-structured Parzen Estimator (TPE) to predict pay porosity (PORPAYX) in infill wells of the Pertamina Hulu Sanga Sanga field. Methods: A DNN model was developed to predict oil well production based on subsurface and production features from a comprehensive dataset of Pertamina Hulu Sanga Sanga reservoir characteristics and production data. Details of our method include: training the model on a robust dataset, hyperparameter tuning using the Tree-structured Parzen Estimator (TPE), and K-fold cross-validation for performance validation. Results: Scaling normalized the data in such a way that every feature had equal influence during model training, enabling better learning and accurate prediction. In contrast, fitting the model using unscaled data resulted in an R² of less than zero (a negative score), meaning that the model could not explain the variability in the data. The mean R² score of the unscaled data model was −0.08496, along with a higher MSE = 0.009057 and RMSE = 0.095148. This was due to the model's failure to process features with varying scales, which prevented proper learning and prediction. Conclusion: Residual plots confirmed that the model trained with scaled data met the assumptions of linearity and normality.]]>
Copyrights © 2026
