محدثه مشهدي نجفي

Accurate prediction of oil production rates in hydrocarbon reservoirs, especially in complex an‎d heterogeneous fields, is essential for optimal reservoir management an‎d economic planning in the oil industry. The goal of this research is to improve production forecasting accuracy by employing advanced machine learning techniques. In this study, data preprocessing (removal of incomplete data using the Isolation Ran‎dom Forest (IRF)), normalization, feature engineering, an‎d model training were performed. To achieve the objective, this study was designed in two separate scenarios for the Volve field. In the first scenario, data from a single well (Well No. 4) was used to focus on accuracy, while in the second scenario, data from five wells were combined to allow the models to learn from a more diverse set of temporal an‎d spatial patterns. The machine learning models tested include Linear-XGBoost-Linear, the Bidirectional Long Short-Term Memory (Bi-LSTM) hybrid model, the Convolutional Neural Network (CNN) hybrid model, an‎d the CNN-LSTM hybrid model, all optimized with Particle Swarm Optimization (PSO) an‎d compared against each other. To eva‎luate the model performance in the scenarios, R², MSE, MAE, an‎d RMSE metrics were used. The eva‎luation results indicated that within-scenario, the CNN-LSTM algorithm outperformed the Linear-XGBoost-Linear algorithm with an R² of 0.9947 compared to 0.9922. Furthermore, when comparing across scenarios, the CNN-LSTM algorithm showed superior performance with an R² of 0.9968 compared to the other models.

پيش‌بيني توليد، داده‌هاي سري زماني، CNN-LSTM ، بهينه‌سازي ازدحام ذرات (PSO)، ميدان نفتي Volve

Production forecasting, Time-series data, CNN-LSTM, Particle Swarm Optimization (PSO), Volve oil field

Mohhadeseh Mashhadi Najafi

Dr. Forough Ameli