PENGEMBANGAN MODEL PREDIKSI PRODUKSI SUCKER ROD PUMP DAN DETEKSI DINI ANOMALI

• Main Author: Winasta Sinisuka, Angelica
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/86173
• Institution: Institut Teknologi Bandung

Petroleum and gasindustry in Indonesia is currently facing challenges in optimizing production in mature fields. Machine learning is one of the potential solutions to predict oil well and to detect early anomalies in well pumps. The purpose of this research is to increase prediction performance in oil well production and to detect anomalies using machine learning models and to identify feature combinations from the combination of two different information in order to get optimal performance. In this research, the data from well production, (surface fluid flow rate), dynamometer card, and pump condition is taken from oil field X, which is located and X in Indonesia. Dynamometer in SRP (Sucker Rod Pump) records and produces graph to present measurement result from the amount of load from 1 pump cycle with respect to the position of the plunger load. The SRP performance decreases from time to time, creating data with time series characteristics. In this research, the model used for machine learning learning are LSTM (Long Short- Term Memory), Bidirectional LSTM, and Phased LSTM to predict production in oil well and also used to early detect anomalies with respect to machine learning algorithms Random Forest, Light Gradient Boosting Machine, and Adaboost to compare the performance of the ensemble algorithms. Model performance is evaluated using Mean Absolute Percentage Error (MAPE), Root Mean Squared Error (RMSE), R-squared, accuracy, precision, recall, and F1-score. Results from this research shows that the LSTM model in combination of dynamometer card features and pump conditions shows the best performance in predicting oil well productions and a score of MAPE 13.54% and R-squared 0.74. From another point of vies, The LSTM Bidirectional in combination with pump condition features and well productions shows the highest score in early anomaly detection with a score of F1 0.92. This research also shows, combination of pump condition features and dynamometer card increases prediction performance in oil well production, however it does not improve or increase early anomaly detection performance. Production of a single oil well can be influenced by the condition of the oil well reservoir and the conditions of the SRP. Therefore, measurement in oil well production must be consistent with respect to time will better facilitate prediction in oil production and early anomaly detections.