Perpustakaan Digital ITB

Indonesia's oil and gas industry is currently facing challenges in optimizing production from mature fields. Machine learning has emerged as a potential and easily applicable solution for predicting oil well production, especially in the presence of operational issues such as problems with sucker rod pumps (SRP). This research is a strategic effort aimed at improving oil well performance by identifying SRP-related issues using several selected machine learning models to achieve optimal performance. To accomplish this, dynamometer card data is required, which is now commonly collected through the SCADA (Supervisory Control and Data Acquisition) system. The implementation involves developing an end-to-end cyber-secure architecture, where a software platform integrates edge gateways and cloud systems to collect dynacard data in real-time from the Remote Terminal Unit (RTU). The surface dynamometer on the SRP records and generates graphs that are then converted to represent the measured load per pump cycle in relation to the plunger load position. Since SRP performance tends to decline over time, the data is treated as time series data. The main machine learning architecture used in this study is a Convolutional Neural Network (CNN). Additionally, pretrained CNN models such as VGG16, MobileNetV2, and NasNet are employed to compare performance and support an ensemble modeling approach. The model performance is evaluated using the binary cross-entropy loss, along with evaluation metrics such as subset accuracy, precision, recall, and sample-based F1-score, tailored for multi-label classification. The results show that the customized CNN model combined with pretrained models provided the best performance in predicting issues in SRP, achieving a subset accuracy of 20%, precision of 73.40%, recall of 72.87%, and an F1-score of 72.87%. Therefore, an intelligent control system integrated with machine learning is expected to be able to monitor pump conditions in real time, maintain production efficiency, and reduce operational expenditures (OPEX) during this era of digital energy transition.