Perpustakaan Digital ITB

Electric Submersible Pump (ESP) is widely used as an artificial lift method in oil production; however, its design and operation often face challenges such as unexpected failures that lead to costly downtime. Conventional ESP design workflows rely heavily on manual calculation or closed commercial software, offering limited transparency and flexibility. This research aims to develop an integrated system that automates ESP design and optimization using a multimodal artificial intelligence framework combining Large Language Models (LLM), Machine Learning (ML), and Retrieval-Augmented Generation (RAG), and applies it to real field cases. The system was built with four main modules: (1) automated ESP design powered by LLM with function-calling for precise backend calculation, (2) a RAG module for contextual failure diagnosis based on historical remarks, (3) a classification model using XGBoost to predict ESP run life, and (4) a user interface for interactive evaluation. The model was trained using 1,443 historical ESP records from the Offshore Southeast Sumatera (OSES) field, with feature engineering applied to capture Best Effective Point deviation, gas indicators, and operational loading. Troubleshooting data was used in raw form and interpreted through retrieval and summarization. Case studies on Well-11 and Well-07 were conducted to test the system’s performance under real-field conditions. The system successfully recommended ESP configurations that either aligned with or improved upon actual field implementations. The RAG module accurately diagnosed underload and power outage events, while the ML model achieved 81.1% soft accu4racy and 61.4% F1-score in run life prediction. Comparative analysis showed that the automated design system provided more efficient sizing and better justification for component selection, particularly in gas-prone wells. In conclusion, the system fulfills its objectives by enabling accurate, transparent, and explainable ESP design, failure reasoning, and predictive analysis. It also demonstrates potential for integration into real-time operations and further development with structured troubleshooting data and economic decision support. Keywords: ESP Design, Large Language Model, Machine Learning, Retrieval-Augmented Generation, Run Life Prediction, ESP Troubleshooting