Perpustakaan Digital ITB

Electrical submersible pumps (ESPs) play a significant role as artificial lift systems in the oil and gas industry, installed in over 33% of operational wells and contributing approximately 60% of global oil production. Despite their widespread use, ESP systems are highly vulnerable to unplanned failures caused by electrical faults, mechanical degradation, and gas interference, often leading to costly production losses and workover operations. Conventional reactive and fixed interval preventive maintenance strategies are insufficient for early fault detection, motivating the development of a more proactive, data driven predictive maintenance approach. This study proposes an AI-based framework for early failure detection of ESP systems integrating downhole sensor forecasting, physics based virtual flow metering, multi classifier ensemble failure prediction, and a decision support dashboard. Three deep learning architectures Long Short Term Memory (LSTM), Bidirectional LSTM (BiLSTM), and Gated Recurrent Unit (GRU) were evaluated to identify the most suitable model for each of six downhole sensor parameters: discharge pressure, intake pressure, intake temperature, motor temperature, average ampere, and vibration. Data from seventeen ESP-equipped wells, recorded at ten minute intervals from April to July 2020, were used for model development and validation. BiLSTM achieved the best forecasting performance for four parameters, while LSTM and GRU were optimal for discharge pressure and vibration, respectively. All selected models produced Mean Absolute Percentage Error (MAPE) values below 10%, confirming highly accurate three day ahead sensor predictions. For wells with incomplete production data, liquid flow rates were estimated using the Power Equilibrium Equation, yielding a deviation of only 5.56% against available well test measurements. For failure classification, five supervised learning algorithms Logistic Regression, Random Forest, Support Vector Machine, K-Nearest Neighbor, and Decision Tree were trained on slope based sensor features with SMOTE oversampling and Optuna hyperparameter optimization. Random Forest and Decision Tree achieved the highest test accuracy of 99.09%, and an ensemble majority voting mechanism was applied to consolidate predictions across all classifiers. This framework detects ESP operating conditions in ten minute intervals over three days and generates prescriptive maintenance recommendations for each predicted failure mode. The key novelty of this study lies in integrating the comparative evaluation and selection of the most suitable deep learning architecture for each sensor, physics based liquid flow rate estimation, and majority voting failure classification into a unified decision support dashboard, advancing beyond previous single model approaches by providing engineers with early warnings and actionable maintenance guidance up to three days before potential ESP failure.