A Machine Learning Predictive Model for the Charging Capacity of Stationary Lithium-Ion Batteries Connected to Renewable Energy Sources at Remote Oil and Gas Fields

Abstract As the oil and gas industry strives to meet its energy transition goals while ensuring reliable and uninterruptable electricity generation to meet its operational needs, where grid access is often limited, Renewable Energy Sources (RES), particularly wind and solar power, have become increasingly vital. Stationary lithium-ion batteries (LIBs) play a pivotal role in the electricity needs at oil and gas sites, providing backup power for critical systems, enabling microgrid stability, and reducing reliance on fossil-fuel-powered generators. This paper addresses the challenge of the limited useful life of LIBs—a key energy storage solution for integrating renewable energy sources for the oil and gas industry applications—by leveraging advanced AI and machine learning techniques. A Machine Learning predictive model is developed using the Long Short-Term Memory (LSTM) algorithm to enhance the accuracy and reliability of battery management systems. The model is trained on an extensive dataset of grid-scale LIBs which includes measurements of Voltage, Current, Cell Temperature, and Environment Temperature. Rigorous data pre-processing steps were undertaken to ensure data integrity and enhance model performance. A sequential time-based splitting technique preserved the temporal dependencies crucial for time-series analysis. Feature engineering identified Voltage and Current as critical predictors, supplemented by timestamps and temperature features. Data augmentation was employed to generate synthetic time series measurements of capacity, further enhancing accuracy, model stability, and generalization. The LSTM model effectively predicts the annual degradation in the charging capacity of grid-scale stationary Lithium-ion batteries subjected to the sporadic currents generated by RES, achieving a Root Mean Square Error of 0.0126 and a Mean Absolute Error of 0.0107, proving its adequacy for handling the complex, non-linear dynamics of time-series data inherent in battery performance metrics. This study addresses a significant gap in research, focusing on predicting the charging capacity of LIBs in offshore and remote locations where the batteries experience sporadic current, as opposed to the constant current and voltage conditions typically studied. It underscores the potential for machine learning-driven advancements in energy storage technology, for enhanced efficiency and consequently more sustainable oil and gas operations.