Anomaly detection in liquid air energy storage systems using autoencoders: a machine learning approach for operational reliability

• Deep learning application for anomaly detection in Liquid Air Energy Storage. • Autoencoder trained on experimental data achieved coefficient of determination 0.91. • It detected both sharp faults and slow-developing anomalies in sensor data. • Cumulative error flagged nitrogen leakage, early, in a fourth unseen experiment. Liquid Air Energy Storage systems play an increasingly critical role in grid flexibility and renewable integration, yet their real time monitoring and anomaly detection remain underexplored. Traditional monitoring methods often struggle to capture subtle or gradually developing anomalies in such complex thermodynamic environments. This study introduces a systematic, unsupervised deep learning framework for anomaly detection in a pilot-scale liquid air energy storage facility, and presents new operational data from real system measurements. In this context, a multivariate autoencoder was trained on normal operating data, with anomalies detected via reconstruction error. 648 autoencoder configurations were evaluated through grid search, and the optimal model was selected using the technique for order of preference by similarity to ideal solution, balancing root mean square error (0.172) and coefficient of determination (0.916). Statistical thresholding based on a 3σ rule enabled detection of point anomalies, including sensor spikes, while feature-wise error analysis helped localize likely sources of fault such as base-layer thermocouples and charge-line flow sensors. To evaluate detection of complex faults, a fourth run with a known nitrogen leak was analyzed. A sustained increase in cumulative reconstruction error revealed a gradual deviation from expected behavior. Further analysis showed that the highest error contributions came from sensors located along the nitrogen delivery line, matching the leak location identified experimentally. These results confirm the model’s ability to detect both abrupt and evolving faults, supporting early diagnostics, and offering a generalizable framework for real-time liquid air energy storage system monitoring and broader thermodynamic applications.