Artificial intelligence for predictive maintenance and safety optimization in green hydrogen infrastructure
Abstract
Green hydrogen is one of the central technologies in the ongoing energy transition, but scaling up hydrogen infrastructure introduces significant challenges in equipment reliability, operational safety, and maintenance planning. This paper presents an Artificial Intelligence (AI) framework for predictive maintenance and safety optimization across the production, storage, and distribution stages of green hydrogen systems. The framework processes multi-sensor data, temperature, pressure, vibration, flow rate, and energy consumption, through an ensemble anomaly detector combining an Isolation Forest and an LSTM Autoencoder, a Bidirectional LSTM model that estimates Remaining Useful Life (RUL) using a monotonicity-penalized loss function, and an XGBoost classifier that assigns real-time risk across three levels. SHAP-based explainability is integrated throughout to support operator interpretation and regulatory audit. The framework is evaluated through a physics-informed simulation of a 1 MW PEM hydrogen facility under twelve fault scenarios, achieving an anomaly detection F1-score of 0.924, an RUL estimation MAE of 6.4 hours (R 2 = 0.877), and a safety classification accuracy of 96.1% with 97.1% recall on critical states. These results indicate that AI can serve as a practical enabler of safer, more cost-effective hydrogen infrastructure.