https://www.mdu.se/

Global energy transition efforts towards decarbonization requires significant advances within the energy sector. In this regard, hydrogen is envisioned as a long-term alternative fuel for gas turbines. Accordingly, the gas turbine industry has expedited their efforts in developing 100% hydrogen compliant burners and associated auxiliary components for retrofitting the existing gas turbines. The utilization of hydrogen in gas turbines has some underlying challenges such as corrosion mainly originating from increased steam content in the hot gas path. In addition to corrosion, the gas turbine compressor is vulnerable to fouling which is the most commonly occurring fault in gas turbine operating over certain time window. Both faults are susceptible to performance and health degradation. To avoid expensive asset loss caused by unexpected downtimes and shutdowns, timely maintenance decision making is required. Therefore, simple, accurate and computationally efficient fault detection and diagnostics models become crucial for timely assessment of health status of the gas turbines. The present study encompassed development of a physics-based performance model of a 100-kWe micro gas turbine running on 100% hydrogen fuel. The model is validated with experimental data acquired from test campaigns at the University of Stavanger. Data synthesized from experimentally validated performance model are utilized further for training machine learning algorithms. To identify an accurate algorithm, various algorithms such as support vector machine, decision tree, random forest algorithm, k-nearest neighbors, and artificial neural network were tested. The findings from fault diagnostics process (classification) revealed that ANN outperformed its counterpart algorithm by giving accuracy of 94.55%. Similarly, ANN also showed higher accuracy in performance degradation estimation process (regression) by showing the MSE of training loss as low as ~0.14. The comparative analysis of all the chosen algorithms in the present study revealed ANN as the most accurate algorithm for fault diagnostics of hydrogen fueled gas turbines. However, there is need to further implement the ensemble machine learning models or deep learning model to explore and expedite the real time fault diagnostic accuracy to avoid false alarms and missed detections in context of hydrogen fuel.