Abstract
This research paper explores the use of machine learning (ML) to create a surrogate model for steady state and transient metal temperature prediction in industrial gas turbines. This work leverage measured performance parameters from a development test engine to build a surrogate model using ML-based approach. The resulting surrogate model significantly reduces the effort required to derive empirical correlations to predict the tie-bolt temperature of the industrial gas turbine which operates in various environmental conditions and power requirements.
The development engine test involves approximately 10,000 measured signals which includes major performance parameters like speed, fuel flow rate, power, IGV angle, station temperature and pressure along with others testing requirements from different disciplines. Hence various statistical techniques such as heatmap, correlation map, auto-lag, and principal component analysis were employed to identify the key influencing performance parameters to train and test the surrogate model.
Several time-series models were studied for this research, starting from simple regression models such as ARIMA, ARIMAX, and VAR to more complex deep learning models like LSTM. The benefits and drawbacks of different machine learning models were assessed, and their relative performance was compared in terms of temperature prediction accuracy.
The LSTM model was found to be the most accurate and better fit for this kind of sequential problem (time series), with temperature prediction accuracy within 2% for steady-state and 4% for transient conditions. This surrogate model forecasting approach can potentially estimate the real-time remaining useful life of critical parts in the fleet engine.
Furthermore, the paper proposes that the machine learning framework developed for predicting metal temperature in industrial gas turbines can be extended to forecast temperatures for other regions using fleet measured data. This can lead to more accurate critical part lifing calculations based on real-time fleet measured data, which can be used to optimize part replacement, inventory management, servicing, and overhaul scheduling. In essence, this research presents a comprehensive analysis of the benefits and drawbacks of different machine learning models and demonstrates the effectiveness of using an LSTM model for metal temperature prediction in industrial gas turbines.
Overall, this research presents a significant step towards improving the reliability of industrial gas turbines through the use of machine learning approach. The proposed framework has the potential to revolutionize the way in which industrial gas turbines are maintained and serviced, leading to significant cost savings and provide valuable insights for the engineers and researchers working in this field.
