Evaluation of Data-Driven Classifiers for an Ignition Forecast of Large Gas Turbines

This paper demonstrates several supervised learning algorithms for ignition prediction of large stationary gas turbines (GTs), based on existing “real world” operational data. We obtain the data set from the mature SGTx-8000Hframe. In this work, we detect both successful and failed ignitions in advance based on various environmental changes. We test seven classification algorithms for their classification ability in the context of engine ignition: Random Forest (RF), Balanced Random Forest (BRF), Gradient Boost (GB), Gaussian Process Classifier (GPC), Multiple Logistic Regression (MLR), AdaBoost and Support Vector Machine (SVM). After a detailed pre-processing including data selection, cleaning and feature engineering, the modelling results are promising: The small fraction of 2% failed ignitions, in the presented data set, is correctly classified with high accuracy. In addition, we reduced the parameter space from 22 to 11 (for the fleet’s 60 Hz cohort) and to 14 (for the fleet’s 50 Hz cohort). In a further step, we reduced to 5 and finally to 2 features while maintaining the model’s accuracy, which boils down the model complexity by the application of 2 robust selection methods: The Pearson’s paired comparing of feature correlation and the Recursive Feature Elimination (RFE). Starting with seven applied classification algorithms on the whole GT fleet’s data set (both 50 Hz and 60 Hz cohorts) we reduced the data set length to a single engine and applied the most promising classifiers accordingly. To check the “real world” applicability and thus the classifier’s performance on unseen data, we apply two different test scenarios: Firstly, we transfer the prediction map for successful ignitions of a 50 Hz GT with good accuracy to unseen data of a different 50 Hz GT. Secondly, we assist the commissioning engineers for the recommissioning after an overhaul of a 60 Hz GT. The algorithm’s recommendation led to constant successful ignitions on this specific test site covering various boundary conditions over the observation period of 2 years.