The current study is based on multiple machine learning algorithms to predict the normal behavior of operational parameters including power generated and blade path temperature spread. The predictions can be used to identify anomalies and probable failures in the gas turbine performance. The data used in the study is taken from multiple heavy-duty gas turbine units of combined cycled utility power plants which are known to contain operational failures. The predictors include operational parameters such as fuel flow, various thermodynamic variables, etc.
In the first step, we cluster the observations into different working modes, because of the heterogeneous behavior of the gas turbine parameters under various modes. Then we consider predicting the operational parameters under each mode respectively, via algorithms including random forest, generalized additive model, and neural networks. The models are trained and parameters are selected based on the overall prediction performance on the validation set.
The comparative advantage based on prediction accuracy and applicability of the algorithms is discussed for real-time use and post processing. The advantage of our method is that they achieve high predictive power and provide insight into the behavior of specific gas turbine variables, e.g.- turbine blade path temperature spread, which are not explicitly known to have any correlation with other thermodynamic variables.