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

Abstract

Prediction of successful ignition is a challenging task that requires knowledge of the local thermochemical state in highly turbulent flow conditions. For typical industrial gas turbine (GTs) conditions operated on active service, these pieces of information are not directly available despite the high impact of ignition on technological and economic performance. The success of ignition mainly depends on the experiences made with the specific GT during long-term operation. Hence, there is a need for reliable prediction models for successful ignition that take the current conditions of the engine into account. We demonstrate the performance of supervised machine learning (ML) models to predict both successful and failed ignitions of GTs based on “real world” fleet data from the SGTx-8000H frame. This study compares the classification ability of seven widely used algorithms, for which we initially select 22 engineering-relevant parameters based on sensor data. We employ correlation and elimination techniques to reduce the parameter space drastically so that with only two input parameters we achieve a high (87%) prediction accuracy. Finally, we highlight the generalizability of the best performing ML model by application to unseen data of a different GT and report, for another case, on a recommissioned GT that is assisted by the ML model and consistently ignites the engine over an extended period of time.