Application of Transfer Learning for the Prediction of Gas Turbine Trip

One of the most disrupting events that reduces gas turbine (GT) availability and also significantly increases maintenance costs is trip. Given its nature, data-driven methodologies, which are becoming more and more widespread for diagnostic and prognostic purposes, are best suited for its prediction. However, one of the major limitations of data-driven models is their partial extrapolability. To overcome such limitation, this paper exploits a transfer learning approach for sharing knowledge about GT trip among different GT fleets. The transfer learning approach applied in this paper employs a Long Short-Term Memory (LSTM) neural network that is pre-trained by using the data of a given fleet and then finetuned on the data of another fleet running in a different site. The case study is represented by a collection of field measurements referring to real-world trips taken from two GT fleets during several years of operation. The field data cover multiple ambient and operating conditions, thus providing a challenging test for the transfer learning approach. The analyses performed in this paper demonstrate that the loss of accuracy between the source domain model (trained and tested on trips that belong to the same domain) and the predictions made on the target domain can be less than 7%. Thus, in general, the results prove that transfer learning is possible and allows comparable (though necessarily lower) accuracy to that of the source domain model.