Session: 04-23 Ignition I

Paper Number: 128055

128055 - Evaluation of Data-Driven Classifiers for an Ignition Forecast of Large Gas Turbines 

Renewable energy integration requires flexible and reliable power plants for backup. Preventing
unsuccessful ignitions is crucial to reduce downtime, stabilize the electrical grid, save fuel gas, hot gas path
parts' lifetime, and thus the operation costs of the power plant. Due to the high degree of automatization,
large gas turbines are equipped with various kinds and amounts of sensors. The sensor data is used to protect
and control specific gas turbine's open- and closed-loop to fulfill the abovementioned demands. Daily, large
amounts of data are generated, evaluated, and used by gas turbine operators or OEMs to optimize the
operation of their gas turbines and other assets. In a continuously changing and cost-driven environment, it is
critical to fully utilize sensor data, quickly interpret it, and optimize operations. Experienced interdisciplinary
experts are vital for interpreting these data. Since these experts' capacity is insufficient for monitoring and
optimizing the whole OEM's gas turbine fleet in real-time, combining domain knowledge and modern data
analysis (e.g., machine learning) is a promising approach. This paper demonstrates several supervised
learning algorithms based on existing operational data for ignition prediction of large stationary gas turbines.
Unsuccessful ignitions are detected in advance based on different environmental changes, significant engine
conditions reflecting the fleet experience, and relevant parameter settings so that preventive measures can be
taken before an ignition fails. We tested five classification algorithms for their detection: multiple logistic
regression, random forest, balanced random forest, gradient boosting, and support vector machine. After a
detailed pre-processing of data selection, cleaning, and feature engineering, the modeling results are
promising: The small fraction of 2 % failed ignitions in the present data can be correctly classified with
adequately high accuracy. In addition, we reduced the parameter space from twelve to three features while
maintaining accuracy, thereby reducing the model complexity. The best-performing algorithm was applied to
give setting recommendations to the commissioning engineers for the recommissioning after an overhaul of a
large stationary gas turbine in commercial operation. The algorithm's recommendation led to repeated
successful ignitions on a specific test site.

Presenting Author: Florian Lang Siemens Energy Global GmbH & Co. KG

Presenting Author Biography: 2009: Diploma in Industrial Engineering (HS Pforzheim, Germany)
Diploma thesis in application of advanced sensor technologies in an industrial environment
2009- 2014: Instrumentation & Controls (I&C) Service engineer and I&C lead engineer for
large gas turbines, generators, and steam turbines (Siemens AG, Karlsruhe, Germany)

2014: Master of Science in Electrical engineering and Information Technology (PRT, Institute of
Control systems engineering, Faculty of Mathematics and Computer Science, University of
Hagen, Germany)
Master thesis: Vibration diagnostics approach in gas turbine dynamics based on FPGA

2014-2019: R&D Combustion application engineer for large gas turbines
(Siemens AG, Mülheim a.d.R., Germany)

2019- present: Program manager for large gas turbines (Siemens Energy Global GmbH & Co.
KG, Berlin, Germany)
2020- present: External researcher at Prof. Dr. Abdulla Ghani's Team (DFM, TU Berlin)

Since 2023: Program management Single Point of contact for SGTx-8000H gas turbines in the region of Asia Pacific (Siemens Energy Global GmbH & Co. KG, Berlin, Germany)

Authors:

Florian Lang Siemens Energy Global GmbH & Co. KG
Maximilian Savtschenko Siemens Energy Global GmbH & Co. KG
Vikas Yadav Technische Universität Berlin, Institut für Strömungsmechanik und Technische Akustik, Data Analysis and Modeling of Turbulent Flows
Abdulla Ghani Technische Universität Berlin, Institut für Strömungsmechanik und Technische Akustik, Data Analysis and Modeling of Turbulent Flows