Gas turbine performance prediction via machine learning

• Published in: Energy (Oxford) Vol. 192; p. 116627
• Main Authors: Liu, Zuming, Karimi, Iftekhar A.
• Format: Journal Article
• Language: English
• Published: Oxford Elsevier Ltd 01.02.2020; Elsevier BV
• Subjects: air; Air compressors; Artificial neural networks; Computer simulation; Correction curves; data collection; Electric power generation; energy; Fault diagnosis; fuels; Gas turbine; Gas turbines; Learning algorithms; Learning theory; Machine learning; Model accuracy; Neural networks; Performance degradation; Performance prediction; power generation; Power plants; prediction; Pressure ratio; Simulation; Surrogate models; temperature; Turbines; Performance prediction; Gas turbine; Surrogate models; Correction curves; Simulation; Machine learning
• ISSN: 0360-5442; 1873-6785
• DOI: 10.1016/j.energy.2019.116627

This paper develops a machine learning-based method to predict gas turbine performance for power generation. Two surrogate models based on high dimensional model representation (HDMR) and artificial neural network (ANN) are developed from real operational data to predict the operating characteristics of air compressor and turbine. Both models capture the operating characteristics well with average errors of less than 1.0%. Moreover, four more holistic models are developed to capture gas turbine part-load and full-load performance. The models for air compressor and turbine are then embedded into a gas turbine simulation program, and all surrogate models are validated using a separate data set. It is shown that the power output, pressure ratio, fuel flow, and turbine exhaust temperature from these models match their measured values well with average and maximum errors of less than 2.0% and 4.3%, respectively. Since holistic ANN models have lower complexity and higher accuracy, the ANN model for predicting full-load performance is used to construct gas turbine performance correction curves. The correction curves along with the ANN model for predicting part-load performance offer an excellent basis for continuous health monitoring and fault diagnosis. The proposed methodology is applicable to any gas turbines and can help power plants to study and quantify performance degradation over time. •A data-driven method is developed for predicting gas turbine performance.•The method builds and uses surrogate models via machine learning from operational data.•Surrogate models predict gas turbine part-load and full-load performance.•Gas turbine power generation capacity under any operating condition is obtained.•Gas turbine performance correction curves are successfully reproduced.