A Thermodynamic based and Data Driven Hybrid Network for Gas Turbine Modeling

TL;DR

This paper introduces a hybrid gas turbine modeling network combining thermodynamic principles with data-driven techniques, improving accuracy over traditional models and pure data-driven methods by accounting for component degeneration and individual differences.

Contribution

It presents a novel hybrid network that integrates physical equations with data-driven component characterization for more accurate gas turbine modeling.

Findings

Achieves about 7% max T6 relative error, outperforming thermodynamic and data-driven models.

Reduces modeling error by 5-8% compared to existing methods.

Utilizes a large dataset of 26,970 flight data points for training and testing.

Abstract

The on-wing engine performance is difficult to track for thermodynamic models because of its inaccurate component maps, and also difficult for data driven methods for their over-fitting to measurement errors. So, we propose a thermodynamic based and data driven hybrid network for gas turbine modeling. Different from thermodynamic models, our network reconstructs the component characteristics in a data-driven way to take component degeneration and individual difference into consideration. Moreover, different from data driven methods, in the training phase, physical based equations and the analytical mathematical description are used to ensure that the optimization converges to the gas turbine's dynamics. A huge number of relaxed quasi steady state flight data to 26970 is used to train and test our hybrid network. The result shows that the accuracy of our hybrid network can reach about 7% measured by max T6 relative error, 5% better than map fitting based thermodynamic model and 8% better than pure data driven method with similar model volume.