System Identification Methodology of a Gas Turbine Based on Artificial Recurrent Neural Networks

Area of Science:

• Engineering
• Artificial Intelligence
• Control Systems

Background:

• Gas turbines (GT) exhibit complex nonlinear dynamics, challenging traditional physics-based modeling.
• Artificial intelligence (AI) techniques, particularly NARX (Nonlinear Autoregressive Network with Exogenous Inputs) models, show promise for GT identification.
• Existing methods lack a systematic approach for developing accurate and robust NARX models for GTs.

Purpose of the Study:

• To propose a systematic, nine-step method for designing robust NARX models for gas turbine identification.
• To address the need for a structured approach in generating accurate AI-based GT identification models.
• To improve the accuracy and reliability of identifying nonlinear dynamic systems like gas turbines.

Main Methods:

• Development of a nine-step systematic methodology for NARX model design.
• Application of the proposed method to a real-time dataset from a SIEMENS TG, model SGT6-5000F (215 MW).
• Utilizing 2305 real-time series data records for model validation and performance evaluation.

Main Results:

• Successful application of the systematic method to a 215 MW SIEMENS gas turbine.
• Achieved a highly accurate NARX model with Mean Squared Error (MSE) of 1.945 × 10^-5.
• Obtained Root Mean Squared Error (RMSE) of 0.4411% and Mean Absolute Percentage Error (MAPE) of 0.0643%, indicating excellent model performance.

Conclusions:

• The proposed systematic method effectively generates robust NARX models for gas turbine identification.
• The validated model demonstrates high accuracy and reliability in capturing the nonlinear dynamics of the gas turbine.
• This structured approach offers a significant advancement for AI-driven identification of complex industrial systems.