Research on Performance Degradation Prediction Method of Heavy-Duty Gas Turbine Based on Data-Physics Fusion Under Uncertainty

Abstract Heavy-duty gas turbines (HDGT) have been widely used for power generation, because of their unique characteristics of high efficiency, low-emission, and flexible in rapid startup and shutdown. However, their main components such as compressor, combustion chamber, and turbine are usually operating under harsh environments such as high temperature, high pressure and high corrosion for a long time. The resulting fatigue, leakage, corrosion, or fouling in various components often lead to their performance degradation over the operating time, lowering the productivity of the gas turbine. Performance prediction becomes a powerful approach to monitor the degradation degree so that maintenance recommendation is provided to effectively ensure the efficiency and productivity of the machine. This paper presents a new data-physics intelligent fusion approach for performance degradation prediction of HDGT under uncertainty. A thermodynamic heat balance model is first established to simulate component characteristics. The actual operating data of a real-world gas turbine are then employed for physics-based model assessment and parameters correction. Next, The deviation between expected performance data and actual heat balance value under the current operating point is utilized to assess the performance degradation. Finally, in order to consider the uncertainties in both the sensor signals and the prediction model, an improved Bayesian-LSTM (long short-term memory) model is proposed to predict multiple performance indicators by seamlessly integrating Bayesian variational inference with LSTM model. This method models each of the weight and bias parameters as a continuous distribution, rather than a constant. About half a year of sensor data collected from a 180 MW real-world gas turbine is utilized in a comparison study with the traditional LSTM models to demonstrate the superiority of the proposed method in predicting the performance trend of gas turbines. The proposed methodology can not only produce the predicted curve, but also obtain the fluctuation range of the prediction output to effectively consider the uncertainty in both the data and model, thus improving the robustness of the proposed model.