A novel gas turbine performance prediction model incorporating the residual connection and feature engineering methods

With the growing share of renewables, gas turbines (GTs) are increasingly operating under part-load conditions. Although existing models can accurately predict performance at both design-and off-design conditions, they often overlook part-load conditions. Besides, during full-load modeling, the mismatch between ideal operational data (with fully opened IGV) and actual operational data (with nearly fully opened IGV) may introduce additional modeling errors. Moreover, existing part-load performance prediction methods are typically implemented based on specific assumptions that only work under certain preconditions. To tackle these problems, a novel GT full-and part-load performance prediction model is proposed. First, the baseline model is established, serving as a solid foundation for subsequent correcting. Then, the IGV correction coefficients are calculated to quantify the baseline model's prediction bias caused by varying IGV positions. Subsequently, a multi-layer perceptron (MLP) enhanced with residual connection and feature engineering techniques is developed to accurately predict the IGV correction coefficients. Due to the small amount of input variables and training samples, feature generation and feature analysis methods are adopted to improve the prediction accuracy. Moreover, the residual connection is incorporated to partially mitigate the gradient vanishing phenomenon. Finally, the trained model is integrated into the physics model to calculate the corrected performance. The actual data adopted from various operating conditions are utilized to verify its efficiency and accuracy. The average prediction error is less than 1.59 %, indicating that the proposed method can accurately predict GT performance under both full- and part-load conditions, thereby providing a solid foundation for fault diagnosis.