Prediction of Aero-Engine Exhaust Gas Temperature Using iTransformer

Exhaust gas temperature (EGT) is one of the key parameters for evaluating the thermal state of an aero-engine during flight operations. Precise EGT predictions are important in both managing engine health and optimizing performance. However, the nonlinear interactions among multiple operating parameters pose challenges for conventional prediction methods. An approach to EGT prediction presented in this paper is based on the iTransformer architecture. Unlike conventional time-series models that process temporal observations as tokens, this algorithm works with every single variable like a unique token and encodes the entire historical sequence of each variable. This design enables the model to capture inter-variable dependencies directly, thus increasing its power to capture multivariate time series. Two different datasets were used to validate the proposed approach and compared to various commonly used prediction models. The results indicate that iTransformer gives smaller prediction errors when the same conditions are met with MAE values of 0.5525°R and 0.4539K both on the two datasets. Despite the overall satisfactory performance, some prediction mismatches can be observed under extreme fast changing conditions of the engine operation. These results demonstrate that the proposed method can effectively characterize the complex interactions among engine variables and has potential for practical EGT prediction applications.