Gas turbine engine system is characterized by tightly coupled gas path, mechanical, and hydraulic subsystems, exhibiting strong nonlinearity and high dynamic behavior. It was found that the power turbine speed showed parameter fluctuations, making the accurate identification of disturbance sources critical for ensuring system safety. When inferring the causes of fluctuations directly from component monitoring data, multiple parameters may exhibit high causal relevance. However, in systems with strong variable correlations and complex feedback couplings, the true source of oscillation is often not the parameter with the highest causal score, making it difficult for conventional methods to locate the disturbance. This study presents a hybrid method that combines a mechanism-based simulation model with transfer entropy-enhanced machine learning to locate the sources of such oscillations. A dynamic model capturing multi-time-scale interactions is used to simulate the system's response under perturbations applied to different subsystems, generating a labeled dataset for supervised learning. Causal relationships among subsystems are quantified using transfer entropy, which is then encoded into a graph-based prior to guide learning. An LSTM-MLP classifier augmented with an attention mechanism is developed to integrate transfer entropy features with time-series data, significantly enhancing source localization performance. Comparison results show that the proposed method improves the identification accuracy from 94.11% to 99.01% on the test dataset. This approach offers a promising solution for disturbance source identification in complex gas turbine systems for parameter oscillation.