To address the challenges of high computational cost and long simulation cycles inherent in traditional numerical simulation methods for analyzing labyrinth seal leakage characteristics. A two-stage artificial neural network-based method was proposed for predicting leakage in labyrinth seals. Using MLP (multilayer perceptron) as the core, a main network prediction model was constructed to achieve preliminary regression prediction of labyrinth seal leakage. To further improve accuracy, a second-stage residual prediction network was introduced to correct the prediction errors of the main network. Leakage data under different labyrinth seal structural parameters were obtained through computational fluid dynamics methods, and these data were used as samples to complete both stages of training and testing. The results Show as follows. The main network prediction model achieved a coefficient of determination R² of 0.985 5 on the test set, demonstrating excellent predictive performance. The overall performance of the prediction model was improved after incorporating the error compensation network, reducing the root mean square error and mean absolute error of the main network by 30.56% and 31.98%, respectively, and increasing R² to 0.992 6. The proposed method maintains high prediction accuracy even with only 800 samples, effectively mitigating the fluctuation in errors that typically occurs with MLP as sample size increases. This study provides a reliable and efficient predictive tool for the rapid design and performance optimization of labyrinth seal structures.