<p>This study investigates the modeling and prediction of tracking errors in CNC machine tool feed systems by proposing an LSTM-Transformer-based neural network framework. The proposed model leverages the strengths of LSTM in capturing local temporal dependencies, while the multi-head attention mechanism of the Transformer is employed to model the global influence of motion state variations on tracking error. This combination yields improved prediction accuracy and generalization capability. To further enhance model performance, a featured trajectory construction method is introduced to adequately excite the dynamic characteristics of the servo system. By systematically enriching the motion state distribution of the training dataset, the prediction model's capacity to capture nonlinear dynamic behaviors is significantly enhanced. Comparative experiments with conventional LSTM and GRU networks demonstrate that the proposed framework, combined with the featured trajectory design strategy, effectively reduces tracking error prediction errors and maintains strong generalization capability under complex trajectories.</p>