Lower limb exoskeleton technology can assist in enhancing human strength and has a wide range of applications. However, exoskeleton motors exhibit dynamic nonlinearity due to electromagnetic inertia and mechanical loads, leading to overshoot that causes mismatches in human-machine collaborative motion responses and increases the risk of biomechanical injury. To address the inertial overshoot issue in exoskeleton motors during the leg-lifting phase of human walking, a two-stage collaborative overshoot prediction and optimization strategy is proposed, integrating inverse system model prediction with forward model optimization. By constructing a reverse system model based on a CNN-LSTM-Attention architecture, the system receives real-time motor target outputs, effectively captures multivariate time-series patterns, and rapidly generates the initial input commands for the exoskeleton motor. A forward optimization model based on a pyramid feature fusion-CNN-bi-LSTM-Transformer (Pyramid-CLT) architecture is constructed. The model employs a gating mechanism together with pyramid-shaped fully connected layers to achieve multi-scale feature integration. The mean squared error between the predicted output and the target output serves as the objective function. To further improve the prediction accuracy, a particle swarm optimization (PSO) algorithm is applied to iteratively refine samples with high mean square error, enabling the generation of precise motor control commands for overshoot compensation. Experimental that the prediction optimization strategy can precisely generate exoskeleton motor input commands based on human movement trajectories, achieving a model correlation coefficient (R2) of 0.985, root mean square error (RMSE) of 0.537, and mean absolute error (MAE) of 0.442; Compared with single-model algorithms and other prediction algorithms, the proposed method achieves real-time dynamic prediction and correction of overshoot, enabling motor output to closely align with human movement trajectories, thereby effectively enhancing human-machine synergy and providing a new method for precise control of lower-limb exoskeletons.