This paper introduces a novel approach to predicting joint torque in a lower limb exoskeleton rehabilitation robot. Initially, a simplified neuromusculoskeletal (NMS) model is constructed to estimate joint torque. This model integrates various human parameters, including surface Electromyography (sEMG) signals, joint angle data, and tendon length. A simulated annealing algorithm (SA) is employed to optimize the accuracy of the NMS model and optimize its parameters. Moreover, to further improve prediction accuracy, a hybrid model comprising a convolutional neural network (CNN), a long short-term memory neural network (LSTM), and an attention mechanism (Attention) is developed (CNN-LSTM-Attention). Leveraging the NMS model as a teacher model, the generated data is utilized as additional training samples for the student network CNN-LSTM-Attention. This process aims to augment the generalization capability and accuracy of the CNN-LSTM-Attention model.