Accurate recognition of human motion intention remains a critical challenge in developing lower limb exoskeleton robots for effective human-robot collaboration. Surface electromyography signals (sEMG), which provide non-invasive measurement of neuromuscular activity, have emerged as crucial bioelectric signal for movement intention recognition due to their ability to reflect muscle activation patterns prior to physical movement. This paper proposes a novel TCN-Transformer-Cross-Attention model based on sEMG signals for hip joint angle prediction in multitasking scenarios (level walking, stair ascent, stair descent, and ramp ascent). The proposed framework employs a temporal convolutional network (TCN) to extract localized temporal features, complemented by a Transformer module to capture global temporal dependencies. A cross-attention mechanism is innovatively integrated to enable synergistic fusion of local and global feature representations, thereby enhancing the model’s capability for comprehensive sEMG signal interpretation. Comparative evaluations against benchmark models (CNN-LSTM, BiLSTM, TCN-BiLSTM, and TCN-Transformer) demonstrate significant improvements in prediction accuracy and stability across multitaskings. The incorporation of cross-attention mechanism yields remarkable performance enhancements, achieving relative improvements of 18.87% in R2 score while reducing RMSE and MAE by 48.61% and 53.91%, respectively, compared to the baseline TCN-Transformer architecture. The results validate the effectiveness of the proposed model in lower limb hip joint angle prediction, and provide a new way for multitasking motion control of lower limb exoskeleton robots.