Real-time surface electromyography (sEMG) decomposition has emerged as a research hotspot, with applications spanning prosthetic control, medical rehabilitation, and beyond. However, the accuracy of decomposition algorithms is often compromised by electrode shifts. Here, we propose a deep learning-based (DL-based) decomposition method specifically for electrode shift interference. First, an electrode-shift-oriented data augmentation strategy is designed to enrich training data diversity for sEMG-to-motor unit spike train mappings. Subsequently, an spatiotemporal DL architecture with robustness to feature shift is constructed to learn mapping knowledge embedded in training data. Simulation evaluation results demonstrate that the proposed method outperforms the traditional DL-based method under 10 mm multidirectional electrode shifts, achieving an F \(_{1}\) -score of \(0.917\pm 0.066\) compared to \(0.207\pm 0.306\) . Moreover, the proposed method requires only \(0.0393\pm 0.0017\) seconds per instance for decomposition, meeting real-time constraints. These outcomes enable reliable acquisition of neural drive information under electrode shift interference, thereby advancing the application of neural control interfaces.