Lightning type recognition plays a crucial role in the fields of lightning detection and meteorological disaster prevention and mitigation. Currently, lightning type recognition faces challenges such as data diversity and complex waveform features, which limit the accuracy and generalization ability of recognition algorithms. A lightning waveform multi-classification algorithm, multi-scale residuals and self-attention (MSRES-SA), based on a multi-scale residual network and attention mechanism, was proposed to enhance the accuracy of lightning waveform recognition. Initially, a multi-scale residual feature extraction module was constructed to extract information from lightning waveforms at different scales in the time dimension, and residual concatenation was employed to enhance the model's representational capability. Subsequently, the attention mechanism was utilized to dynamically weight important features, capturing long-range dependencies within the waveform sequence. Experimental results demonstrate that the MSRES-SA algorithm achieves an average recognition accuracy of 99.35%, outperforming the baseline model in the recognition of multiple lightning waveform categories. The effectiveness of the multi-scale residual module and the attention module in the lightning waveform recognition task is further validated through ablation experiments.