Abstract: To address the challenges of strong non-stationarity, insufficient feature representation, and difficulty in mining deep discriminative information in ultrasonic signals from partial discharge (PD) in dry-type transformers, a novel PD pattern recognition method based on a dual-branch CNN-TCN-Attention model is proposed. The convolutional neural network (CNN) branch employs convolutional kernels of different sizes to extract local multi-scale features from the signal, while the temporal convolutional network (TCN) branch utilizes causal convolutions and residual connections to capture global temporal features. An attention mechanism adaptively integrates the features from both branches. Experimental results show that the proposed method achieves a final recognition accuracy of 96.52%, outperforming single-structure models and other commonly used combination models.