The increasing integration of non-linear devices such as new energy generation and a large number of electric vehicle charging stations into the power grid has led to increasingly prominent power quality problems. However, current methods face challenges in the classification of power quality disturbances, with complex steps and low accuracy when considering disturbance signals. To address these issues, this paper proposes the TCN-LSTM model, which combines a temporal convolutional network (TCN) with long short-term memory (LSTM). The TCN network excels in capturing local features of time series, while the LSTM is proficient in digging long-term dependencies within the time series. The fusion of both enables the model to effectively capture both local characteristics and global relationships of the signals. To validate the model’s performance, a classification test is conducted on 14 types of power quality disturbance signals with varying signal-to-noise ratios. Finally, the results demonstrate that the TCN-LSTM model exhibits strong noise resistance. In comparison to existing deep network models, the model proposed in this paper achieves higher classification accuracy.