Traditional power quality disturbance (PQD) classification methods often rely on a limited set of disturbance types for training, making it challenging to accurately identify previously unseen complex and multiple disturbance types. To address this issue, this paper proposes a novel PQD classification method based on time-frequency domain fusion and confidence enhancement model. Firstly, the PQD signal is transformed using the fast Fourier transform to obtain its spectral information. Then, a temporal convolutional network and a convolutional neural network are employed to extract features from the time and frequency domains, respectively. The extracted features are fused to enhance the overall feature representation. Within a multi-label learning framework, class labels are introduced to differentiate between single and multiple disturbance types, and confidence scores are predicted to determine the presence of each disturbance label. Finally, to further improve the ability of model to identify unseen multiple disturbance types, a label enhancement factor is designed to optimize the confidence is tribution for multiple disturbances without affecting the recognition performance of known PQD types. Simulation results show that the proposed method achieves an dentification accuracy of over 96. 75% for multiple disturbance types not included in the training set, even when trained only on single and dual disturbance samples. In real-world tests, the method maintains a recognition rate above 91.67% for unknown disturbance types, demonstrating strong generalization capabilities. The proposed method offers high application value in real-world scenarios where power grid operating conditions are variable and disturbance patterns are complex and superimposed.