Most existing supervised power quality disturbances (PQD) classification methods cannot be applied practically due to the difficulty of obtaining sufficient labeled field data. In this paper, a unsupervised domain adaptation (UDA) method is introduced to tackle this problem. Moreover, based on the characteristics of the PQD classification problem, a new deep transfer model called MCDA-pqd (Metric based deep Convolutional Domain Adaptative neural network in power quality disturbances classification) is proposed. By comparing with previous works, two main contributions can be drawn. Firstly, in addition to the signal-to-noise ratio term, we propose to use different noise distributions to simulate possible realworld scenarios. Secondly, the proposed model can utilize domain adaptation technic to complete unsupervised training process on any possible target domain (simulated realword scenario). Experimental results demonstrate the effectiveness and reliability of both the dataset generation method and the proposed model.