The traditional WiFi fingerprint-based indoor positioning methods show poor performance due to the fluctuation of received signal strength (RSS) caused by environmental changes and heterogeneous hardware. Recent studies reveal that domain adaptation is a suitable strategy to overcome the above drawbacks by reducing the discrepancy between two domains. However, most existing methods can only learn some shallow representation features, which just reduces the domain discrepancy to some extent. Comparatively, deep neural networks (DNN) can learn deep transferable features to manifest invariant factors underlying different domains. Thus, we propose a Deep Mean Correlation Alignment (DMCA) domain adaptation network to reduce the domain discrepancy efficiently. DMCA can learn transferable features from domain shift minimization by aligning the mean embeddings of the hidden representation in a reproducing kernel Hilbert space, and aligning the second-order statics of hidden representation of different domains. Experimental results demonstrate that our proposed approach outperforms other existing methods in accuracy and robustness.