The current research on non-technical loss(NTL) detection is mainly aim on residential customers, however, the related methods are not suit industrial and commercial customers. According to this problem, a deep learning based NTL detection method by embedding the principle of electricity measurement is proposed. Firstly, various NTL is analyzed and the phenomena show that only smart meter data is not enough for detecting NTL. Hence, smart meter data and some principles of electricity measurement are organized which describe the inherent relationship among electricity magnitudes as samples for deep learning. Secondly, an improved hybrid residual neural network is proposed to extract advanced features of NTL from massive smart meter data for detecting NTL. The experiment results show that the approach in this paper has achieved significant improvement on all metrics by comparing with the baselines.