Line-of-sight (LOS) or non-line-of-sight (NLOS) identification is of vital significance to the localization of mobile sensors in intelligent substations for power Internet of Things. This article investigates the LOS/NLOS identification in substation scenarios, based on deep learning networks and feature fusion methods. Channel measurement data in high-voltage substation environments with LOS and NLOS cases are collected, and both original and manually extracted channel features are obtained to generate data sets. A novel LOS/NLOS identification model is proposed, which employs a deep neural network and a self-attention network to separately learn the information contained in the manually extracted channel features and the original channel feature. This model also applies a hybrid fusion method to capture correlation between the channel features and mitigate data inundation risk caused by the dimension difference of input features. The results of performance evaluation show that the proposed model not only has the identification accuracy as high as 98.95%, but also possesses good noise robustness and acceptable computational complexity.