Deep Transfer Learning for Power Substation Recognition with Google Earth

Abstract

Power substation is a vital component of power transmission systems. However, their location information, longitude and latitude, maybe incorrect due to human errors. For example, an employee of a power grid company may carelessly input wrong location information to the system database. In practice, these errors residue in database and are difficult to be corrected or even unaware of because human naturally are insensitive to these long numbers. Correct location information is critical and it will ensure destroyed power substations to be quickly pinpointed and repaired after earthquake. In this paper, we develop a deep transfer learning approach to automatically verify power substation location information based upon Google Earth (GE)images. We first extract an image patch for a substation from GE using the recorded location information in database. We then apply the deep transfer learning model to verify whether the image patch contains a power substation. We validated the proposed method on more than 1000 substations in Yunnan Electric Power Grid. Our experimental results showed that the Xception deep transfer model achieved the best accuracy of 98.5% among four compared machine learning methods with an image size of 299⨯299 pixels. The developed model is a promising tool to verify the location information of power substations in database.