With the continuous development of smart grid technology, people’s demand for smart electricity consumption is increasing, and electricity consumption identification is a key aspect of achieving smart electricity consumption. Therefore, to promote the development of electricity consumption identification, we have studied load identification methods in low-voltage (LV) station areas and proposed a set of load identification models based on deep reinforcement learning. Each model consists of a non-intrusive load monitoring (NILM) device, an improved adaptive density peak (ISDPC) model, a new end-to-side neural network called GhostNet, and a data processing and analysis module specific to 10 kV power transmission. Considering the complexity and diversity of loads in the 10 kV power transmission system, we employ ISDPC algorithm to perform cluster analysis on load characteristic data and use GhostNet for load identification. Additionally, we preprocess and extract features from the data specific to 10 kV power transmission to improve the accuracy and effectiveness of the identification. Finally, we compare our results with those of the k-means clustering algorithm, Euclidean distance load curve clustering, and other algorithms to demonstrate the superiority of our method in terms of clustering and identification accuracies.