With the continuous increase in the scale of power customers, traditional power customer service systems can no longer meet customer service needs. To address the issues of poor understanding of user intentions and low logical coherence in power customer service dialogues, a power customer service dialogue method based on retrieval enhanced generation technology is proposed. Firstly, language and text cleaning was carried out to reduce the impact of data noise on the power customer service dialogue model. Secondly, natural language processing is used for semantic analysis to grasp the real demands of power customers. Based on this, retrieval enhanced generation technology is adopted, combined with external data retrieval and the advantages of generative models, to respond to power customer service conversations. Finally, simulation verification was conducted on the conversation data of power customers from a certain provincial power company, and the results showed that the proposed method has better response performance and shorter response time compared to bidirectional recurrent neural networks.