Abstract: This study investigates the problem of missing and anomalous multi-source heterogeneous data in smart grids and proposes an intelligent processing and question answering framework that integrates reinforcement learning with knowledge graphs. The system employs a multimodal encoder for data preprocessing and constructs a comprehensive power grid knowledge graph to integrate diverse information sources. Leveraging deep reinforcement learning (DQN), the data completion task is formulated as an agent-driven path exploration process within the knowledge graph, with a reward mechanism explicitly incorporating domain-specific knowledge of power systems. Furthermore, the framework supports natural language question answering, where semantic parsing and graph-based querying are used to generate precise responses. Experimental evaluations demonstrate that the proposed method substantially enhances data quality and question answering accuracy, thereby offering robust support for intelligent power grid management.