The “dual-high” characteristics of the novel power system pose new challenges to reactive power and voltage control in power systems. This paper proposes a multi-scale reactive power and voltage control method for power systems based on multi-source data fusion and knowledge graphs. It integrates the logical reasoning and autonomous learning capabilities of knowledge graphs and deep deterministic policy gradient algorithms to construct an efficient and interpretable reactive power and voltage decision-making framework. Based on knowledge graphs, the relationships between grid entities and components are mapped from high-dimensional space to low-dimensional vector space. The translational embedding (TransE) model is used to quantify the fit degree of triples, and symbolic logical reasoning is performed through production rules and confidence calculations to generate a set of recommended execution actions. The recommended actions output by the knowledge graphs are integrated into the deep deterministic policy gradient algorithm framework. A hybrid reward function is designed considering network loss and voltage deviation, reducing the blind randomness of reinforcement learning during exploration and improving the convergence speed of the policy network. Experiments verify the effectiveness of the proposed method. The algorithm achieves significant multi-index effects in network loss optimization, voltage stability, and control performance after introducing knowledge graphs.