Against the backdrop of the rapid development of smart grids, the scale and complexity of power systems have surged, leading to differences in the representation of the same grid entity across different business systems and application scenarios. This heterogeneity severely hinders the fusion and analysis of grid data. To address the challenge of entity alignment in large-scale power grids and support data fusion and intelligent decision-making, this study proposes a multilevel entity representation alignment technology based on contrastive learning. This research introduces an innovative multilevel entity representation alignment method. Specifically, a Multilevel Power Grid Entity Representation model (MPGER) is constructed, comprising an explicit entity alignment module, an implicit alignment module, and a contrastive learning module. Explicit entity encoding is achieved through a graph attention network, while implicit entity encoding is realized using a multi-head attention mechanism. Meanwhile, a contrastive learning strategy optimizes entity representations. The node representation learning method designed based on graph neural networks incorporates attention mechanisms to effectively handle the heterogeneous and dynamic characteristics of power grids. Additionally, the proposed entity alignment algorithm comprehensively considers the semantic, structural, and attribute similarities of entities, adopting an iterative optimization strategy to improve alignment accuracy. Experiments were conducted using three datasets: public datasets DWY100K, DBP15K, and the Nanjing Grid NanJ-Grid dataset. The results show that on the DWY100K dataset, MPGER achieves a 3.9% improvement in F1-score compared to state-of-the-art (SOTA) methods; on the DBP15K dataset, the F1-score increases by 4.2%; and on the NanJ-Grid dataset, the F1-score improves by 1.2%. Ablation experiments further validate the effectiveness of the model’s explicit and implicit alignment modules.