Groundwater serves as both an economic and strategic resource, as well as a crucial factor in maintaining ecosystems. Therefore, the accuracy of groundwater level simulation plays a significant role in sustainable groundwater resource utilization and management. Machine learning methods have been widely applied in groundwater level simulation for capturing the nonlinear relationships between input variables and target variables. However, traditional machine learning methods do not consider the spatial relationships between stations, making it challenging to handle unstructured monitoring data with irregular station distributions. In this study, we employ Graph Neural Networks (GNN) to simulate the dynamic changes in groundwater levels, where each monitoring station can be represented as a node in the graph, and the connections between nodes are determined by adjacency matrices. The proposed model is applied and evaluated using monitoring data from a typical groundwater depression zone in Hebei Province. Compared with three benchmark models, Random Forest (RF), Support Vector Machine (SVM), and Multi-layer Perceptron (MLP), the proposed model demonstrates better performance in the defined evaluation metrics. Additionally, the proposed model has higher data utilization efficiency compared to single-station models as it can simultaneously simulate the groundwater level changes at all monitoring stations in the modeling system.