Power transmission and distribution systems are heavily reliant on power cables. The stable operation of power grids is directly tied to cable operational safety. Cable temperature is utilized as the primary metric for operational status assessment. However, anomaly identification is typically restricted to simple threshold-based alarms. These alarms are dependent on strict upper and lower temperature limits. Complex anomalies are frequently missed by this conventional approach. Significant scientific challenges are encountered in cable temperature anomaly detection. These challenges are characterized by complex temporal trends, strong spatial coupling, and severe ambient temperature interference. A joint modeling method based on a Transformer and a Graph Neural Network (GNN) is proposed to address these issues. Long-term dependencies and variation trends of temperature sequences are extracted by the Transformer. Dynamic features like periodicity, trends, and mutations are depicted by this component. Topological correlations and local relevance among measurement points are identified by the GNN. The underlying patterns of spatial propagation and coordinated variation are captured. Additionally, an ambient temperature correction module is designed for environmental constraints. Ambient temperature conditions are incorporated by this module through spatial alignment and attention fusion.