With the formation of a "double high" power system characterized by high proportions of new energy and power electronic devices, subsynchronous oscillation (SSO) events in power systems have become more frequent. This paper proposes a method for locating SSO sources based on transient energy flow and spatiotemporal deep learning. The proposed deep learning method, composed of a Graph Convolutional Network (GCN) and a Transformer network, aims to accurately locate SSO sources under limited observability of the power system. First, the propagation mechanism of transient energy in complex power grids is studied, laying the foundation for modeling the disturbance source location problem as a classification problem. Then, the GCN method is used to construct graph data based on node features and topology information, extracting spatial features of the oscillations. Simultaneously, the Transformer method extracts temporal correlations of oscillation data across multiple nodes. A spatiotemporal deep learning network model is trained by integrating spatial and temporal features. Finally, case studies on the modified IEEE 39-bus system validate the effectiveness of the proposed method. The case analysis shows that the proposed T-GCN model has higher localization accuracy and noise resistance and performs well despite poor system observability and topological changes.