Abstract: To address the challenges of poor observability, coupling of multiple oscillation sources, and noise interference in locating sub-synchronous oscillation (SSO) sources in high-penetration renewable energy grids, this paper proposes an intelligent oscillation source localization method that integrates transient energy flow spatiotemporal characteristics with a Transformer model. First, based on the transient energy flow signals absorbed by the grid's positive damping elements, adaptive variational modal decomposition (AVMD) is used to dynamically optimize modal parameters and suppress modal aliasing. Combined with singular value decomposition (SVD), the principal components are selected to reconstruct the signal, enhancing the robustness of key features; Second, the reconstructed signals are converted into high-dimensional spatiotemporal feature images using a topology-constrained Markov transition field (T-MTF), and a dataset representing the energy correlation characteristics of wind farm nodes is constructed. Finally, a spatiotemporal Transformer model is designed, integrating a multi-head attention mechanism to focus on key source node features, embedding busbar location encoding to model the power grid topology, and employing a residual feature selection module to enhance noise resistance. Through improved IEEE-39 node renewable energy system simulation verification, it is demonstrated that the proposed method significantly improves localization accuracy and robustness compared to traditional methods under scenarios with strong noise, low observability, and multiple oscillation sources, providing an effective solution for real-time localization of oscillation sources in renewable energy grids.