This　paper　proposes　a　novel　Graph　Recurrent　Neural　Network　(GRNN)　based　approach　for　detecting　eavesdropping　attacks　in　smart　grid　wireless　communication　systems　enabled　by　Simultaneous　Wireless　Information　and　Power　Transfer　(SWIPT).　By　leveraging　the　graph-centric　nature　of　GRNNs,　the　proposed　method　effectively　learns　the　topological　structure　and　edge　features　of　Wireless　Sensor　Networks　(WSNs),　enabling　the　detection　of　eavesdropping　attacks　in　dynamic　WSNs.　This　paper　mathematically　models　the　Channel　State　Information　(CSI)　under　man-in-the-middle　eavesdropping　attacks　based　on　Physical　Layer　Security　(PLS)　in　SWIPT　networks.　Moreover,　this　paper　sets　up　a　real-world　testbed　to　create　training　and　testing　datasets.　The　proposed　GRNN　model　can　handle　large-scale　complex　topologies　and　dynamic　eavesdropping　networks,　accurately　detect　eavesdropping　behaviors,　and　enhance　the　security　of　information　transmission　in　WSNs.　Simulation　results　demonstrate　that,　compared　with　algorithms　such　as　Support　Vector　Machine　(SVM),　K-Nearest　Neighbors　(KNN),　Convolutional　Neural　Network　(CNN),　Graph　Convolutional　Network　(GCN),　and　Gated　Recurrent　Unit　(GRU),　the　proposed　method　exhibits　stronger　robustness　under　complex　attack　scenarios,　achieving　a　detection　accuracy　of　over　95%.　This　paper　provides　a　novel　and　effective　graph　learning　solution　for　smart　grid　wireless　communication　security,　which　is　of　great　significance　to　ensure　the　stable　and　reliable　operation　of　smart　grids.　IEEE