Optical　computing　power　network　(OCPN)　has　emerged　as　a　key　enabler　for　the　era　of　AI.　However,　OCPN　operators　find　it　difficult　to　accommodate　increasing　bulk　data　transfers　over　end-to-end　lightpaths　due　to　the　spatiotemporal　volatility　of　background　traffic.　In　this　paper,　we　propose　a　graph-classifier-aided　store-and-forward　(SnF)　scheduling　method,　namely　GCS,　for　bulk　data　transfers　in　the　OCPN.　Instead　of　formulating　a　static　optimization　model,　GCS　formulates　the　SnF　problem　as　a　routing　model　on　a　multilayer　graph.　Instead　of　searching　the　entire　graph　directly,　GCS　decomposes　it　into　multiple　fixed-route　subgraphs　and　leverages　a　GCN　graph　classifier　to　exclude　infeasible　subgraphs　as　well　as　narrow　down　the　search　scope　of　subgraphs.　GCS　not　only　reduces　the　complexity　of　the　problem　and　the　difficulty　of　the　model　learning.　Studies　show　that　GCS　obtains　lower　blocking　probability　and　shorter　searching　time　than　the　existing　methods.