Edge　computing　power　optical　network　(ECPON)　has　emerged　as　a　solution　for　providing　last-mile　AI　access,　bringing　with　it　an　urgent　need　for　efficient　resource　allocation,　which　cannot　be　achieved　without　accurate　traffic　prediction　and　estimation　of　the　network　performance.　However,　privacy　concerns　and　data　heterogeneity　prohibit　the　conventional　prediction　and　estimation　approaches　from　becoming　practical　solutions　for　the　ECPON.　In　this　paper,　a　hierarchical-federated-learning-aided　adaptive　upstream　transfer　scheme　is　presented.　On　one　hand,　it　combines　shape-based　traffic　clustering,　wavelet　analysis　and　GRU　into　an　HFL-aided　predictor　to　perceive　future　traffic　fluctuation.　On　the　other　hand,　it　uses　HFL-aided　estimators　to　perceive　the　ECPON　performance　metrics.　As　such,　it　can　allocate　resources　to　adapt　to　traffic　fluctuation　in　advance.　Simulations　show　that　the　proposed　scheme　offers　near-optimal　allocation　compared　with　the　conventional　schemes.