The　forecasting　of　top　oil　temperature　can　provide　a　significant　basis　for　insulation　aging　assessment　and　fault　warning　of　UHV　transformers.　This　paper　proposes　a　prediction　method　for　top-level　oil　temperature　of　ultra-high　voltage　(UHV)　transformer　based　on　conditional　mutual　information　(CMI)　as　well　as　long-term　and　short-term　time-series　network　(LSTNet).　Based　on　historical　data,　including　top　oil　temperature,　dissolved　gases　in　oil,　winding　temperature,　winding　current,　environmental　temperature　and　other　nine　parameters,　the　CMI　method　is　used　to　select　the　characteristics　with　greater　information　gain　for　the　top　oil　temperature　forecast　to　reduce　the　input　characteristic　dimension.　On　this　basis,　LSTNet　is　applied　to　extract　the　long-term　periodic　law　and　short-term　nonlinear　variation　contained　in　the　characteristic　variables.　CMI-LSTNet　model　is　established　to　realize　the　prediction　of　top　oil　temperature　at　multiple　parts　of　UHV　transformer.　The　results　show　that　the　addressed　method　not　only　can　describe　the　change　tendency　of　top　oil　temperature　of　UHV　transformer　effectively,　but　also　has　higher　prediction　accuracy,　compared　with　the　existing　typical　forecasting　methods.　©　2022　Power　System　Technology　Press.　All　rights　reserved.