Precise　forecasting　of　the　thermal　parameters　is　a　critical　factor　for　the　safe　operation　and　fault　incipient　warning　of　the　ultra-high　voltage　(UHV)　transformers.　In　this　work,　a　novel　multi-step　forecasting　method　based　on　the　long-　and　short-term　time-series　network　(LSTNet)　with　the　conditional　mutual　information　(CMI)　is　proposed　for　the　UHV　transformer.　To　improve　the　computational　efficiency　and　eliminate　the　redundancy,　the　CMI-based　feature　selection　algorithm　is　applied　to　analyse　the　correlation　between　the　original　monitoring　parameters　and　construct　the　optimal　feature　subset.　LSTNet,　which　is　composed　of　a　convolutional　layer,　recurrent　layer　and　recurrent-skip　layer,　is　utilized　to　capture　both　the　short-term　nonlinear　characteristics　and　the　long-term　periodic　characteristics.　The　LSTNet　model　is　established　to　forecast　the　variation　tendency　of　the　oil　and　winding　temperatures　for　different　locations　in　the　UHV　transformer.　The　results　show　that　the　proposed　method　significantly　enhances　the　accuracy　in　both　one-step　and　multi-step　thermal　parameters　forecasting　and　achieves　better　performance　in　terms　of　the　RMSE　and　MAE　compared　with　other　existing　methods.