The　top　oil　temperature　is　a　critical　indicator　for　assessing　the　insulation　performance　and　load　capacity　of　a　power　transformer.　Hence,　predicting　the　top　oil　temperature　is　essential　for　the　predictive　maintenance　of　transformers　and　ampacity　prediction.　In　this　paper,　a　long-term　and　short-term　prediction　method　based　on　comprehensive　thermal　factors　and　CNN-LSTM-Attention　is　proposed　to　improve　the　prediction　accuracy　of　top　oil　temperature　estimation　and　prove　the　physical　interpretability.　Firstly,　the　coefficient　method　is　employed　to　identify　the　key　characteristics　that　affect　the　top　oil　temperature　of　transformers.　The　thermal　parameter　analysis　is　performed　based　on　the　conventional　models　to　identify　a　comprehensive　thermal　factor.　Secondly,　the　local　feature　is　extracted　using　convolutional　neural　networks　(CNN).　The　long　short-term　memory　network　(LSTM)　is　employed　to　capture　the　temporal　information.　Thirdly,　the　attention　mechanism　is　used　to　allocate　weights　reasonably.　Finally,　the　top　oil　temperature　prediction　model　is　constructed　with　the　CNN-LSTM-Attention　algorithm,　allowing　the　long-term　and　short-term　predictions　on　a　220　kV　transformer.　The　results　show　that　the　minimum　value　of　the　mean　absolute　percentage　error　(MAPE)　for　the　long-term　prediction　is　0.9749　%,　while　it　is　0.6429　%　for　the　short-term　prediction,　which　is　better　than　the　1.4963　%　and　1.2082　%　of　the　conventional　models.　©　2025　The　Author(s)