The　serious　global　issue　has　arisen　from　the　escalating　frequency　and　intensity　of　extreme　weather　events,　requiring　collaborative　efforts　to　predict　such　catastrophes　accurately　and　to　prevent　their　potentially　disastrous　repercussions.　Consequently,　the　global　agenda　has　elevated　the　priority　of　forecasting　extreme　weather　phenomena　and　devising　strategies　to　mitigate　their　adverse　effects.　However,　extreme　weather　data　exhibit　high　levels　of　noise　and　are　often　accompanied　by　the　common　data　loss　issues.　Therefore,　the　current　research　tries　to　address　how　to　effectively　capture　the　spatio-temporal　characteristics　of　the　data　and　incorporate　global　information　to　mitigate　this　problem.　In　the　paper,　we　introduced　SA-ConvLSTM,　a　novel　extreme　weather　prediction　method　based　on　Convolutional　Long　Short-Term　Memory　(ConvLSTM)　with　the　Self-Attention　(SA)　Mechanism.　The　proposed　method　harnessed　the　power　of　ConvLSTM　to　synergize　spatial　and　temporal　feature　extraction　capabilities,　and　effectively　extract　long-range　spatio-temporal　patterns　present　in　the　weather　time　series.　Additionally,　the　employs　a　SA　mechanism　to　excel　in　modeling　global　dependencies　and　adaptive　feature　weighting,　which　can　increase　effectiveness　in　capturing　long-range　patterns　and　irregularities　within　time　series　data.　The　experimental　results　on　a　genuine　dataset　from　the　Japanese　Advanced　Himawari-8　geostationary　satellite　demonstrated　that　our　approach　has　a　great　potential　for　extreme　weather　forecasting　tasks.　©　2023　IEEE.