Short-term　weather　forecasting　is　of　great　significance　to　people＇s　lives,　especially　in　terms　of　transportation.　Since　daily　precipitation　can　be　regarded　as　a　nonlinear　and　non-stationary　time　series,　it　is　difficult　to　predict　it.　Based　on　the　contrastive　learning　method,　A　multimodal　representation　method　for　forecasting　daily　precipitation　is　proposed.　Specifically,　an　encoder　is　used　to　convert　the　original　weather　data　into　feature　representation,　then　it　is　optimized　through　contrastive　learning,　and　the　optimized　features　are　used　for　prediction.