Forecasting　cargo　throughput　is　an　essential　albeit　challenging　task　for　national　and　port　optimisation　decision-making,　resource　allocation,　and　control　planning.　To　this　end,　a　novel　forecasting　model　is　developed　for　mixed-frequency　data　called　attention-DeepAR-MIDAS　(ADM)　by　introducing　the　mixed　data　sampling　(MIDAS)　technique　and　attention　mechanism　into　the　DeepAR　forecasting　algorithm　in　this　study.　The　proposed　ADM　model　is　specifically　designed　with　an　attention　mechanism　to　accurately　identify　and　prioritise　the　most　influential　variables,　both　endogenous　and　exogenous,　over　time.　Hence,　it　can　effectively　use　the　nonlinear　information　of　mixed-frequency　data,　which　is　conducive　to　port　throughput　forecasting.　Furthermore,　the　ADM　model　possesses　both　long-term　and　short-term　high-precision　forecasting　capabilities,　enabling　multi-step　probability　forecasting　and　better　tracking　of　abnormal　changes　in　endogenous　and　exogenous　variables　of　port　throughput,　fitting　their　fluctuation　trends.　By　analysing　the　differences　in　model　performance　before　and　after　improvement　based　on　forecast　accuracy　metrics,　probability　interval　testing,　and　DM　testing　methods,　the　ADM　model　achieves　accurate　forecasting.　Finally,　China＇s　monthly　port　throughput　forecast　results　also　illustrate　the　superiority　of　the　ADM　model,　which　provides　decision-makers　with　more　timely,　accurate,　and　comprehensive　forecasts.