Rapid　and　accurate　short-term　passenger　flow　prediction　plays　an　important　and　far-reaching　role　in　passenger　flow　control　and　early　warning.　In　fact,　the　short-term　passenger　flow　presents　the　characteristics　of　non-linearity　and　randomness.　Traditional　machine　learning　algorithms　can　hardly　meet　current　predictions.　In　this　paper,　the　random　forest(RF)　is　used　to　calculate　the　feature　importance　to　filter　the　extracted　features　and　remove　the　redundant　features,　and　we　apply　the　Long　Short-Term　Memory　network(LSTM)　algorithm　model　to　predict　the　short-term　passenger　flow　of　the　metro.　First,　we　calculate　the　out-of-bag(OOB)　error　of　the　features　by　RF　based　on　the　characteristics　of　bootstrap　sampling　and　regression　tree,　and　calculate　OOB　error　again　after　adding　noise.　According　to　the　two　OOB　errors,　the　feature　importance　can　be　obtained　through　related　formulas,　and　some　features　can　be　filtered.　RF　can　effectively　reduce　redundant　features　to　participate　in　calculation　and　improve　operating　efficiency.　Second,　we　apply　the　LSTM　model　to　predict　passenger　flow　every　10　minutes　for　each　station　and　use　the　important　features　selected　by　RF　as　the　model　inputs.　LSTM　has　an　excellent　effect　in　dealing　with　problems　that　are　highly　related　to　time　series,　and　it　is　very　suitable　for　prediction　on　time　series　issues.　The　proposed　model　is　evaluated　with　real　metro　card　data,　the　prediction　performance　compares　to　single　RF,　LSTM,　and　other　algorithm　models.　The　experiment　results　show　that　the　accuracy　of　the　RF　combined　LSTM　model　algorithm　is　better　than　that　of　other　existing　models　such　as　RF　and　LSTM　model.　It　shows　good　prediction　accuracy　and　has　far-reaching　significance　in　the　field　of　passenger　flow　prediction.　©　2020　IEEE.