This　study　presents　two　dynamic　models,　namely　recurrent　neural　network　and　long　short-term　memory　(LSTM)　models,　for　predicting　PM2.5　concentrations　in　Taiwan　by　using　PM2.5　time　series　obtained　at　air　quality　monitoring　stations　and　weather　information　obtained　at　neighboring　weather　stations.　The　proposed　models　can　efficiently　predict　PM2.5　by　incorporating　a　learned　memory　structure　with　a　forgetting　gate.　To　evaluate　the　predictive　performance　of　the　proposed　models,　large-scale　databases　established　by　Taiwan＇s　Environmental　Protection　Administration,　and　Central　Weather　Bureau　were　used;　these　databases　include　hourly　data　from　77　air　quality　monitoring　stations　and　580　weather　stations　over　a　1-year　period.　The　results　demonstrated　that　the　proposed　models　outperformed　three　traditional　machine　learning　methods　(gradient　boosting,　support　vector　machine,　and　classification　and　regression　tree　models)　by　27.12%　and　33.69%　on　average　in　terms　of　the　coefficient　of　determination　and　root　mean　square　error,　respectively.　A　geographical　divergence　analysis　was　conducted　to　compare　predictive　performance　in　different　regions.　The　results　revealed　that　the　most　significant　improvement　in　predictive　performance　was　achieved　in　central　Taiwan.　The　seasonal　and　pollution　effect　on　predictive　performance　were　reduced　by　the　LSTM　and　the　source　distribution　of　PM2.5　emission　in　Taiwan　was　also　analyzed.