Accurate　prediction　of　PM2.5　concentration　at　the　hour　scale　can　improve　the　capabilities　of　air　pollution　forecast.　Traditional　models　for　PM2.5　prediction　rely　on　a　number　of　selected　factors,　while　the　selection　of　factors　is　hard　to　conduct.　In　this　paper,　we　propose　a　new　model　using　a　Stack　Sparse　Auto-Encoder　(SSAE)　and　a　Long-Short　Term　Memory　(LSTM)　neural　network　for　PM2.5　concentration　prediction,　referred　to　as　SSAE-LSTM.　We　use　the　SSAE　method　to　automatically　extract　high-level　features　with　respect　to　PM2.5　based　upon　several　factors　regarding　time,　space,　weather　and　air　pollution.　The　extracted　features　are　then　used　to　build　a　time-series　LSTM　model　that　is　able　to　extract　effective　spatial-temporal　features.　The　proposed　SSAE-LSTM　model　was　evaluated　using　datasets　of　air　pollution　and　weather　collected　from　71　air　monitoring　sites　in　Beijing-Tianjing-Hebei　urban　areas　from　2016　to　2018,　and　compared　with　existing　methods.　Experimental　results　show　that　the　predicted　results　by　the　proposed　method　had　a　higher　accuracy　than　that　of　existing　methods,　having　an　IA　index　over　0.99　for　all　testing　datasets,　while　the　corresponding　RMSE　and　MAE　were　dropped　to　13.98　and　7.90,　respectively.　Furthermore,　we　applied　the　SSAE-LSTM　to　the　datasets　of　71　air　monitoring　sites　covering　the　spring,　summer,　autumn,　and　winter　seasons.　Results　showed　that　the　predicted　values　were　highly　correlated　with　ground-truth　values　for　all　seasons,　with　coefficients　of　0.86,　0.92,　0.96,　0.93,　respectively.　Moreover,　the　predicted　results　of　the　Wanshouxigong　site　in　Beijing　revealed　that　the　SSAE-LSTM　model　can　be　effectively　used　for　PM2.5　prediction　at　various　air　quality　conditions.　©　2020,　Science　Press.　All　right　reserved.