To　accurately　predict　the　concentration　of　PM2.5　in　the　atmosphere,　this　paper　establishes　LSSVR　prediction　model　based　on　historical　data　of　atmospheric　PM2.5　concentration.　The　parameters　of　LSSVR　model　are　optimized　by　particle　swarm　optimization　algorithm　(PSO).　According　to　PM2.5　concentration　data　per　hour　and　meteorological　conditions　from　June　to　August　2017　in　Beijing,　other　PM2.5　concentration　prediction　models　are　established,　which　include　ANN　prediction　model　and　$\varepsilon$-SVR　prediction　model.　By　comparing　the　prediction　errors　of　these　three　prediction　models,　the　calculated　mean　absolute　error　of　the　ANN　prediction　model　was　25.24%,　the　mean　absolute　percent　error　of　$\varepsilon$-SVR　is　10.39%,　and　the　mean　absolute　percent　error　of　PSO-LSSVR　model　is　4.95%.　The　simulation　results　show　that　the　PSO-LSSVR　model　is　better　than　ANN　model　and　$\varepsilon$-SVR　model,　and　the　PSO-LSSVR　model　has　less　computational　time　and　reduces　the　complexity　of　the　algorithm.　Therefore,　the　proposed　PSO-LSSVR　algorithm　is　effective　and　reliable　by　predicting　PM2.5　concentration.　©　2019　IEEE.