Data　aggregation　is　a　key　technology　in　wireless　sensor　networks　(WSNs),　enabling　WSNs　to　be　widely　used　in　many　fields　such　as　environment　monitoring.　It　can　not　only　reduce　redundant　data　transmission　during　data　gathering,　but　also　effectively　improve　data　accuracy.　However,　poor　anti-interference　ability,　limited　energy　and　other　reasons　existed　in　sensor　nodes　make　data　error-prone,　thus　affecting　quality　of　the　aggregation　results.　In　this　paper　we　propose　a　prediction-based　fault-tolerant　aggregation　algorithm,　which　is　implemented　on　a　constructed　aggregation　tree.　Using　temporal-spatial　correlation　between　data,　the　algorithm　can　distinguish　different　data　types.　And　then　we　make　a　specific　analysis　of　faulty　data　that　may　occur　during　data　gathering　and　pointedly　put　forward　different　fault-tolerant　solutions　using　the　PSO-BPNN-based　prediction　model.　Experimental　results　show　that　the　proposed　algorithm　can　effectively　predict　data　and　exclude　the　impact　of　faulty　data,　ensuring　real-time　and　fault-tolerant　requirements.