In　recent　years,　unmanned　aerial　vehicle　(UAV)　autonomous　flight　technology　has　been　applied　in　many　fields.　However,　in　the　process　of　autonomous　operation,　the　UAV　may　deviate　from　the　set　flight　path　due　to　various　disturbance　factors,　which　results　in　mission　failure.　In　order　to　find　the　abnormal　situation　in　time　and　take　corresponding　measures,　it　is　necessary　to　monitor　the　operation　state　of　the　UAV.　Predicting　the　UAV　flight　path　is　the　main　monitoring　method　at　present;　however,　the　accuracy　and　real-time　of　the　existing　prediction　methods　are　limited.　Therefore,　this　paper　proposes　an　error　compensation　Bessel　bidirectional　long　short-term　memory　real-time　path　prediction　model　deployed　in　ground　stations.　First,　because　of　inconsistency　of　the　units　in　all　directions　of　the　original　positioning　information　provided　by　global　positioning　system,　the　Bessel　geodetic　coordinate　transformation　is　introduced　to　unify　the　units　of　three-dimensional　coordinate　data,　so　as　to　improve　the　prediction　accuracy.　Second,　considering　the　problems　of　poor　data　quality　and　data　missing　in　the　operation　process,　the　least　square　fitting　method　is　used　to　supplement　and　correct　the　positioning　coordinate　data　to　obtain　more　reliable　and　accurate　path　observation　values　as　the　model　input.　Finally,　a　deep　learning　path　prediction　model　based　on　bi-directional　long　short-term　memory　(BiLSTM)　network　is　constructed,　and　the　appropriate　network　parameters　are　determined　with　the　prediction　accuracy　and　time　as　evaluation　indicators.　In　order　to　further　improve　the　prediction　accuracy,　a　compensator　based　on　proportional　integral　differential　error　control　theory　is　designed　according　to　the　output　characteristics　of　the　BiLSTM　network,　which　is　used　for　providing　compensation　values　for　the　prediction　results　of　the　model.　The　training　and　testing　results　using　the　actual　flight　data　of　UAV　operation　show　that,　under　the　experimental　environment　built,　the　model　proposed　in　this　paper　can　complete　the　UAV　flight　path　prediction　with　root　mean　square　error　＜　1　meter　within　0.1　second,　and　has　better　performance　and　higher　prediction　accuracy　than　other　neural　network　models.