Anomalies　in　vaccine　data　are　often　encountered　when　projecting　the　active　cases　in　Africa.　This　issue　could　severely　impact　the　accuracy　of　expected　active　case,　thereby　affecting　future　government　intervention　protection　policies　etc.　In　this　paper,　a　new　method　to　accurately　predict　active　cases　considering　anomalies　is　presented.　Firstly,　an　autoencoder　is　used　to　detect　the　vaccine　anomalies,　Then,　in　order　to　impute　missing　values　in　the　complete　African　vaccination　data　a　small-scale　k　nearest　neighbor　(ssKNN)　with　upper　and　lower　bounds　is　developed.　As　an　extension　of　K-nearest　neighbor　interpolation,　ssKNN　could　guarantee　the　interpolation　quality　and　efficiency　by　limiting　the　filling　range.　To　be　specific,　the　upper　and　lower　bounds　of　missing　values　are　found　based　on　the　proposed　method,　and　then　the　missing　values　could　be　filled　according　to　the　upper　and　lower　bounds　as　well　as　the　Euclidean　distance.　Then,　the　transmission　rate　ss　is　estimated　using　long　short-term　memory　(LSTM),　which　is　an　essential　parameter　in　a　classical　mechanistic　model　for　projecting　the　number　of　future　active　cases.　Finally,　the　effectiveness　of　proposed　method　is　verified　on　dataset　of　15　worst-affected　countries　by　COVID-19　epidemic　in　Africa　demonstrating　its　efficacy　in　prediction　of　active　cases　of　COVID-19　with　vaccine　anomalies.