As　one　of　the　key　technologies　of　intelligent　transportation　systems,　short-term　traffic　volume　prediction　plays　an　increasingly　important　role　in　solving　urban　traffic　problems.　In　the　last　decade,　many　approaches　were　proposed　for　the　traffic　volume　prediction　from　different　perspectives.　However,　most　of　these　approaches　are　based　on　a　large　amount　of　historical　data.　When　there　are　only　finite　collected　traffic　data,　they　cannot　be　well　trained,　so　the　prediction　accuracy　of　these　approaches　will　be　poor.　In　this　paper,　a　tensor　model　is　proposed　to　capture　the　change　patterns　of　continuous　traffic　volumes.　From　collected　traffic　volume　data,　the　element　data　are　extracted　to　update　the　corresponding　elements　of　the　tensor　model.　Then,　a　tucker　decomposition　and　gradient　descent　based　algorithm　is　employed　to　impute　the　missing　elements　of　the　tensor　model.　After　missing　element　imputation,　the　tensor　model　can　be　directly　applied　to　the　short-term　traffic　volume　prediction　through　searching　the　corresponding　elements　of　the　model　and　the　storage　cost　of　the　model　is　low.　Our　model　is　evaluated　on　real　traffic　volume　data　from　PeMS　dataset,　which　indicates　that　our　model　has　higher　traffic　volume　prediction　accuracy　than　other　approaches　in　the　situation　of　finite　traffic　volume　data.　©　2023,　Systems　Engineering　Society　of　China　and　Springer-Verlag　GmbH　Germany.