Container　throughput　has　been　perceived　as　one　of　the　most　important　factors　of　port　economic　development.　Accurate　forecasting　of　container　throughput　can　not　only　improve　the　efficiency　of　container　operation　but　also　meet　the　requirements　of　financial　trading.　This　paper　proposes　a　secondary　decomposition　(SD)　learning　approach,　which　is　integrated　with　a　complementary　ensemble　empirical　mode　decomposition　(CEEMD),　sample　entropy　(SE),　wavelet　packet　decomposition　(WPD),　extreme　learning　machine　(ELM)　and　kernel　extreme　learning　machine　(KELM),　to　implement　monthly　container　throughput　forecasting.　The　CEEMD　is　conducted　to　decompose　the　original　container　data　into　several　intrinsic　mode　function　components　and　a　residual　term.　The　complexity　of　the　subseries　is　measured　by　SE.　Subsequently,　the　WPD　is　used　to　further　smooth　the　most　complex　component.　Then,　ELM　is　executed　to　forecast　the　primary　decomposition　results,　and　KELM　is　employed　to　predict　the　secondary　decomposition　results.　The　final　ensemble　results　can　be　obtained　by　integrating　all　of　the　forecasting　results.　The　empirical　results　indicate　that　the　SD　learning　approach　is　an　excellent　tool　for　forecasting　nonlinear　and　nonstationary　container　throughput.　(C)　2019　Elsevier　B.V.　All　rights　reserved.