Compared　with　the　traditional　SVM,　LS_SVM　improves　the　speed　of　the　large　sample　study,　however,　LS_SVM　loses　its　sparsity,　which　is　an　advantage　of　the　the　traditional　SVM.　For　LS_SVM,　almost　all　of　the　samples　are　support　vectors.　When　a　new　sample　added　all　samples　are　re-involved　in　the　calculation,　and　that　greatly　increases　the　amount　of　the　computation.　In　response　to　these　issues,　this　paper　presents　an　algorithm　called　the　LSSVM　sliding　window　pruning　algorithm,　which　is　based　on　DTW.　With　the　combination　of　local　modeling,　the　algorithm　selects　the　most　similar　samples　with　the　current　sample　point　in　the　process　of　the　establishment　of　the　on-line　model　of　the　batch　process　as　the　training　samples.　It　can　remove　those　samples　that　are　relatively　ineffective,　which　can　restore　LSSVM　＇s　sparsity.　Taking　the　cell　concentration　in　the　fermentation　process　of　E.　Coli　for　example,　this　paper　proves　that　the　method　can　establish　a　more　accurate　on-line　prediction　model　in　the　case　of　a　smaller　number　of　samples.　Compared　with　the　off-line　LSSVM,　the　on-line　model　has　a　higher　accuracy　and　a　better　dynamic　adaptability.　©　(2013)　Trans　Tech　Publications,　Switzerland.