This　paper　presents　an　improved　regression　algorithm　of　sliding　window　least　squares　support　vector　machine　(the　Sliding　Window　LS-SVM).　This　method　simplifies　the　data　within　the　sliding　window,　and　selects　the　similar　data　for　local　modeling　from　a　database　of　historical　batches　to　predict　the　data　within　the　sliding　window.　Combined　with　local　modeling,　the　improved　sliding　window　LS-SVM　algorithm　is　very　effective　to　predict　the　cell　concentration　in　the　penicillin　fermentation　process.　©　2013　AACC　American　Automatic　Control　Council.