To　overcome　the　deficiency　of　Support　Vector　Machine　(SVM)　for　regression,　dynamic　epsilon-SVM　method　was　proposed.　To　establish　precise　mathematical　models,　a　new　modeling　method　was　introduced,　combining　self-organizing　feature　map　(SOFM)　with　the　dynamic　epsilon-SVM.　Firstly,　SOFM　was　used　as　a　clustering　algorithm　to　partition　the　whole　input　space　into　several　disjointed　regions;　then,　the　dynamic　epsilon-SVM　modeled　for　these　partitioned　regions.　This　method　was　illustrated　by　modeling　penicillin　fermentation　process　with　plant　field　data.　Results　show　that　the　method　achieves　significant　improvement　in　generalization　performance　compared　with　other　methods　based　on　SVM.