This　paper　develops　a　new　predictive　approach　to　improve　the　static　VAr　compensator　(SVC)　performance　in　the　electric　arc　furnaces　(EAFs).　The　proposed　method　models　the　reactive　power　consumption　pattern　in　the　EAF　for　a　half-cycle　ahead　to　improve　the　SVC　compensation　process.　Given　this,　a　new　nonparametric　approach　based　on　lower　upper　bound　estimation　method　and　support　vector　regression　(SVR)　is　developed　to　construct　prediction　intervals　(PIs)　around　the　reactive　power　consumption　pattern　in　the　SVC.　The　proposed　method　makes　use　of　the　PI　concept　to　model　the　uncertainties　of　reactive　power　and,　thus,　avoid　the　flicker　issues.　Owing　to　the　high　complexity　and　nonlinearity　of　the　proposed　problem,　a　new　optimization　method　based　on　the　krill　herd　(KH)　algorithm　is　proposed　to　adjust　the　SVR　setting　parameters,　optimally.　Also,　a　three-stage　modification　method　is　suggested　to　increase　the　krill　population　and　avoid　the　premature　convergence.　The　feasibility　and　performance　of　the　proposed　method　are　examined　using　experimental　data　gathered　from　the　Mobarakeh　Steel　Company,　Iran.　©　1982-2012　IEEE.