Five　typical　partial　discharge　(PD)　models　of　oil-paper　insulation　are　made,　and　the　time　series　of　the　Models＇　PD　pulse　amplitude　are　extracted　from　the　measurements.　Then　the　time　series　are　pre-processed　and　further　fitted　by　the　auto　regressive　(AR)　model,　and　the　coefficients　of　the　AR　model　are　taken　as　feature　vector　for　pattern　recognition　of　the　five　partial　discharge　models　by　BP　neural　network.　The　AR　models　with　different　order　are　adopted　to　fit　the　time　series,　and　higher　recognition　accuracy　of　over　80%　is　achieved　by　the　4th-or　6th-order　AR　model.