Derived　from　an　effective　strategy　-　direct　and　multiple-input　multiple-output　strategy,　a　modular　neural　network　based　on　a　bi-level　particle　swarm　optimization　algorithm　(BLPSO-MNN)　is　proposed　in　the　present　study　to　improve　the　accuracy　for　multi-step　time　series　prediction.　While　a　binary　particle　swarm　optimization　algorithm　is　designed　for　the　external　layer　to　optimize　the　task　division　of　prediction　horizons,　a　multi-objective　particle　swarm　optimization　algorithm　is　designed　for　the　internal　layer　to　trade　off　between　the　prediction　accuracy　and　structural　complexity　for　each　subnetwork　in　modular　neural　network.　Besides,　a　set　of　fuzzy　If-Then　rules　is　proposed　to　determine　the　historical　information　to　be　input　to　subnetworks.　Thus,　the　structure　of　BLPSO-MNN,　including　the　number　of　modules　as　well　as　the　subnetwork　structure,　is　self-determined　accordingly.　Numerous　experiments　are　conducted　for　18-step-ahead　time　series　prediction　to　evaluate　the　performance　of　BLPSO-MNN.　Experimental　results　show　that,　although　the　prediction　accuracy　decreases　when　the　prediction　horizon　is　large,　the　overall　performance　of　BLPSO-MNN　is　superior　over　all　comparative　models　with　greater　improvement　for　larger　horizons,　indicating　it　is　suitable　for　a　long-term　prediction.　Besides,　the　set　of　fuzzy　rules　balances　the　prediction　accuracy　against　the　structural　complexity　caused　by　the　subnetwork　inputs.