In　order　to　improve　the　classification　accuracy　of　complex　power　quality　disturbances　(PQDs),　this　paper　proposes　a　PQDs　classification　method　based　on　the　optimized　parallel　model　of　features　merging.　This　method　uses　fully　convolutional　networks　(FCN)　and　long　short-term　memory　(LSTM)　to　mine　the　high-dimensional　spatial　and　temporal　features　of　PQDs　in　a　feature　merging　manner.　The　global　max　pooling　(GMP)　and　time　series　reshape　(TSR)　optimization　methods　are　used　to　improve　the　classification　performance　of　the　model.　In　order　to　verify　the　effectiveness　of　the　proposed　method,　this　paper　builds　a　classification　model　based　on　the　Keras　framework,　establishes　a　PQDs　database　with　72　types　of　disturbances,　and　conducts　simulation　experiments.　The　proposed　method　has　an　average　classification　accuracy　of　92.38%　in　a　20dB　white　noise　environment,　which　has　higher　noise　robustness　and　classification　accuracy　than　other　mainstream　deep　learning　classification　methods.　In　addition,　the　classification　test　is　carried　out　on　10　types　of　PQDs　sampled　by　the　hardware　platform,　and　a　total　of　100　groups　of　signals　are　all　correctly　classified,　which　further　verifies　the　reliability　of　the　proposed　method.　©2023　Chin.Soc.for　Elec.Eng.