Because　the　prediction　of　thermal　comfort　is　a　complex,　nonlinear　process,　which　is　not　convenient　for　the　application　of　real-time　control　of　air　conditioning,　neural　network　can　quickly　establish　the　nonlinear　mapping　between　the　input　and　output　of　PMV　index,　which　becomes　a　new　research　direction　of　PMV　index　solution.　PSO　is　used　to　optimize　GRNN　parameters　and　build　a　prediction　model,　which　can　find　the　optimal　smooth　factor　of　GRNN　network　in　the　shortest　time,　so　that　the　trained　network　can　meet　the　requirements　of　both　learning　ability　and　learning　speed.　The　results　show　that　the　generalized　regression　neural　network　inherits　all　the　advantages　of　RBF　network,　and　is　superior　in　approximation　ability　and　learning　speed.　From　the　experiment,　when　the　SPREAD　value　is　the　same,　the　prediction　accuracy　of　GRNN　network　is　slightly　better　than　that　of　RBF　network.　The　prediction　accuracy　of　GRNN　network　can　also　be　improved　obviously　after　PSO　optimization.　©　2022　IEEE.