Although　renewable　energy　such　as　photovoltaic　and　wind　power　can　be　integrated　into　the　grid　to　meet　the　growing　power　demand,　the　unpredictable　changes　of　renewable　energy　and　the　large　number　of　non-linear　loads　resulting　in　complex　power　quality　disturbances　(PQDs).　The　robust　and　accurate　classification　of　complex　PQDs　signals　in　this　situation　is　considered　to　be　challenged.　In　this　paper,　a　novel　multimodal　feature　fusion　method　for　PQDs　classification　method　based　on　the　Stockwell　transform　(ST)　and　deep　learning　(DL)　is　proposed,　which　can　combine　both　the　time　and　frequency　domain　features　of　PQDs　signals.　Specifically,　the　ST　is　performed　on　PQDs　signals　to　obtain　corresponding　ST　contour　images,　which　are　fed　into　a　channel　and　spatial　attention　mechanism　integrated　ResNet18　network　to　extract　frequency　domain　features.　Meanwhile,　a　convolutional　neural　network　(CNN)　is　adopted　to　extract　time　domain　features　of　reshaped　PQDs　signals.　Finally,　the　extracted　time　and　frequency　domain　features　are　fused　to　compensate　the　transform　errors　and　boost　PQDs　recognition　accuracy.　The　simulation　results　demonstrate　the　robustness　and　high　accuracy　of　the　proposed　multimodal　feature　fusion　method　compared　with　benchmark　methods.　©　2023　IEEE.