Deep　neural　network　has　been　widely　applied　for　seizure　prediction,　and　its　appropriate　input　is　the　key　to　prediction　effect.　However,　the　phase　synchronization　which　may　happen　before　seizure　onset　has　seldom　been　received　sufficient　attention.　In　this　paper,　a　novel　seizure　prediction　method　is　proposed.　The　multi-channel　electroencephalogram　(EEG)　signals　are　preprocessed　for　denoising,　and　wavelet　transform　(WT)　is　employed　to　extract　time-frequency　features.　Then,　the　phase　locking　value　(PLV)　of　each　channel　of　filtered　EEG　is　calculated　based　on　two　frequency　bands　(i.e.　2-40　Hz　and　80-128　Hz),　and　it　is　used　for　weighting　time-frequency　features,　obtaining　the　fused　features.　Finally,　a　three-dimensional　convolution　neural　network　(3DCNN)　is　constructed　and　the　fused　features　are　used　as　input　to　achieve　epilepsy　prediction.　Experiments　are　conducted　based　on　the　CHB-MIT　dataset,　and　the　average　prediction　accuracy　of　10　subjects　reaches　99.14%.　The　results　show　that　the　fusion　of　PLV　and　time-frequency　features　can　reflect　the　characteristics　of　Epilepsy,　and　it　matches　with　3DCNN　well.　©　2023　IEEE.