Rapid　and　accurate　prediction　of　peak　ground　acceleration　(PGA)　is　an　important　basis　for　determining　seismic　damage　through　on-site　earthquake　early　warning　(EEW).　The　current　on-site　EEW　uses　the　feature　parameters　of　the　first　arrival　P-wave　to　predict　PGA,　but　the　selection　of　these　feature　parameters　is　limited　by　human　experience,　which　limits　the　accuracy　and　timeliness　of　predicting　peak　ground　acceleration　(PGA).　Therefore,　an　end-to-end　deep　learning　model　is　proposed　for　predicting　PGA　(DLPGA)　based　on　convolutional　neural　networks　(CNNs).　In　DLPGA,　the　vertical　initial　arrival　3-6　s　seismic　wave　from　a　single　station　is　used　as　input,　and　PGA　is　used　as　output.　Features　are　automatically　extracted　through　a　multilayer　CNN　to　achieve　rapid　PGA　prediction.　The　DLPGA　is　trained,　verified,　and　tested　using　Japanese　seismic　records.　It　is　shown　that　compared　to　the　widely　used　peak　displacement　(Pd)　method,　the　correlation　coefficient　of　DLPGA　for　predicting　PGA　has　increased　by　12-23%,　the　standard　deviation　of　error　has　decreased　by　22-25%,　and　the　error　mean　has　decreased　by　6.92-19.66%　with　the　initial　3-6　s　seismic　waves.　In　particular,　the　accuracy　of　DLPGA　for　predicting　PGA　with　the　initial　3　s　seismic　wave　is　better　than　that　of　Pd　for　predicting　PGA　with　the　initial　6　s　seismic　wave.　In　addition,　using　the　generalization　test　of　Chilean　seismic　records,　it　is　found　that　DLPGA　has　better　generalization　ability　than　Pd,　and　the　accuracy　of　distinguishing　ground　motion　destructiveness　is　improved　by　35-150%.　These　results　confirm　that　DLPGA　has　significant　accuracy　and　timeliness　advantages　over　artificially　defined　feature　parameters　in　predicting　PGA,　which　can　greatly　improve　the　effect　of　on-site　EEW　in　judging　the　destructiveness　of　ground　motion.