Based　on　the　characteristic　analysis　of　the　high　power　amplifier　(HPA)　in　wide-band　CMMB　repeater　stations,　a　novel　neural　network　was　proposed　which　can　respectively　process　the　memory　effect　and　the　nonlinear　of　power　amplifier.　The　novel　model　based　on　real-valued　time-delay　neural　networks(RVTDNN)　uses　the　Levenberg-Marquardt　(LM)　optimization　to　iteratively　update　the　coefficients　of　the　neural　network.　Due　to　the　new　parameters　w0　in　the　novel　NN　model,　the　modified　formulas　of　LM　algorithm　were　provided.　Next,　in　order　to　eliminate　the　over-fitting　of　LM　algorithm,　the　Bayesian　regularization　algorithm　was　applied　to　the　predistortion　system.　Additionally,　the　predistorter　of　CMMB　repeater　stations　based　on　the　indirect　learning　method　was　constructed　to　simulate　the　nonlinearity　and　memory　effect　of　HPA.　Simulation　results　show　that　both　the　NN　models　can　improve　system　performance　and　reduce　ACEPR　(adjacent　channel　error　power　ratio)　by　about　30　dB.　Moreover,　with　the　mean　square　error　less　than　10-6,　the　coefficient　of　network　for　FIR-NLNNN　is　about　half　of　that　for　RVTDNN.　Similarly,　the　times　of　multiplication　and　addition　in　the　iterative　process　of　FIR-NLNNN　are　about　25%　of　that　for　RVTDNN.