With　the　development　of　intelligent　transportation　systems,　the　potential　need　for　city-level　traffic　prediction　is　emerging.　However,　some　large-scale　traffic　forecasting　methods　divide　urban　areas　into　grids,　which　cannot　achieve　nodelevel　prediction.　Meanwhile,　mainstream　graph　convolution　models　have　limitations　on　the　scale　of　the　input　graph.　For　this　problem,　we　propose　cluster　graph　convolution　networks　with　spectral　graph　wavelet　kernels.　The　proposed　method　divides　the　large　graph　into　subgraphs　and　then　builds　localized　spatial　dependency　with　spectral　graph　wavelet　convolution.　The　model　could　achieve　node-level　traffic　prediction　on　a　large　graph　with　an　end-to-end　framework.　The　experiments　are　conducted　on　the　historical　dataset　of　Beijing,　covering　more　than　8,000　road　segments.　The　experimental　results　demonstrate　the　effectiveness　of　the　proposed　method.