The　established　reference　spectrum　(ASTM　AM1.5)　based　on　mid-latitudes　is　still　the　most　frequently　used　standard　in　design　and　simulation　in　various　fields　worldwide,　however,　the　solar　spectrum　varies　with　climate　and　time,　which　significantly　affects　the　evaluations　of　solar　energy　utilization.　This　study　proposes　a　novel　reference　spectrum　analysis　method　to　establish　several　local　reference　spectra,　in　this　case,　based　on　147,374　spectra　in　Beijing,　which　can　characterize　solar　resources　throughout　the　year.　The　dataset　consists　of　spectra　of　direct　normal　irradiance,　and　a　deep　learning　autoencoder　is　adopted　to　extract　the　intrinsic　features　of　each　spectrum　for　clustering.　10　clusters　are　finally　divided　by　agglomerative　hierarchical　clustering,　and　10　local　reference　spectra　were　identified　from　each　cluster　by　correlation　analysis.　The　results　show　that　the　shape　of　new　reference　spectra　shifts　to　the　short　wave　as　the　irradiance　increases,　and　the　average　photon　energy　increases　from　1.069　eV　to　1.456　eV.　The　new　local　reference　spectra　are　compared　with　the　standard　spectrum　AM1.5D　to　show　the　influence　of　shape　variations,　and　it　reveals　61%　maximum　differences　among　photovoltaic　applications.　The　superiority　verified　new　reference　spectra　could　be　further　used　to　optimize　local　solar　technologies　design　and　guide　photovoltaic　material　development.