The　decision　-making　process　for　pavement　maintenance　from　a　scientific　perspective　is　based　on　accurate　predictions　of　pavement　performance.　To　improve　the　rationality　of　pavement　performance　indicators,　comprehensive　consideration　of　various　influencing　factors　is　necessary.　To　this　end,　four　typical　pavement　performance　indicators　(i.e.,　Rutting　Depth,　International　Roughness　Index,　Longitudinal　Cracking,　and　Alligator　Cracking)　were　predicted　using　the　Long　Term　Pavement　Performance　(LTPP)　database.　Two　types　of　data,　i.e.,　local　input　variables　and　global　input　variables,　were　selected,　and　S　-ANN　and　L　-ANN　models　were　constructed　using　a　fully　connected　neural　network.　A　comparative　analysis　of　the　predictive　outcomes　reveals　the　superior　optimization　of　the　L　-ANN　model.　Subsequently,　by　incorporating　structures　such　as　self　-attention　mechanism,　a　novel　predictive　approach　based　on　the　Transformer　architecture　was　proposed.　The　objective　is　to　devise　a　more　accurate　predictive　methodology　for　pavement　performance　indices,　with　the　goal　of　guiding　pavement　maintenance　and　management　efforts.　Experimental　results　indicate　that,　through　comparative　analysis　of　three　quantitative　evaluation　metrics　(root　mean　square　error,　mean　absolute　error,　coefficient　of　determination),　along　with　visual　scatter　plots,　the　predictive　model　employing　the　fused　Transformer　architecture　demonstrates　higher　robustness　and　accuracy　within　the　domain　of　pavement　performance　prediction　when　compared　to　the　L　-ANN　model.　This　outcome　substantiates　the　efficacy　and　superiority　of　the　model　in　terms　of　predictive　performance,　establishing　it　as　a　reliable　tool　for　accurately　reflecting　the　evolution　of　asphalt　pavement　performance.　Furthermore,　it　furnishes　a　theoretical　reference　for　determining　optimal　preventive　maintenance　timing　for　pavements.