This　paper　proposes　an　interpretable　deep　learning　method　based　on　generative　data　augmentation　for　forecasting　carbon　allowance　prices　in　the　EU　Emissions　Trading　System　(ETS)　Phase　4.　Utilizing　TimeGAN,　we　generate　near-real　expanded　data　to　enhance　the　training　sets.　Temporal　Fusion　Transformer　(TFT)　is　used　to　quantify　the　contribution　of　impact　factors.　The　results　show　that　the　augmentation　effectively　improved　the　prediction　accuracy.　Interpretability　analysis　reveals　that　Brent　crude　oil,　NBP　natural　gas,　and　Rotterdam　coal　are　the　top　three　contributors.　Our　findings　offer　a　strong　approach　for　the　new　phase　price　forecasting,　helping　market　participants　and　policymakers　in　informed　decision-making.　©　2024