With　the　rapid　development　of　smart　grids,　accurate　load　prediction　is　essential　for　stable　operation　and　optimal　scheduling.　This　paper　addresses　errors,　missing　values,　and　anomalies　in　electrical　load　data　using　a　conditional　generative　adversarial　network　(CGAN)　with　dual　self-attention　(SELF)　for　data　reconstruction.　The　model　simplifies　time-series　complexity　and　historical　load　patterns,　eliminating　the　need　for　intricate　spatiotemporal　modeling.　Based　on　the　reconstructed　data,　a　short-term　load　forecasting　method　is　proposed　using　a　bidirectional　temporal　convolutional　network　(BiTCN),　bidirectional　gated　recurrent　unit　(BiGRU),　and　self-attention.　This　model　processes　forward　and　backward　time-series　information　in　parallel,　extracting　multi-scale　features　for　more　accurate　predictions.　In　order　to　accurately　describe　the　response　behavior　of　users　under　different　electricity　price　differentials,　a　logistic　demand　response　(DR)　model　considering　time　lag　factors　is　introduced.　The　model　defines　optimistic　and　pessimistic　response　curves,　effectively　reflecting　the　actual　range　of　user　responses　to　price　incentives,　thus　enhancing　the　practicality　of　load　forecasting　in　decision　support.　Experimental　results　demonstrate　that　the　proposed　method　not　only　enhances　the　accuracy　and　stability　of　load　forecasting　but　also　provides　robust　technical　support　for　the　stable　operation　of　smart　grids.　©　2025