Data-driven　methods　have　been　widely　used　in　electricity　load　forecasting　to　improve　forecasting　accuracy.　However,　when　the　electricity-selling　company　accesses　new　customers,　the　applicability　of　conventional　data-driven　methods　is　somewhat　limited　due　to　the　lack　of　customers＇　historical　electricity　consumption　data.　To　solve　this　problem,　the　paper　proposes　a　short-term　power　load　forecasting　method　based　on　Domain　adversarial　Transfer　Network　(DATN).　The　model　utilizes　the　Transformer　model　as　a　feature　extractor　to　capture　dynamic　features　and　time　dependencies　in　the　load　data.　Subsequently,　the　load　forecaster　accurately　predicts　the　future　load　profile　based　on　these　features.　Adversarial　learning　through　the　domain　discriminator　and　feature　extractor　ensures　that　the　model　learns　deep　domain　invariant　features　while　combining　Multi-Kernel　Maximum　Mean　Discrepancy　(MK-MMD)　and　Correlation　Alignment　(CORAL)　to　reduce　further　the　distribution　difference　between　source　and　target　domain　data.　The　model　is　validated　on　the　electricity　consumption　data　of　industrial　users　in　a　southern　province,　and　the　experimental　results　show　that　the　method　has　good　prediction　accuracy　and　adaptability　in　small　sample　scenarios.　©　2025　Power　System　Technology　Press.　All　rights　reserved.