Balanced　in-core　power　levels　in　nuclear　power　plants　(NPPs)　are　critical　for　safety,　whereas　power　tilt　disrupts　this　balance,　reducing　safety　margins　and　posing　risks.　Early　warning　for　power　tilt　offers　an　effective　way　of　optimizing　monitoring.　Due　to　abnormal-sample　scarcity　and　security　concerns,　common　data-driven　models　train　on　the　simulated　data　generated　by　simulators.　However,　achieving　a　satisfactory　effect　in　practices　is　difficult　because　simulators　imperfectly　emulate　reality.　Thus,　we　propose　a　power　tilt-oriented　early　warning　method　called　simulation–reality　spatial–temporal　model　(SR-STM).　Motivated　by　the　physical　model　in　NPPs,　a　knowledge-guided　hierarchical　graph　is　designed　to　characterize　spatial　correlations　among　local　power　levels　for　SR-STM’s　input.　The　SR-STM　uses　a　lightweight　spatial–temporal　network　(LST-Net)　as　a　feature　extractor,　balancing　precision,　and　efficiency.　To　bridge　sim-real　interdomain　discrepancies,　SR-STM　utilizes　node-alignment　adversarial　learning　(NAAL)　for　fine　weight　tuning　in　subdomain,　and　eigenvalue-based　scale　alignment　(ESA)　for　sim-real　feature　proximity.　Forecasting　local　power　levels　using　the　SR-STM,　dynamic　metrics　and　alarm　limits　are　calculated　and　compared　to　perform　the　early　warning　task.　The　online　experimental　prototype　verifies　that　SR-STM　surpasses　various　state-of-the-art　methods　in　terms　of　early　warning　and　sim-real　cross-domain　tasks.　©　2014　IEEE.