As　nuclear　power　plants　(NPPs)　undertake　more　peak　regulation　tasks　to　handle　high　new　energy　penetration　and　overcapacity,　precise　forecasting　of　in-core　power　distributions　is　essential　for　optimal　control　and　safe　operation.　However,　current　works　lack　an　effective　strategy　for　predicting　high-resolution　power　distributions　and　neglect　in-core　spatial　correlations.　This　study　proposes　a　spatial-temporal　hierarchical-directed　network　(ST-HDN)　for　forecasting　power　distributions,　whose　prediction　strategy　is　guided　by　the　physical　model.　To　characterize　spatial　correlations　and　causal　relationships　among　physical　quantities,　the　hierarchical-directed　graph　is　designed　and　combined　with　neutron　and　power　signals　for　input　to　the　ST-HDN.　Concretely,　the　ST-HDN　integrates　three　sub-modules:　a　temporal-differencing　layer　to　enhance　representation　of　subtle　variations;　a　multi-dilated　convolutional　network　to　extract　dynamic　temporal　features;　and　a　graph　convolutional　network　to　propagate　spatial　adjacent　information,　further　predicting　power　nodes　at　various　positions.　The　predicted　power　nodes　are　post-processed　to　derive　future　power　distributions.　Experiments　on　two　peak　regulation　scenarios　from　a　real-world　NPP　illustrate　that　the　ST-HDN　outperforms　various　state-of-the-art　methods　in　10-,　20-,　and　30-min　ahead　forecasting.