The　holomorphic　embedding　method　(HEM)　is　a　newly　developed　recursive　algorithm,　which　has　the　advantages　of　not　depending　on　initial　value　selection　and　providing　an　analytical　solution.　However,　in　the　actual　power　flow　calculation　of　large　power　grids　with　heavy　loads,　HEM　needs　to　calculate　high-order　power　series　coefficients,　which　is　prone　to　numerical　problems,　resulting　in　convergence　problems　in　power　flow　calculation.　To　improve　the　convergence　characteristics　of　HEM,　this　paper　proposes　a　holomorphic　embedding　power　flow　method　based　on　dynamic　power　restart　(DPRHEM).　This　method　is　based　on　the　initial　value　flexibility　of　the　fast　and　flexible　holomorphic　embedding　power　flow　method　(FFHEM)　and　incorporates　the　proposed　dynamic　update　restart　mechanism　based　on　changes　in　power　residuals　to　dynamically　update　and　transmit　the　＇germ＇　of　the　holomorphic　embedding　power　flow　method.　It　restricts　the　power　series　coefficients　of　the　analytic　function　to　low　orders,　so　that　the　power　state　constantly　approaches　the　target　power　state　and　ensures　stable　and　reliable　convergence　during　power　flow　calculation　in　large　power　grids　with　heavy　loads.　To　a　certain　extent,　this　improves　the　efficiency　of　power　flow　calculation.　Finally,　this　paper　compares　the　performance　of　NRM,　HEM　and　other　restart-based　HEMs　on　the　case9　and　case13659pegase　systems　of　matpower8.0b1,　as　well　as　the　actual　East　China　regional　power　grid.　The　results　show　that　DPRHEM　has　better　convergence　and　higher　computational　efficiency　in　large　power　grids　with　heavy　loads.　©　2025　The　Author(s)