In　cellular　networks,　the　unbalanced　load　among　base　stations　always　results　in　the　disparity　of　spectrum　efficiency　and　bad　user　experience　in　heavy-loaded　cells.　In　this　paper,　an　intelligent　cell　individual　offset　(CIO)　adjustment　strategy　is　proposed　to　control　the　number　of　users　served　in　the　cell　aiming　of　minimizing　the　variance　of　the　physical　resource　block　(PRB)　usage　ratio　among　the　cells.　Firstly,　a　multi-graph　convolutional　and　gated　recurrent　unit　network　(MGCN-GRU)　is　proposed　to　predict　the　active　users,　where　an　MGCN　is　developed　to　aggregate　spatial　features,　and　a　GRU　is　used　to　capture　temporal　features.　Secondly,　a　multilayer　perceptron　is　used　to　model　the　relationship　between　the　PRB　usage　ratio,　active　users　number　and　cellular　resources.　Finally,　the　optimal　CIO　is　obtained　by　the　genetic　algorithm　to　achieve　load　balancing.　Simulation　results　show　a　good　load　balancing　performance,　and　a　better　prediction　performance　achieved　by　the　proposed　MGCN-GRU.　©　2022　IEEE.