Multi-energy　load　forecasting　plays　a　pivotal　role　in　strategy　management　and　operational　planning　in　integrated　energy　systems　(IESs).　However,　the　inherent　complexity　and　high　dynamicity　of　IESs　bring　significant　challenges,　including　modeling　the　coupling　relationships　between　relevant　features　and　designing　effective　forecasting　models.　The　traditional　statistical　methods　for　feature　correlation　analysis　struggle　to　reveal　the　dynamic　correlation　of　feature　sequences,　and　the　classic　data-driven　machine　learning　methods　for　multi-energy　load　forecasting　suffer　from　a　lack　of　generalization.　To　address　these　issues,　inspired　by　the　successful　applications　of　pre-trained　large　models　in　specific　tasks,　this　paper　introduces　a　large　language　model　(LLM)　into　the　energy　field　and　proposes　a　novel　multi-energy　load　forecasting　model　based　on　GPT2,　called　EnergyGPT,　which　utilizes　the　knowledge　and　capabilities　of　LLMs　trained　on　a　massive　text　stream　corpus　to　enhance　generalization　and　performance　in　multi-energy　load　forecasting.　Additionally,　EnergyGPT　adopts　a　dynamic　self-attention　module　to　adjust　the　importance　of　sequential　features.　The　proposed　EnergyGPT　is　evaluated　on　the　real　load　data　from　the　Tempe　campus　provided　by　Arizona　State　University＇s　online　platform,　compared　with　a　set　of　the　latest　load　forecasting　methods.　The　experimental　results　show　that　EnergyGPT　outperforms　all　the　comparisons　with　the　best　performance　in　four　seasons,　which　validates　the　effectiveness　of　EnergyGPT　and　shows　the　potential　of　large　models　for　cross-domain　applications.　©　2025　Elsevier　Ltd