Driven by the “dual carbon” goals and energy security imperatives

China is accelerating the construction of the new power system. This transition urgently necessitates the integration of intelligent methodologies to enhance system flexibility

security

and economic efficiency. The AI revolution

exemplified by foundation models

holds immense potential to revolutionize traditional analysis and decision-making paradigms. However

the core scientific problems in applying foundation models to the power domain require systematic investigation. Also

application paradigms remain to be clarified

and the overall research framework has yet to be established. Thus

this paper first elucidates the necessity of introducing foundation model technologies into the new-type power systems and systematically analyzes two key scientific problems. Then

from the perspectives of general large models and domain-specific foundation models

it explores the overall architecture and prospective application paradigms in power systems. Based on these technological pathways

an evolutionary road-map is proposed to advance from single-model intelligence toward multi-agent collaborative intelligence. Furthermore

a suite of domain- specific base models is established to support core power-system functions

including forecasting

operational planning

stability analysis and control

and panoramic simulation. On this basis

the paper analyzes the major challenges confronting large-model research for power systems and proposes a prospective research framework and key technological directions for future development.