The high volatility and nonlinearity of new power systems pose significant challenges to the accuracy and adaptability of mechanism-based analysis and optimization control methods. Data-driven machine learning methods have partially addressed these challenges but still suffer from weaknesses such as limited interpretability and reliability. Large models

which have shown preliminary capabilities in learning physical laws

offer a new pathway to tackling the complexities of power systems. Given the unique characteristics of power systems

such as spatiotemporal complexity and well-established mechanistic knowledge

it is imperative to explore the construction methodologies for domain-specific large models tailored to power systems. This paper first summarizes the development and capabilities of large model technologies

followed by an analysis of the core requirements for building specialized large models for power systems. Based on this

the key challenges in constructing large scientific models for power systems are identified. Finally

the paper discusses the development trends and prospects of scientific large models for power systems. Through this research

we aim to explore the fundamental requirements and construction methodologies for power system-specific large models

enabling them to extract underlying physical knowledge from massive data and establish a new paradigm for the intelligent operation of power systems.