Abstract: Data-driven has become the core paradigm for the construction and digital transformation of new power systems, and related algorithms have shown superior engineering effects and application potential in multiple power system fields, such as load forecasting, condition-based maintenance, and multi-agent scheduling. However, the actual of engineering data often faces problems such as insufficient and imbalanced samples, which restricts the ultimate effectiveness of data-driven algorithms. Therefore, the few-shot learning is needed to address this challenge. This paper explores the few-shot learning technologies from three levels of data, features, and models. It reviews and analyzes the current application status of related technologies in scenario generation, fault diagnosis, and transient stability assessment of power systems. The shortcomings and challenges faced by fewshot learning technologies in new power systems are further presented.