Abstract: In recent years, with the rapid development of renewable energy and the accelerated promotion of new power system construction, the uncertainty of power systems has become more prominent, posing huge challenges for modeling and optimization scheduling. Machine learning techniques can effectively utilize vast historical data to provide new theoretical basis for optimizing stable and fast solutions. This paper provides a detailed analysis of the progress in this emerging interdisciplinary field. First, for general optimization problems, based on the interaction between machine learning and optimization computing, the basic algorithm framework is summarized into three categories: machine learning end-to-end optimization solving, machine learning enhanced optimization solving algorithms, and machine learning and power system optimization joint driving solving. Their basic principles and applicable problem forms are explained respectively. Then, the progress of related technology applications in power system optimization is reviewed and the basic methods and application effects are summarized. Finally, the development trends of learning-based optimization methods and their application prospects in new power systems are explored, with the aim of providing references and inspirations for future research work in this emerging field.