To ensure the generalization performance of machine learning (ML) models

the statistical learning theory typically assumes that data are independent and identically distributed (i.i.d.). However

due to the openness of the power systems and the dynamic evolution of actual data

data distribution shifts may be unavoidable

resulting in significantly performance degradationwhen well-trained models are deployed in practical applications. This paper tackles the problems of ML in scenarios where the training and test dataset are not i.i.d.

known as out-of-distribution generalization (OODG). First

the analysis framework of the OODG problem is established for power systems applications. Then

through the combination of theoretical analysis and case studies

the quantitative methods for assessing data distribution shifts

strategies for improving and evaluating the MLs’ OODG performance

are explored. By deepening the understanding and research of OODG

the viability and application values of ML can be further enhanced.