In this paper, we propose a method for creating a reduced-order model (ROM) for temperature prediction that can be applied to nonlinear time-variant systems. Our nonlinear time-variant ROM is an extension of sparse identification of nonlinear dynamical systems, first proposed in 2016. Three machine learning methods are developed for automatically deriving a thermal network model from time series data. Link relationships between temperature nodes and parameter settings, such as thermal resistance and heat capacity, are automatically inferred by machine learning. Because the proposed model targets phenomena that can be represented as a thermal network model, it can deal with forced convection with velocity as a variable. The model can also handle natural convection and radiation, where thermal resistance is a function of temperature. The model can be created using a very small amount of time series data. Unlike black-box AI, the proposed model is easy to interpret; in other words, it enables us to understand the heat transfer process. Furthermore, because it is based on a physical model, it has high prediction accuracy when extrapolated. The effectiveness of the proposed method is demonstrated using the results of thermo-fluid analysis for an air-cooled power module. In addition, because the proposed method can be created using a very small amount of time series data, has high extrapolation accuracy, and is easy to interpret, it is expected not only that design parameters can be fine-tuned and actual loads can be taken into account, but also that condition-based maintenance can be realized through real-time simulation.