This paper proposes a deep learning-based method to accelerate the traditional simulation process of the thermal field in switchgear cabinet. We use finite element method (FEM) simulations to generate detailed thermal field distribution data and construct a dataset that captures the relationship between input current and high-dimensional temperature distributions. To handle the large-scale data efficiently and enhance the model’s computational performance, the proper orthogonal decomposition (POD) technique is applied for dimensionality reduction. The reduced-order data is then used to train a deep learning model that simulates the complex mapping between current and temperature. This approach achieves model order reduction and is validated by comparison with traditional FEM simulations, demonstrating significant improvements in efficiency without sacrificing accuracy.