Facial expression recognition (FER) has wide applications in various domains such as healthcare, human-computer interaction, and more. However, the performance of existing FER algorithms is often compromised in low-light environments due to inconsistent data distribution between low-light and normal-light expression images. To address this challenge, we propose a supervised domain adaptation method with alignment regularization (ARSDA-FER). Our approach incorporates a multi-stage feature alignment module to maintain consistency between the source and target domains, reducing their distribution differences. In addition, we use the prediction alignment module to constrain and refine the results, mitigating the bias in cross-domain prediction through embedding-level optimization. Furthermore, a class-aware alignment module is employed to minimize the spatial distance between class centers of the same class across source and target domains. Our method achieves an accuracy of 86.96% on the RAF-DB and 83.33% on the FERplus. Extensive experiments demonstrate that our proposed low-light FER method has superior performance, with a 1.7% improvement compared to state-of-the-art methods, revealing its applicability and value in real-world scenarios.