The website fingerprinting (WF) classification focuses on identifying websites accessed by clients over encrypted network tools. By leveraging machine learning technology, methods have demonstrated high accuracy. However, many existing methods heavily rely on specific feature distributions, leading to poor performance on unseen data or data affected by concept drift. In this paper, we design a representation framework based on multi-temporal domain feature adaptation for website fingerprinting identifications, which consists of pre-training and fine-tuning stages. The framework extracts features from multi-temporal traffic data to generate robust representations and requires only a small amount of data for fine-tuning, making it applicable to data with different feature distributions. Experiments demonstrate that this approach effectively addresses concept drift.