A network manager could infer the website which the client is visiting from encrypted website traffic via website fingerprinting (WF). Existing WF methods primarily rely on designing classifiers that distinguish distinct traffic features from different websites. However, the website content updates and background noise make it challenging to identify the website correctly in real-world environment. Our study exploits the fact that although website content changes dynamically, the layout of website and the types of transmission resources remain stable. Building on this insight, we propose a novel multi-level flow representation enhanced WF, namely MFRWF, to obtain stable website features that are affected less by the dynamic network environment. We describe the most representative features of the interaction about resource request between clients and servers from different level. Furthermore, we improve the attention mechanism to capture the stable website layout pattern through the most relevant flow-level relationship among website flows. To evaluate the performance of MFRWF, we construct a real-world website traffic dataset and conduct comprehensive experiments. The evaluations demonstrate that MFRWF not only outperforms state-of-the-art WF methods with the best accuracy in ideal closed-world scenario but also exhibits strong robustness against website content updates and background noise in the real-world scenario.