Currently, most crowd counting methods have outstanding performance under normal weather conditions. However, they often struggle to maintain their performance in extreme and adverse weather conditions due to significant differences in the domain and a lack of adverse weather images for training. To address this issue and enhance the model's robustness in adverse weather, we propose a two-stage crowd counting method. Specifically, in the first stage, we introduce a multi-queue MoCo contrastive learning strategy to tackle the problem of weather class imbalance. This strategy facilitates the learning of weather-aware representations by the model. In the second stage, we propose to refine the representations under the guidance of contrastive learning, enabling the conversion of the weather-aware representations to the normal weather domain. While significantly improving the robustness, our method only marginally increases the weight of the model. In addition, we also create a new synthetic adverse weather dataset. Extensive experimental results show that our method achieves competitive performance.