Correlated Time Series (CTS) refer to a set of time series that are recorded to monitor entities that interact with each other in social and production systems. Correlated Time Series Forecasting (CTSF) plays an important role in various applications by enabling people to make informed and effective decisions. Accurate forecasting of CTS relies significantly on modeling appropriate correlations among channels. However, the correlations among channels are constantly changing over time due to the complexity and uncertainty of the external environment. Previous works typically introduce relevant channels in a relatively fixed way for feature representation, which is unable to effectively capture and learn the dynamically changing correlations among channels, thus resulting in suboptimal prediction performance. To this end, we propose CEDMix, a Contrast-Enhanced Dynamic channel Mixing time series forecaster, which adopts an encoder-only architecture. First, we design a Dynamic Channel Mixing module (DCM) to capture the time-varying correlations. During different time windows, DCM introduces present highly relevant channels into a customized 2D form for each channel, and adopts a parameter-efficient inception block to extract correlations by time steps. Second, to further enhance DCM, we propose dynamic channel-wise contrasting, which can discriminate the constantly changing correlations between the relevant and irrelevant channels. Extensive experiments conducted on seven datasets from diverse domains demonstrate that, compared with recent state-of-the-art models, CEDMix achieves an average reduction of 16.44% in MAE and 16.98% in MSE.