Time-series forecasting plays a pivotal role in many areas. In recent years, Transformer based on Channel-Independent(CI) and Channel-Dependent(CD) strategy have achieved remarkable progress on this task. However, each of these approaches has its own limitations: CI strategy, while enhancing model robustness, neglects the intrinsic relationships among variables; CD strategy, although capturing inter-variable interactions, simultaneously introduces redundant information from unrelated channels, thereby degrading predictive performance. To address this issue, we introduce FECCD. This approach follows a two-stage modeling strategy: CI first, then channel correlation discovery. First, we propose a novel phase-normalized frequency attention to capture independent information from each variable, effectively eliminating cross-variable interference. Subsequently, we introduce a mask-based channel correlation discovery mechanism. It generates dynamic masks from the attention scores computed in the first stage to actively select effective inter-variable interactions and suppress interference from irrelevant channels. Experimental results demonstrate that the proposed FECCD approach significantly outperforms than state-of-the-art(SOTA) in predictive accuracy, exhibiting strong modeling capabilities in the domain of time-series forecasting.