Transformer-based Channel Dependency (CD) models generally demonstrate outstanding performance in multivariate time series forecasting (MTSF) tasks. However, CD-based attention mechanisms are often disrupted by weakly correlated sequences when capturing global features and tend to overlook local temporal dependencies. To address these challenges, we propose ECDformer: an Enhanced Channel-Dependent Transformer for multivariate time series forecasting. Specifically, we design an improved Channel-Dependent method (CD+) that transforms multivariate forecasting from a many-to-many approach to multiple many-to-one predictions in the variable dimension. In the temporal dimension, CD+ enhances attention on adjacent time steps, thereby improving the modeling of local features. The framework comprises three key components: a) Relevant Sequence Selection (RSS): Groups variables based on their correlations. b) Local Attention (LA): Extracts local temporal features. c) Multivariate Time Series Forecasting Output (MO): Concatenates the output sequences. Extensive experiments on diverse datasets demonstrate the effectiveness of ECDformer in improving forecasting accuracy.