Time series forecasting (TSF) is pivotal across numerous domains Many popular technologies have been adopted to improve such time series forecasting, such as input patching and heterogeneity modeling. However, effective forecasting still faces two core challenges: potential semantic break by standard patching strategies, and the pattern drift from inherent non-stationarity. Prevailing works, such as Transformer-based or static CNNs models, often exhibit limitations in computational efficiency or adaptation to time-varying dynamics. To overcome these problems, we introduce a novel pure convolutional architecture: the Context-aware Convolutional Network with Dynamic Weight (CDConvNet). In CDConvNet, two key mechanisms are proposed to improve predictive performance. First, a context-aware patching mechanism utilizes global temporal context to guide the generation of adaptive patch boundaries, effectively tackling the semantic break problem. Second, a convolutional encoder augmented by a dynamic weight generator that modulates channel features. It significantly boosts the model’s capacity to adapt to various patterns in time series. Comprehensive experiments on nine widely used datasets demonstrate that CDConvNet substantially outperforms existing state-of-the-art(SOTA) models in 81.94%, exhibiting strong performance particularly in long-term time series forecasting tasks. Our work proves the potential capacity in pure convolutional architectures for complex TSF.