Deep learning-based models have achieved remarkable success in time series forecasting. However, they still exhibit significant shortcomings when handling complex time series data, especially in modeling seasonal and trend components. For instance, such models often fail to adequately capture the nonlinear variations in trend components. To address these challenges, this paper proposes a model based on variable decomposition and convolutional attention (VDCA). The VDCA comprises modules for variable decomposition, component modeling, and residual learning. The variable decomposition module utilizes convolutional operations to accurately extract trend and seasonal components. During the component modeling phase, linear transformations are applied to the seasonal components, while a convolutional attention mechanism is employed to extract features from the trend components. Experimental results show that VDCA achieved the best performance in 72 out of 88 experiments conducted on seven publicly available datasets, demonstrating exceptional generalization ability and forecasting accuracy. This provides a more effective new approach.