Abstract: Addressing the challenges of high computational complexity and sensitivity to noise in complex nonlinear models for time series forecasting tasks, this paper proposes a novel forecasting model combining linear models, Fourier decomposition, and a Temporal Convolutional Network (TCN). The linear model is employed for its simplicity and efficiency in trend modeling, demonstrating performance comparable to more complex models; Fourier decomposition leverages its advantages in extracting trend and seasonal components from time series data to accurately capture periodic patterns; and the TCN efficiently models long-term dependencies through causal and dilated convolutions, possessing parallel computing capabilities superior to models like Transformer. Furthermore, this study proposes an adaptive dynamic weighted loss function, combining MSE and MAE, to balance prediction accuracy and stability through a dynamic weighting strategy. Experiments on multiple benchmark datasets, covering multivariate-to-univariate, univariate-to-univariate, and multivariate-to-multivariate forecasting scenarios, validate the model's performance. The results show that the model outperforms current state-of-the-art methods across different prediction lengths, demonstrating its superiority and generalization ability.