Time Series Prediction Model Combining Temporal Convolutional Network and Fourier Decomposition

To tackle the high computational complexity and noise sensitivity of complex nonlinear models for time series forecasting, this paper proposes FDTCNLinear, a novel model integrating Temporal Convolutional Network (TCN) and Fourier decomposition. The model first leverages Fourier transform to decompose time-series signals into long-term trend and seasonal-cyclic components. It then adopts a linear layer for efficient extrapolation of long-term trends, and employs a multi-scale TCN to mine long- and short-term temporal dependencies hidden within seasonal components. In addition, this paper designs an adaptive Dynamic Weighted Hybrid Loss (DWHL) function. By dynamically balancing the weights of Mean Squared Error (MSE) and Mean Absolute Error (MAE), this loss alleviates disturbances caused by outliers. Experiments on multiple public benchmark datasets verify that the proposed model achieves remarkably higher prediction accuracy than state-of-the-art Transformer-based models and linear baselines across all prediction horizons.