TIMixer: 장기 시계열 예측을 위한 경량 CNN-MLP 하이브리드 모델

초록

Long-term time series forecasting (LSTF) is essential in various domains, including energy, transportation, and weather. While transformer-based models have shown strong performance, their high computational complexity and large parameter sizes limit their use in real-time and resource-constrained environments. In this paper, we propose TIMixer, a CNN-MLP hybrid architecture. TIMixer substitutes multi-head attention with a single linear layer and utilizes depth-wise separable convolution to effectively capture both global and local temporal patterns. Under identical settings, including input length (512), prediction horizons (96–720), and training conditions, we evaluate TIMixer on six benchmark datasets (ETT, Weather, and Electricity). Experimental results demonstrate that TIMixer achieves comparable or superior MSE/MAE while employing approximately one-fourth the parameters and one-sixth the computational cost of PatchTST. Additionally, TIMixer reduces CPU inference time by 36% compared to TSMixer-K on the Weather dataset.

키워드