基於適應性多尺度時頻域多層感知器於時間序列預測之架構設計

Abstract

準確預測多變量時間序列對於工業應用至關重要，然而，由於數據中固有的短期瞬態變化、長期相依性、隱藏的週期性，以及嚴格的計算效率要求，這項任務充滿挑戰。為了解決這些問題，我們提出了AMTF-MLP，一個創新的純多層感知器(MLP)架構，它透過自適應融合機制，整合了多尺度時域混合器與頻域頻譜學習。AMTF-MLP透過並行分支處理信號：一個帶有層級化區塊混合器的時域分支，用以捕捉局部與全域的時間模式；以及一個帶有頻譜MLP的頻域分支，用以建模週期性。 在多樣化的公開基準上進行的大量實驗，驗證了我們模型的有效性與通用性。在長期預測方面，AMTF-MLP展現了高度的競爭力，與iTransformer和PatchTST等主流Transformer模型相比，其均方誤差(MSE)降低了9.3%至20.4%。在高頻率的短期預測場景中，它取得了優異的成果，在PEMS數據集上，其MSE分別比AMD和iTransformer等強力競爭對手降低了24.3%和40.4%。 此模型優異的跨域性能，同時也體現在計算效率上。在其高效能的MLP同類模型中，AMTF-MLP展現了優異的記憶體用量，記憶體消耗比AMD少1.80倍，同時維持著快1.5倍的訓練速度。消融實驗證實，我們設計的每個組件都對模型 的效能至關重要。憑藉其線性複雜度與強大的實證結果，AMTF-MLP為真實世界的預測系統提供了一個強大且實用的解決方案。

Accurately predicting multivariate time series is essential for industrial applications, yet it poses significant challenges due to short-term transients, long-term dependencies, and hidden periodicities, alongside stringent computational efficiency requirements.To address these issues, we introduce AMTF-MLP, an innovative pure Multi-Layer Perceptron (MLP) architecture that integrates multi-scale time-domain mixers and frequency domain spectral learning, unified through adaptive fusion. AMTF-MLP processes the signal in parallel branches: a time-domain branch with hierarchical patch mixers to capture local and global temporal patterns, and a frequency-domain branch with a spectral MLP to model periodicities. Extensive experiments on diverse public benchmarks validate our model's effectiveness and versatility. In long-term forecasting, it delivers highly competitive performance, reducing Mean Squared Error (MSE) by 9.3% to 20.4% compared to prominent Transformer models like iTransformer and PatchTST. In high-frequency short-term scenarios, it achieves leading results, reducing MSE by 24.3% and 40.4% against strong competitors like AMD and iTransformer, respectively, on the PEMS datasets. The model's excellent cross-domain performance is also reflected in its computational efficiency: among its high-performance MLP peers, it exhibits the most superior memory efficiency, consuming 1.8 × less memory than AMD, while maintaining a training speed 1.5 × faster. Ablation studies confirm that each component of our design is critical to the model’s efficacy. With its linear complexity and strong empirical results, AMTF-MLP presents a powerful and practical solution for real-world forecasting systems.