ASGMamba: Adaptive Spectral Gating Mamba for Multivariate Time Series Forecasting

TL;DR

ASGMamba tackles long-horizon multivariate time series forecasting under resource constraints by introducing Adaptive Spectral Gating (ASG) to condition the linear Mamba state-space backbone on local spectral energy. A multi-scale patching scheme with identity-preserving embeddings captures diverse temporal dynamics, while an adaptive fusion mechanism combines predictions from scales to handle heterogeneous variables. The approach preserves strict $O(L)$ complexity and reduces memory usage, enabling high-throughput forecasting on HPC hardware. Empirical results across nine real-world benchmarks show state-of-the-art or competitive accuracy with strong robustness to noise, especially in volatile, low-SNR environments, and clear efficiency gains over Transformer-based models.

Abstract

Long-term multivariate time series forecasting (LTSF) plays a crucial role in various high-performance computing applications, including real-time energy grid management and large-scale traffic flow simulation. However, existing solutions face a dilemma: Transformer-based models suffer from quadratic complexity, limiting their scalability on long sequences, while linear State Space Models (SSMs) often struggle to distinguish valuable signals from high-frequency noise, leading to wasted state capacity. To bridge this gap, we propose ASGMamba, an efficient forecasting framework designed for resource-constrained supercomputing environments. ASGMamba integrates a lightweight Adaptive Spectral Gating (ASG) mechanism that dynamically filters noise based on local spectral energy, enabling the Mamba backbone to focus its state evolution on robust temporal dynamics. Furthermore, we introduce a hierarchical multi-scale architecture with variable-specific Node Embeddings to capture diverse physical characteristics. Extensive experiments on nine benchmarks demonstrate that ASGMamba achieves state-of-the-art accuracy. While keeping strictly $$\mathcal{O}(L)$$ complexity we significantly reduce the memory usage on long-horizon tasks, thus establishing ASGMamba as a scalable solution for high-throughput forecasting in resource limited environments.The code is available at https://github.com/hit636/ASGMamba

Figures (8)

• Figure 1: System Architecture of ASGMamba. The framework employs a multi-scale parallel design ($P=\{8, 16, 32\}$). The core innovation is the Adaptive Spectral Gating (ASG) module within the Scale Block. By computing local spectral energy density via patch-level FFT, ASG generates a data-dependent gate $\mathbf{G}$. This gate acts as a frequency-selective filter, modulating the input $\mathbf{U}$ before it enters the Mamba encoder, thereby preventing high-frequency noise from contaminating the latent state space.
• Figure 2: Long-term forecasting performance comparison with an input length of $L=96$ for prediction horizons $T \in \{96, 192, 336, 720\}$. The best results are in red and the second best are blue. 'Avg' denotes the average performance.
• Figure 3: (continued)
• Figure 4: Visual comparison of forecasting performance on the Etth2 dataset with a horizon of $T=96$.
• Figure 5: Ablation results for different ASGMamba modules on ETTh1 and Weather. The table displays MSE and MAE values, as well as the relative performance deterioration (%) for each ablated part.