ASGMamba: Adaptive Spectral Gating Mamba for Multivariate Time Series Forecasting

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

💡 Research Summary

The paper “ASGMamba: Adaptive Spectral Gating Mamba for Multivariate Time Series Forecasting” addresses the critical challenge of Long-term Multivariate Time Series Forecasting (LTSF) in high-performance computing (HPC) environments, such as real-time energy grid management and large-scale traffic simulation. It identifies a key dilemma in existing solutions: Transformer-based models suffer from quadratic computational complexity, limiting scalability on long sequences, while efficient linear State Space Models (SSMs) like Mamba often fail to distinguish valuable high-frequency signals from stochastic noise, leading to inefficient use of their latent state capacity—a problem termed “state inefficiency.”

To bridge this gap, the authors propose ASGMamba, a novel forecasting framework designed for resource-constrained supercomputing environments. The core innovation is the integration of a lightweight Adaptive Spectral Gating (ASG) mechanism into the Mamba backbone. Unlike traditional methods that apply global Fourier transforms (O(L log L)), ASG operates on fixed-size local patches, maintaining strict linear complexity (O(L)). For each patch, it computes the local spectral energy density via FFT and generates a data-dependent gating signal. This gate acts as a frequency-selective filter, dynamically suppressing noise-dominated components before the input is processed by the Mamba’s selective scan mechanism. This prevents high-frequency noise from saturating the SSM’s state, allowing it to focus its limited capacity on modeling robust long-term temporal dynamics.

Furthermore, ASGMamba employs a hierarchical multi-scale architecture with three parallel branches using different patch sizes (8, 16, 32) to capture temporal patterns at various granularities. An overlapping patching strategy with 50% overlap is used to preserve boundary information and reduce spectral leakage. To compensate for the loss of semantic context caused by the efficient Channel-Independent (CI) processing strategy, the model introduces learnable Node Embeddings, which inject variable-specific physical characteristics into the shared backbone. Finally, a weighted fusion mechanism combines the predictions from each scale branch based on learned importance weights.

The proposed method is rigorously evaluated on nine real-world benchmark datasets. Experimental results demonstrate that ASGMamba achieves state-of-the-art accuracy, outperforming strong baselines including Transformer-based models (e.g., FEDformer, Autoformer, PatchTST) and recent SSM-based models (e.g., TimeMachine). Its superiority is particularly evident in long-horizon forecasting tasks. Crucially, complexity analysis confirms that ASGMamba maintains linear time and memory complexity while significantly reducing GPU memory footprint and inference latency compared to Transformer models on long sequences. This validates its practicality and scalability for high-throughput forecasting in resource-limited HPC environments.

In conclusion, ASGMamba successfully bridges spectral analysis and linear recurrence, offering a scalable and accurate solution to LTSF. By conditioning the SSM’s state evolution on local spectral properties, it effectively solves the state inefficiency problem, making it a promising approach for real-world applications where computational efficiency and robustness to noise are paramount. The code is publicly available to ensure reproducibility and foster future research.