AGformer: Adaptive filtering and gated fusion-attention for multivariate time series forecasting

Multivariate time series forecasting is a fundamental problem in sequence modeling, with broad applications spanning energy systems, meteorology, finance, and transportation. Despite significant progress in Transformer-based, TCN-based, and MLP-based architectures, three persistent neural network design challenges remain unresolved: 1) non-adaptive frequency-domain filtering that applies identical spectral parameters across heterogeneous variables; 2) inadequate utilization of informative short-term fluctuations that are discarded as noise yet encode critical event-driven dynamics; and 3) insufficient exploration of inter-scale dependencies in multi-scale temporal representations. To address these challenges, this paper proposes AGformer, a novel forecasting architecture integrating three complementary neural modules: 1) A variable-oriented Adaptive Multi-k Filter (AKF) that dynamically generates per-variable spectral filtering parameters from learned frequency-domain features, eliminating the need for manual global hyperparameter tuning. 2) A Dual-Gate Fusion-Attention (DGFA) mechanism that synergistically models long-term trends and short-term fluctuations via complementary temporal and content gating, recovering event-driven signals that are otherwise discarded by standard denoising. 3) A Multi-Scale Embedding (MSED) layer that fuses temporal patterns across resolutions through attention-based inter-scale weighting, capturing cross-resolution hierarchical dependencies that simple concatenation or summation cannot model. Extensive experiments on nine representative benchmark datasets across electricity, weather, finance, and transportation domains demonstrate AGformer's consistent superiority, achieving average reductions of 8.1%-23.6% in MSE and MAE relative to state-of-the-art baselines across nine benchmarks.