AdaptiveFrequencyNet: a long-term time series forecasting approach based on learnable frequency decomposition and cross-scale attention fusion Authors: JiaoJiao Cui, Wei Li, Xin Wang Published: 2026 ABSTRACT Long-term time series forecasting depends on modeling variation across time scales together with persistent temporal dependence. Existing frequency decomposition methods partially address multi-scale modeling but suffer from fixed frequency decomposition, simplistic feature fusion, and insufficient cross-scale interaction. Fixed splits can also shift the predictive distribution away from the input scale, which hurts long-horizon consistency. We describe AdaptiveFrequencyNet (AFNet), an architecture that combines learnable multi-scale filter banks with cross-scale attention fusion. WAD applies depthwise convolutions with branch-wise softmax mixing (a learnable time-domain bank, not a discrete wavelet transform (DWT) or hand-crafted Fourier subband split). AMSFF couples self-attention, multi-scale convolutions, and gating. CSAF uses cross-attention for progressive mixing across scales. On six standard benchmarks under a fixed multivariate protocol (shared lookback), AFNet is compared with nine strong baselines, including recent Transformer-style models iTransformer and Pathformer under the same evaluation procedure. AFNet attains competitive test errors at a moderate parameter budget (Table 4). Outcomes are dataset- and horizon-dependent: among tabulated dataset averages, iTransformer reports the lowest MSE on Electricity (0.170) with AFNet at 0.174, whereas on ETTh1 average MSE for AFNet (0.443) is close to Pathformer (0.439). On Weather, Pathformer and FreTS average lower MSE than AFNet in the same table. The reported numbers should be read only under this protocol. AMSFF and CSAF use standard multi-head attention; each attention block scales as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$O(T^2)$$\end{document} in the processed sequence length T and is therefore not linear in T. Ablation experiments under the same protocol support the contribution of each module. Introduction Research background Time series forecasting underpins applications in finance1,2, meteorology3,4, energy5,6, and transportation7,8. Neural forecasting models9 have improved markedly, yet long-horizon prediction remains difficult when short transients, slow drifts, and nonlinear couplings must be handled jointly. Traditional Recurrent Neural Networks10 and Long Short-Term Memory networks11 perform well in short-term forecasting but suffer from issues such as gradient vanishing and information loss in long-term forecasting. The Transformer architecture12 partially addresses the long-range dependency problem through self-attention mechanisms, but directly applying Transformers to time series forecasting still has limitations such as high computational complexity and insufficient capture of local patterns. Development of frequency decomposition methods Frequency decomposition treats a series as superposed components at different rates. Autoformer13 and FEDformer14 exemplify long-horizon models that rely on fixed FFT splits and relatively shallow fusion, which makes it hard to adapt band boundaries to the dataset or to mix scales nonlinearly (Tables 1, 2, 3). Research motivation and contributions The main limitations of existing frequency decomposition methods are: (1) Fixed frequency decomposition: using fixed FFT cannot adaptively select important frequency components based on data characteristics; (2) Simple feature fusion: single convolutions and simple residual connections are insufficient to effectively capture complex temporal dependencies; (3) Limited cross-scale interaction: fixed interpolation and simple addition cannot achieve effective interaction between different frequency components. We build AdaptiveFrequencyNet (AFNet) from three blocks:Weighted adaptive decomposition (WAD): a learnable multi-scale depthwise filter bank with K parallel branches per scale and softmax weights over branches (feature reweighting over parallel filters, not a discrete wavelet transform (DWT) or hand-designed Fourier subband split). Each scale uses its own kernel width; the stack replaces fixed FFT-style splits with data-driven temporal filtering.Adaptive multi-scale feature fusion (AMSFF): multi-head self-attention plus parallel multi-scale temporal convolutions, combined with a gate that mixes the two paths.Cross-scale attention fusion (CSAF): cross-attention with a fixed low-to-high traversal so coarse scales update finer ones; learned weights mix attended and residual branches at each step. Weighted adaptive decomposition (WAD): a learnable multi-scale depthwise filter bank with K parallel branches per scale and softmax weights over branches (feature reweighting over parallel filters, not a discrete wavelet transform (DWT) or hand-designed Fourier subband split). Each scale uses its own kernel width; the stack replaces fixed FFT-style splits with data-driven temporal filtering. Weighted adaptive decomposition (WAD): a learnable multi-scale depthwise filter bank with K parallel branches per scale and softmax weights over branches (feature reweighting over parallel filters, not a discrete wavelet transform (DWT) or hand-designed Fourier subband split). Each scale uses its own kernel width; the stack replaces fixed FFT-style splits with data-driven temporal filtering. Adaptive multi-scale feature fusion (AMSFF): multi-head self-attention plus parallel multi-scale temporal convolutions, combined with a gate that mixes the two paths. Adaptive multi-scale feature fusion (AMSFF): multi-head self-attention plus parallel multi-scale temporal convolutions, combined with a gate that mixes the two paths. Cross-scale attention fusion (CSAF): cross-attention with a fixed low-to-high traversal so coarse scales update finer ones; learned weights mix attended and residual branches at each step. Cross-scale attention fusion (CSAF): cross-attention with a fixed low-to-high traversal so coarse scales update finer ones; learned weights mix attended and residual branches at each step. Complexity. AMSFF and CSAF rely on standard multi-head attention; the dominant cost per block scales as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$O(T^2)$$\end{document} in the attended length T. Frequency decomposition and fusion layers and Table 4 separate this asymptotic cost from measured parameters and MACs. AFNet stacks depthwise multi-branch filtering, multi-head attention, multi-scale convolutions, cross-attention, and gating in one forecasting pipeline. All numerical statements below refer to the experiments and protocol in Section 4.2. Related Work Time series forecasting methods Time series forecasting methods have evolved from statistical methods to deep learning approaches. Early methods were mainly based on statistical models such as autoregressive models15, moving average models16, and autoregressive integrated moving average models17. While these methods have solid theoretical foundations, they have limitations in handling nonlinear relationships and long-term dependencies. The rise of deep learning has brought considerable advances to time series forecasting. Recurrent Neural Networks and their variants, LSTM11 and GRU18, model sequential dependencies through recurrent structures and perform well in short-term forecasting. However, RNN-based methods suffer from issues such as gradient vanishing, training difficulties, and low computational efficiency. The Transformer architecture12 enables parallel training and long-range modeling via self-attention and is now standard in NLP. Informer19 and Autoformer13 adapt Transformers to time series with sparse or frequency-aware variants. FEDformer14 adds explicit frequency attention. PatchTST20 tokenizes patches for long histories. Later designs include iTransformer21, Pathformer22, TimesNet23, and TimeMixer24. Section 4.2 links these lines of work to Table 2, where iTransformer and Pathformer appear as extra baselines under the AFNet protocol. Frequency decomposition methods Frequency decomposition is a classical tool in signal processing and has been reused extensively in deep forecasting25,26. Forecast-oriented variants map the series to components that emphasize different temporal scales; wavelets, empirical mode decompositions, and Fourier transforms are common building blocks. In deep learning, frequency decomposition methods typically decompose a time series into multiple frequency components and then model each component separately. Autoformer13 proposes a series decomposition mechanism that decomposes a time series into trend and seasonal components. FEDformer14 introduces a frequency-domain attention mechanism to model time series in the frequency domain. FreTS25 transforms time series to the frequency domain and designs dual-scale frequency-domain MLPs (inter-series and intra-series) to explore the global view and energy compression characteristics of frequency components. FiLM26 combines the historical information retention capability of Legendre polynomials with the high-frequency noise filtering characteristics of Fourier transform, introduces low-rank approximation to reduce computational overhead, and achieves high-precision long-term time series forecasting in noisy scenarios. Although frequency decomposition methods have achieved good results in long-term forecasting, they still have limitations: (1) Fixed FFT decomposition cannot be adaptively adjusted based on data characteristics; (2) Simple feature fusion struggles to capture complex temporal dependencies; (3) Fixed cross-scale fusion strategies limit the flexibility of information interaction. Attention mechanisms in time series Attention layers score pairwise time positions and are now standard in neural forecasting27,9. Self-attention improves long-range modeling relative to purely local operators but costs quadratic time in the sequence length, unlike fixed-depth convolutional stacks whose per-layer cost grows linearly with T. Multi-head variants run several attention maps in parallel. In time series work, attention is used along time, in frequency bins, and across scales. CSAF belongs to the third group: cross-attention between components with learned fusion weights and a fixed low-to-high schedule instead of a fully symmetric attention graph. Multi-scale feature learning Multi-scale feature learning is standard in deep networks for encoding patterns at several temporal spans22. Convolutions, pooling pyramids, and multi-branch designs are common implementations. Different time scales in time series correspond to different patterns: short-term scales capture local fluctuations and detailed information, while long-term scales capture trends and periodic patterns. The goal of multi-scale learning is to simultaneously model features at these different scales. Existing methods usually use fixed-scale convolution kernels or pooling windows, which may limit adaptability to heterogeneous data. The AMSFF module uses learnable multi-scale convolutions and weighted fusion to adjust emphasis across scales in a data-driven manner. Methodology Overall architecture Figure 1 outlines AdaptiveFrequencyNet with four stages: (1) average-pooling pyramids for multi-rate inputs; (2) normalization and Conv1D embeddings per scale; (3) stacked frequency decomposition and fusion blocks; (4) a linear readout followed by denormalization.Fig. 1Overall architecture of AdaptiveFrequencyNet. Overall architecture of AdaptiveFrequencyNet. Multi-scale input processing AFNet builds a pooled pyramid so that each level sees a shorter sequence at coarser resolution. Figure 2 sketches the downsampling stack.Fig. 2Multi-scale input processing architecture. Multi-scale input processing architecture. Multi-scale downsampling We use 1D average pooling operations for hierarchical downsampling to generate multi-scale representations. Starting from the original sequence \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{X}^{(0)} = \textbf{X}$$\end{document}, we iteratively downsample:1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \textbf{X}^{(i)} = \left( \text {AvgPool1D}_{w}\left( (\textbf{X}^{(i-1)})^T\right) \right) ^T, \quad i = 1, 2, \ldots , L \end{aligned}$$\end{document}where w is the downsampling window size (usually set to 2), and L is the number of downsampling layers. This yields the multi-scale input set:2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \mathcal {X} = \{\textbf{X}^{(0)}, \textbf{X}^{(1)}, \ldots , \textbf{X}^{(L)}\} \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{X}^{(0)}$$\end{document} corresponds to the highest frequency (original resolution), and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{X}^{(L)}$$\end{document} corresponds to the lowest frequency (maximum downsampling). Physical interpretation Different frequency levels capture different temporal patterns: High-frequency component \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{X}^{(0)}$$\end{document}: Captures short-term fluctuations, detailed information, and rapid changes. Mid-frequency components \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{X}^{(i)}, i \in [1, L-1]$$\end{document}: Capture medium-term patterns, periodicity, and local trends. Low-frequency component \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{X}^{(L)}$$\end{document}: Captures long-term trends, overall patterns, and global characteristics. Data embedding and normalization Each scale is normalized to stabilize training, reshaped for channel-independent processing, and embedded with a shallow Conv1D so that later WAD/CSAF layers operate on fixed-width feature maps. Figure 3 sketches this block.Fig. 3Data embedding and normalization architecture. Data embedding and normalization architecture. Reversible instance normalization For each frequency level \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{X}^{(i)}$$\end{document}, we apply reversible instance normalization to eliminate scale differences between different variables and improve training stability. For each variable j, we compute the mean \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mu ^{(i)}_j$$\end{document} and standard deviation \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sigma ^{(i)}_j$$\end{document}, then perform normalization:3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \tilde{X}^{(i)}_{j,t} = \frac{X^{(i)}_{j,t} - \mu ^{(i)}_j}{\sigma ^{(i)}_j} \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\epsilon = 10^{-5}$$\end{document} is a constant for numerical stability. The statistics \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mu ^{(i)}_j$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sigma ^{(i)}_j$$\end{document} are saved for the subsequent denormalization process. Variable independence and data embedding To handle multivariate time series, we reshape the normalized data into a channel-independent format, then use 1D convolution to map scalar values to a high-dimensional feature space:4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \textbf{E}^{(i)} = \text {Conv1D}\left( \text {Reshape}\left( \tilde{\textbf{X}}^{(i)}\right) \right) \in \mathbb {R}^{(B \cdot C) \times T_i \times d_{model}} \end{aligned}$$\end{document}where the convolution kernel size is 3, padding is 1 to maintain sequence length, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$d_{model}$$\end{document} is the model dimension (typically 64). This reshaping enables each variable’s time series to be processed independently, helping the model learn variable-specific patterns. This yields the multi-scale embedding set:5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \mathcal {E}_0 = \{\textbf{E}^{(0)}, \textbf{E}^{(1)}, \ldots , \textbf{E}^{(L)}\} \end{aligned}$$\end{document} Frequency decomposition and fusion layers The Frequency Decomposition and Fusion Layer (FDFL) is the core module of AFNet. Each FDFL consists of two cascaded submodules: Weighted Adaptive Decomposition (WAD) and Cross-Scale Attention Fusion (CSAF). By stacking \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$N_{layers}$$\end{document} FDFLs, the model can gradually extract and refine multi-scale representations. The overall framework of the FDFL is shown in Figure 4.Fig. 4Frequency decomposition and fusion layer architecture. Frequency decomposition and fusion layer architecture. For the l-th FDFL layer, its processing can be expressed as:6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \mathcal {E}_{l}^{decomp} = \text {WAD}_l(\mathcal {E}_l) \end{aligned}$$\end{document}7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \mathcal {E}_{l+1} = \text {CSAF}_l(\mathcal {E}_{l}^{decomp}) \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {E}_l$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {E}_{l+1}$$\end{document} denote the multi-scale embedding sets at layer l and (l+1), respectively. Through multi-layer stacking, we have:8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \mathcal {E}_0 \xrightarrow {\text {FDFL}_0} \mathcal {E}_1 \xrightarrow {\text {FDFL}_1} \mathcal {E}_2 \xrightarrow {} \cdots \xrightarrow {\text {FDFL}_{N_{layers}-1}} \mathcal {E}_{N_{layers}} \end{aligned}$$\end{document} Weighted adaptive decomposition (WAD) Terminology. WAD stands for Weighted Adaptive Decomposition. In practice it is a learnable multi-scale depthwise filter bank: K parallel convolutions per scale (shared kernel width at that scale) followed by softmax weights over branches, i.e. branch-wise reweighting of time-domain filter outputs. The module is not a discrete wavelet transform, not tied to a fixed orthonormal wavelet basis, and not a rigid Fourier subband partition, although the learned responses can emphasize complementary temporal structure similarly to adaptive filtering. The WAD module builds multi-scale representations and propagates residual information from coarse to fine scales. It replaces a fixed FFT split with learnable temporal filtering. The overall layout is shown in Figure 5.Fig. 5Weighted adaptive decomposition (WAD): learnable multi-scale depthwise filtering with branch-wise softmax mixing and gated residuals between aligned scales. Weighted adaptive decomposition (WAD): learnable multi-scale depthwise filtering with branch-wise softmax mixing and gated residuals between aligned scales. For each frequency level \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$i \in \{0,\ldots ,L\}$$\end{document} we instantiate one filter bank. At every scale the bank holds \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$K{=}4$$\end{document} parallel depthwise convolutions that share the kernel width \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$k_i = 7 + 2i$$\end{document}; the index k therefore labels parallel filters, not four different widths \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\{7,9,11,13\}$$\end{document} on the same scale. Across scales the widths differ, which yields a scale-dependent bank over the pyramid. A branch-mixing head applies global average pooling over time, a two-layer MLP, and softmax over the K outputs (weights for the parallel filters, not a Fourier band selector):9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \textbf{a}^{(i)} = \text {Softmax}\left( \text {MLP}\left( \frac{1}{T_i} \sum _{n=1}^{T_i} \textbf{E}^{(i)}_{:,n,:}\right) \right) \in \mathbb {R}^{(B \cdot C) \times K} \end{aligned}$$\end{document}Let \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{F}^{(i)}_k$$\end{document} be the output of the k-th depthwise filter (kernel \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$k_i$$\end{document} shared across k). The per-scale tensor aggregated for the residual path is10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \textbf{E}^{(i)}_{\textrm{sum}} = \sum _{k=1}^{K} \textbf{a}^{(i)}_k \odot \textbf{F}^{(i)}_k, \end{aligned}$$\end{document}where ⊙ denotes broadcasted element-wise multiplication with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{a}^{(i)}_k$$\end{document} expanded over time and feature dimensions via standard broadcasting. Branch count \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$K{=}4$$\end{document} . Four parallel filters per scale strike a practical balance between response diversity for softmax mixing and training cost when the bank is stacked across scales and layers. Smaller K tends to narrow the set of motifs; larger K adds parameters with quickly saturating gains in informal pilots. We therefore keep \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$K{=}4$$\end{document} as the default and rely on Table 3 for module-level evidence rather than sweeping K in the main text; Appendix B lists a small ETTh1 pilot over \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$K \in \{3,4,5\}$$\end{document}. We do not assert global optimality of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$K{=}4$$\end{document} on every dataset or horizon. To progressively transfer information from low-frequency scales to higher-frequency ones, the model processes scales from low to high (after reversing the multi-scale list), aligns the coarser representation to the finer length with linear interpolation plus a light MLP-based residual on the interpolated tensor, and forms a gated difference between adjacent scales. Concretely, let \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\tilde{\textbf{E}}^{(i)}_{\textrm{up}}$$\end{document} denote the upsampled and aligned coarser feature at the length of scale i+1. The gated residual reads11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \textbf{E}^{(i+1)}_{\textrm{res}} = \sigma \left( \text {MLP}_{gate}\left( \bar{\textbf{E}}^{(i+1)}\right) \right) \odot \left( \textbf{E}^{(i+1)}_{\textrm{sum}} - \tilde{\textbf{E}}^{(i)}_{\textrm{up}}\right) , \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{E}^{(i+1)}_{\textrm{sum}}$$\end{document} is defined as above, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\bar{\textbf{E}}^{(i+1)}$$\end{document} is the temporal mean of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{E}^{(i+1)}_{\textrm{sum}}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {MLP}_{gate}$$\end{document} maps to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$d_{model}$$\end{document} logits, and σ is the sigmoid (element-wise gate). The construction emphasizes a gated residual between aligned scales rather than an unconstrained additive path. CSAF module The Cross-Scale Attention Fusion (CSAF) module fuses information across frequency components using cross-attention and learned interpolation weights, walking scales from low to high frequency. Figure 6 shows the wiring.Fig. 6CSAF module architecture. CSAF module architecture. Low-to-high traversal Processing low-frequency scales before high-frequency ones lets the carry tensor accumulate coarse structure before adding detail, which matches the coarse-to-fine pyramid from Section 3.2. Reversing the order changes the path-dependent mixing; we keep the low-to-high schedule to limit the experimental surface area. Appendix B shows a pilot where reversing the loop hurts ETTh1 accuracy. First, for each frequency component \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{E}^{(i)}_{decomp}$$\end{document}, we apply multi-scale adaptive fusion for feature extraction:12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \textbf{E}^{(i)}_{fused} = \textrm{AMSFF}\left( \textbf{E}^{(i)}_{decomp}\right) \end{aligned}$$\end{document}Starting from the lowest frequency component (index L), we progressively process higher frequency components. For each higher frequency component i, we upsample the processed lower frequency component to the current high-frequency scale and perform scale alignment. Then, we use a cross-attention mechanism where the high-frequency component serves as Query, and the low-frequency component serves as Key and Value:13\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \textbf{E}^{(i)}_{attended} = \text {Softmax}\left( \frac{\textbf{Q}^{(i)} (\textbf{K}^{(i)})^T}{\sqrt{d_k}}\right) \textbf{V}^{(i)} \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{Q}^{(i)} = \textbf{E}^{(i)}_{fused} \textbf{W}_Q$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{K}^{(i)} = \textbf{E}^{(i)}_{aligned} \textbf{W}_K$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{V}^{(i)} = \textbf{E}^{(i)}_{aligned} \textbf{W}_V$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$d_k = d_{model} / H$$\end{document}, and H is the number of attention heads (capped at 8 when \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$d_{model}$$\end{document} is small). Here \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{E}^{(i)}_{aligned}$$\end{document} is the upsampled low-frequency tensor used only as keys and values; it is distinct from the attention output. Multi-head cross-attention follows the usual batch-first layout. The attention output \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{E}^{(i)}_{attended}$$\end{document} is then mixed with the pre-attention branch \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{E}^{(i)}_{fused}$$\end{document} (not with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{E}^{(i)}_{aligned}$$\end{document}):14\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \textbf{E}^{(i)}_{final} = \lambda ^{(i)} \textbf{E}^{(i)}_{fused} + \left( 1 - \lambda ^{(i)}\right) \textbf{E}^{(i)}_{attended}, \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda ^{(i)} = \sigma \big (\text {MLP}_{weight}(\text {Mean}_{t}[\text {Concat}(\textbf{E}^{(i)}_{fused},\textbf{E}^{(i)}_{attended})])\big )$$\end{document}: the gate depends only on \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{E}^{(i)}_{fused}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{E}^{(i)}_{attended}$$\end{document}, the same tensors that appear in the linear mixture. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {Concat}$$\end{document} is along the feature dimension (width \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2d_{model}$$\end{document}), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {Mean}_{t}$$\end{document} averages over time, and the sigmoid map is broadcast over time. The carry tensor for the next step is updated as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{E}_{\textrm{carry}} \leftarrow \textbf{E}^{(i)}_{final}$$\end{document} (low-to-high schedule), consistent with the CSAF description above. Adaptive multi-scale feature fusion The Adaptive Multi-Scale Feature Fusion (AMSFF) block processes one frequency component at a time, mixing a self-attention path with parallel depthwise convolutions and a gate. Figure 7 gives the layout.Fig. 7Adaptive multi-scale feature fusion module architecture. Adaptive multi-scale feature fusion module architecture. First, we apply a two-layer MLP for channel mixing (with a residual connection and layer normalization), then capture long-range dependencies via multi-head self-attention. For each attention head h, we compute:15\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \textbf{E}_{head,h} = \text {Softmax}\left( \frac{\textbf{Q}_h \textbf{K}_h^T}{\sqrt{d_k}}\right) \textbf{V}_h \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{Q}_h$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{K}_h$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{V}_h$$\end{document} are projections of the channel-mixed features. The outputs of all heads are concatenated and projected to obtain \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{E}_{attn}$$\end{document}. To capture multi-scale local patterns, we use three depthwise temporal convolutions with kernel sizes (3, 5, 7) in parallel, corresponding to short-term, medium-term, and long-term local patterns, respectively. Kernel set \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\{3,5,7\}$$\end{document} . Odd widths \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\{3,5,7\}$$\end{document} give symmetric receptive fields at short, medium, and longer spans without a single very wide kernel. Softmax-normalized depthwise weights allocate mass across the three paths while keeping the convolutional branch light relative to self-attention. Other tuples may work on some benchmarks; we leave a broader search aside and record a small ETTh1 pilot in Appendix B. Learnable weights are normalized with softmax for weighted fusion:16\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \textbf{E}_{temporal} = \sum _{k \in \{3,5,7\}} w_k \cdot \text {Conv1D}_k\left( \textbf{E}_{attn}\right) \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{w} = \text {Softmax}\left( \boldsymbol{\theta }_{scale}\right)$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{\theta }_{scale}$$\end{document} are learnable scale weight parameters. Finally, we fuse the attention features and temporal convolution features using a gating mechanism:17\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \textbf{E}_{out} = \textbf{g} \odot \textbf{E}_{attn} + (1 - \textbf{g}) \odot \textbf{E}_{temporal} \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{g} = \sigma \left( \text {MLP}_{gate}\left( \bar{\textbf{E}}_{combined}\right) \right)$$\end{document} is computed from the concatenated features. Computational complexity AMSFF uses the same multi-head self-attention operator as in Transformers12. With sequence length T, dot-product attention costs \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$O(T^2)$$\end{document} time and memory in standard implementations, so AMSFF is not linear in T. CSAF applies multi-head cross-attention between aligned scales and inherits the same \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$O(T^2)$$\end{document} scaling in the attended length. We therefore report both asymptotic order in T and empirical counts (parameters, MACs, and optional wall-clock under the profiler settings of Table 4). Total runtime also depends on the number of pyramid levels, FDFL layers, and heads; the full AFNet stack has no closed-form linear-time guarantee. Prediction output After processing through all FDFL layers, we use the output at the highest frequency (original resolution) for final prediction. Its overall framework is shown in Fig. 8.Fig. 8Prediction output architecture. Prediction output architecture. Time dimension mapping and feature projection We use a linear layer to map the sequence length from input length T to prediction length P, then project the feature dimension from \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$d_{model}$$\end{document} to the output dimension (typically 1):18\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \textbf{H} = \text {Linear}_{T \rightarrow P}\left( \textbf{E}^{(0)}_{N_{layers}}\right) \in \mathbb {R}^{(B \cdot C) \times P \times d_{model}} \end{aligned}$$\end{document}19\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \textbf{Y}_{flat} = \textbf{H} \textbf{W}_{proj} + \textbf{b}_{proj} \in \mathbb {R}^{(B \cdot C) \times P \times 1} \end{aligned}$$\end{document} Reshaping and Denormalization We reshape the output into a multivariate format and apply denormalization to restore the original data scale:20\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \hat{\textbf{Y}} = \text {Reshape}\left( \textbf{Y}_{flat}\right) ^T \cdot \boldsymbol{\sigma }^{(0)} + \boldsymbol{\mu }^{(0)} \in \mathbb {R}^{B \times P \times C} \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{\mu }^{(0)}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\boldsymbol{\sigma }^{(0)}$$\end{document} are the statistics saved during input normalization. Experiments Datasets The experiments use ETTh1, ETTh2, ETTm1, ETTm2, Weather, and Electricity, which are standard long-term forecasting benchmarks with diverse sampling rates and dimensionalities. Table 1 summarizes basic statistics.Table 1Statistical overview of the datasetsParameterETTh1ETTh2ETTm1ETTm2ElectricityWeatherNumber of features777732121Number of timesteps174201742069680696802630452696 Table 1Statistical overview of the datasetsParameterETTh1ETTh2ETTm1ETTm2ElectricityWeatherNumber of features777732121Number of timesteps174201742069680696802630452696 Statistical overview of the datasets Experimental design For baseline models we use nine published long-horizon forecasters: (1) Transformer-based methods: Autoformer13, FEDformer14, PatchTST20, iTransformer21, Pathformer22. (2) MLP-based method: Dlinear28. (3) CNN-based method: MICN29. (4) Frequency-based methods: FreTS25, FiLM26. All runs share lookback T = 96, horizons F=[96, 192, 336, 720], multivariate targets, and MSE/MAE on the test split. Reproducibility and baseline provenance AFNet. Numbers for AFNet come from the open-source release bundled with this manuscript. Training follows the public long-term forecasting recipe (model tag AFNet) with random seed 2048 unless noted, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$d_{model}{=}64$$\end{document}, two encoder layers, and dataset-specific pooling depth and window from the configuration files. Splits match the Autoformer / Time-Series-Library convention for ETTh, ETTm, Weather, and Electricity. Baselines. Columns for Autoformer, FEDformer, PatchTST, DLinear, MICN, FreTS, and FiLM are compiled from reference implementations and prior papers so that multivariate \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$T{=}96$$\end{document}, horizons \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\{96,192,336,720\}$$\end{document}, and test MSE/MAE line up with the community reporting style. iTransformer and Pathformer were rerun with their official code under the same splits, lookback, horizons, and metrics22,21. Columns may come from different machines and library builds, so small gaps versus published cards are expected; anyone reproducing the table should rerun the cited codebases with matched splits. The repository includes a short checklist for AFNet. Additional recent forecasting baselines Long-term forecasting continues to accumulate Transformer-style architectures. Besides the classical baselines above, we add iTransformer21 and Pathformer22 to Table 2 using the same multivariate recipe as AFNet. TimesNet23 and TimeMixer24 are discussed here for context: their public recipes often change lookback, tokenization, or preprocessing, so we do not paste their headline numbers next to ours without a matched rerun. Our own runs cover AFNet trained under Section 4.2 with lookback \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$T{=}96$$\end{document}. Table 2 collects baselines under that same multivariate contract; for iTransformer and Pathformer the entries are fresh runs from the official repositories22,21. TimesNet and TimeMixer citations24,23 are qualitative pointers only. Quantitative statements are tied to this protocol rather than to an exhaustive sweep of competitor-specific budgets. For stricter comparisons, retrain every candidate with the same code template (for example Time-Series-Library); Section 4.2 documents how we reproduced AFNet. Main results Protocol Table 2 fixes lookback \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$T{=}96$$\end{document} and horizons \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$F\in \{96,192,336,720\}$$\end{document} (Section 4.2). The text below stays within that setting. Table 2 lists multivariate test MSE and MAE; lower is better. Across the table, AFNet matches or beats many baselines on individual dataset–horizon cells. Aggregate error AFNet achieves competitive MSE and MAE on many horizons (96–720) across the six datasets. On Electricity, its tabulated dataset-average MSE (0.174) is slightly above iTransformer (0.170) but below most other listed methods; on Weather it remains usable across horizons yet does not attain the lowest dataset-average MSE (Pathformer 0.239, FreTS 0.245 versus AFNet 0.257 in Table 2). Long horizons As the forecast horizon extends (e.g., to 720 steps), AFNet often degrades less than unstable baselines. On ETTh2, MSE rises from 0.292 to 0.437 for AFNet, whereas MICN moves from 0.372 to 0.824 under the same table. Per-dataset averages Dataset-average MSE across the four horizons reads as follows: ETTh1: AFNet achieves an average MSE of 0.443, competitive with several baselines in Table 2 (Pathformer 0.439; iTransformer 0.454) and lower than Dlinear (0.461) among the listed classical baselines. ETTh2: AFNet obtains an average MSE of 0.387; several baselines report lower averages (e.g., Pathformer at 0.344, iTransformer at 0.383), while AFNet remains competitive with PatchTST (0.391) and clearly stronger than unstable large-error baselines such as MICN on this split. Weather: AFNet records an average MSE of 0.257, which is competitive with several baselines but is not the best in Table 2; Pathformer and FreTS achieve lower average MSE values of 0.239 and 0.245, respectively. Electricity: AFNet averages 0.174, second only to iTransformer (0.170) among the methods in Table 2. Cross-domain behaviour Errors vary by domain, as expected for a single fixed architecture; we do not extrapolate beyond the six benchmarks. Under the reported protocol, AFNet is competitive with the baselines in Table 2 on several splits, with Electricity averages near the tabulated minimum and Weather averages behind Pathformer/FreTS. These values depend on (T, F) and the sourcing notes in Experimental design.Table 2Multivariate comparison experimental results. iTransformer and Pathformer are evaluated as additional baselines under the same protocol as AFNet (Section Experimental design; lookback \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$H{=}T{=}96$$\end{document}, horizons \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$F \in \{96,192,336,720\}$$\end{document}) using their official implementations22,21. Other baseline columns follow the compilation caveat described in Section Experimental design.DatasetHorizonAFNetPatchTSTMICNDlinearFreTSFiLMFEDformerAutoformeriTransformerPathformerMSEMAEMSEMAEMSEMAEMSEMAEMSEMAEMSEMAEMSEMAEMSEMAEMSEMAEMSEMAEETTh1960.3820.4020.460.4470.4260.4460.3970.4120.3950.4070.4380.4330.3950.4240.4490.4590.3860.4050.3820.4001920.4310.4290.5120.4770.4540.4640.4460.4410.490.4770.4940.4660.4690.470.50.4820.4410.4360.4400.4273360.4720.4520.5460.4960.4930.4870.4890.4670.510.480.5470.4950.490.4770.5210.4960.4870.4580.4540.4327200.4870.4860.5440.5170.5260.5260.5130.510.5680.5380.5860.5380.5980.5440.5140.5120.5030.4910.4790.461Avg0.4430.4420.5160.4840.4750.480.4610.4570.4910.4750.5160.4830.4980.4840.4960.4870.4540.4480.4390.430ETTh2960.2920.3390.3080.3550.3720.4240.340.3940.3220.3640.3220.3640.3580.3970.3460.3880.2970.3490.2790.3311920.3970.4080.3930.4050.4920.4920.4280.4790.4510.4140.4050.4140.4290.4390.4560.4520.3800.4000.3490.3803360.4230.4310.4270.4360.6070.5550.5910.5410.4660.4450.4350.4450.4960.4870.4820.4860.4280.4320.3480.3827200.4370.4450.4360.450.8240.6550.8390.6610.4850.4570.4450.4570.4630.4740.5150.5110.4270.4450.3980.424Avg0.3870.4060.3910.4110.5740.5310.5630.5190.4330.420.4020.420.4370.4490.450.4590.3830.4070.3440.379ETTm1960.3240.3620.3520.3740.3650.3870.3460.3740.3370.3740.3530.370.3790.4190.5050.4750.3340.3680.3160.3461920.3680.380.390.3930.4030.4080.3820.3910.3820.3980.3890.3870.4260.4410.5530.4960.3770.3910.3660.3703360.3990.4180.4210.4140.4360.4310.4150.4150.420.4230.4210.4080.4450.4590.6210.5370.4260.4200.3860.3947200.4780.4550.4610.4490.4890.4620.4730.4510.490.4710.4810.4410.5430.490.6710.5610.4910.4590.4600.432Avg0.3920.4040.4060.4070.4230.4220.4040.4080.4070.4170.4120.4020.4480.4520.5880.5170.4070.4100.3820.386ETTm2960.1780.2640.1830.270.1970.2960.1930.2930.1860.2750.1830.2660.2030.2870.2550.3390.1800.2640.1700.2481920.2410.3050.2550.3140.2840.3610.2840.3610.2590.3230.2480.3050.2690.3280.2810.340.2500.3090.2380.2953360.3090.3380.3090.3470.3810.4290.3820.4290.3490.3860.3090.3430.3250.3660.3390.3720.3110.3480.2930.3317200.4020.4010.4120.4040.5490.5220.5580.5250.5590.5110.410.40.4210.4150.4330.4320.4120.4070.3900.389Avg0.2820.3270.290.3340.3530.4020.3540.4020.3390.3740.2880.3280.3050.3490.3270.3710.2880.3320.2730.316Weather960.170.2160.1860.2270.1980.2610.1950.2520.1710.2270.1950.2360.2170.2960.2660.3360.1740.2140.1560.1921920.2310.2620.2340.2650.2390.2990.2370.2950.2180.280.2390.2710.2760.3360.3070.3670.2210.2540.2060.2403360.2780.2950.2840.3010.2850.3360.2820.3310.2650.3170.2890.3060.3390.380.3590.3950.2780.2960.2540.2827200.3490.3410.3560.3490.3510.3880.3450.3820.3260.3510.360.3510.4030.4280.4190.4280.3580.3470.3400.336Avg0.2570.2790.2650.2850.2680.3210.2650.3150.2450.2940.2710.290.3090.360.3380.3820.2580.2780.2390.263Electricity960.140.2350.190.2960.180.2930.210.3020.1710.260.1980.2740.1930.3080.2010.3170.1450.2360.1450.2361920.1620.2520.1990.3040.1890.3020.210.3050.1770.2680.1980.2780.2010.3150.2220.3340.1520.2500.1670.2563360.1750.2660.2170.3190.1980.3120.2230.3190.190.2840.2170.30.2140.3290.2310.4430.1760.2710.1860.2757200.2190.3110.2580.3520.2170.330.2580.350.2280.3160.2780.3560.2460.3550.2540.3610.2060.3050.2310.309Avg0.1740.2660.2160.3180.1960.3090.2250.3190.1920.2820.2230.3020.2140.3270.2270.3380.1700.2660.1820.269 Table 2Multivariate comparison experimental results. iTransformer and Pathformer are evaluated as additional baselines under the same protocol as AFNet (Section Experimental design; lookback \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$H{=}T{=}96$$\end{document}, horizons \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$F \in \{96,192,336,720\}$$\end{document}) using their official implementations22,21. Other baseline columns follow the compilation caveat described in Section Experimental design.DatasetHorizonAFNetPatchTSTMICNDlinearFreTSFiLMFEDformerAutoformeriTransformerPathformerMSEMAEMSEMAEMSEMAEMSEMAEMSEMAEMSEMAEMSEMAEMSEMAEMSEMAEMSEMAEETTh1960.3820.4020.460.4470.4260.4460.3970.4120.3950.4070.4380.4330.3950.4240.4490.4590.3860.4050.3820.4001920.4310.4290.5120.4770.4540.4640.4460.4410.490.4770.4940.4660.4690.470.50.4820.4410.4360.4400.4273360.4720.4520.5460.4960.4930.4870.4890.4670.510.480.5470.4950.490.4770.5210.4960.4870.4580.4540.4327200.4870.4860.5440.5170.5260.5260.5130.510.5680.5380.5860.5380.5980.5440.5140.5120.5030.4910.4790.461Avg0.4430.4420.5160.4840.4750.480.4610.4570.4910.4750.5160.4830.4980.4840.4960.4870.4540.4480.4390.430ETTh2960.2920.3390.3080.3550.3720.4240.340.3940.3220.3640.3220.3640.3580.3970.3460.3880.2970.3490.2790.3311920.3970.4080.3930.4050.4920.4920.4280.4790.4510.4140.4050.4140.4290.4390.4560.4520.3800.4000.3490.3803360.4230.4310.4270.4360.6070.5550.5910.5410.4660.4450.4350.4450.4960.4870.4820.4860.4280.4320.3480.3827200.4370.4450.4360.450.8240.6550.8390.6610.4850.4570.4450.4570.4630.4740.5150.5110.4270.4450.3980.424Avg0.3870.4060.3910.4110.5740.5310.5630.5190.4330.420.4020.420.4370.4490.450.4590.3830.4070.3440.379ETTm1960.3240.3620.3520.3740.3650.3870.3460.3740.3370.3740.3530.370.3790.4190.5050.4750.3340.3680.3160.3461920.3680.380.390.3930.4030.4080.3820.3910.3820.3980.3890.3870.4260.4410.5530.4960.3770.3910.3660.3703360.3990.4180.4210.4140.4360.4310.4150.4150.420.4230.4210.4080.4450.4590.6210.5370.4260.4200.3860.3947200.4780.4550.4610.4490.4890.4620.4730.4510.490.4710.4810.4410.5430.490.6710.5610.4910.4590.4600.432Avg0.3920.4040.4060.4070.4230.4220.4040.4080.4070.4170.4120.4020.4480.4520.5880.5170.4070.4100.3820.386ETTm2960.1780.2640.1830.270.1970.2960.1930.2930.1860.2750.1830.2660.2030.2870.2550.3390.1800.2640.1700.2481920.2410.3050.2550.3140.2840.3610.2840.3610.2590.3230.2480.3050.2690.3280.2810.340.2500.3090.2380.2953360.3090.3380.3090.3470.3810.4290.3820.4290.3490.3860.3090.3430.3250.3660.3390.3720.3110.3480.2930.3317200.4020.4010.4120.4040.5490.5220.5580.5250.5590.5110.410.40.4210.4150.4330.4320.4120.4070.3900.389Avg0.2820.3270.290.3340.3530.4020.3540.4020.3390.3740.2880.3280.3050.3490.3270.3710.2880.3320.2730.316Weather960.170.2160.1860.2270.1980.2610.1950.2520.1710.2270.1950.2360.2170.2960.2660.3360.1740.2140.1560.1921920.2310.2620.2340.2650.2390.2990.2370.2950.2180.280.2390.2710.2760.3360.3070.3670.2210.2540.2060.2403360.2780.2950.2840.3010.2850.3360.2820.3310.2650.3170.2890.3060.3390.380.3590.3950.2780.2960.2540.2827200.3490.3410.3560.3490.3510.3880.3450.3820.3260.3510.360.3510.4030.4280.4190.4280.3580.3470.3400.336Avg0.2570.2790.2650.2850.2680.3210.2650.3150.2450.2940.2710.290.3090.360.3380.3820.2580.2780.2390.263Electricity960.140.2350.190.2960.180.2930.210.3020.1710.260.1980.2740.1930.3080.2010.3170.1450.2360.1450.2361920.1620.2520.1990.3040.1890.3020.210.3050.1770.2680.1980.2780.2010.3150.2220.3340.1520.2500.1670.2563360.1750.2660.2170.3190.1980.3120.2230.3190.190.2840.2170.30.2140.3290.2310.4430.1760.2710.1860.2757200.2190.3110.2580.3520.2170.330.2580.350.2280.3160.2780.3560.2460.3550.2540.3610.2060.3050.2310.309Avg0.1740.2660.2160.3180.1960.3090.2250.3190.1920.2820.2230.3020.2140.3270.2270.3380.1700.2660.1820.269 Multivariate comparison experimental results. iTransformer and Pathformer are evaluated as additional baselines under the same protocol as AFNet (Section Experimental design; lookback \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$H{=}T{=}96$$\end{document}, horizons \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$F \in \{96,192,336,720\}$$\end{document}) using their official implementations22,21. Other baseline columns follow the compilation caveat described in Section Experimental design. Lookback fairness and comparison scope Fixing \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$T{=}96$$\end{document} for all models in Table 2 gives every method the same history length on each split, which is the usual reporting convention for these benchmarks. Architectures differ in how much they gain from longer contexts; PatchTST, for instance, is often tuned with histories of 336 or 512 steps, where attention over patches behaves differently. Scope. The table is one controlled snapshot, not a search over each competitor’s preferred lookback. Readers who need other histories should rerun the public code with matched splits. Appendix B mentions optional sensitivity work (including other T) for future releases. Ablation studies We ablate Weighted Adaptive Decomposition, Cross-Scale Attention Fusion, and Adaptive Multi-Scale Feature Fusion.Weighted adaptive decomposition (WAD): AFNet(FFT) swaps the learnable bank for a fixed FFT decomposition.Cross-scale attention fusion: AFNet(FixedInterp) replaces CSAF with fixed interpolation upsampling.Adaptive multi-scale feature fusion: AFNet(SingleConv) replaces AMSFF with one depthwise convolution. Weighted adaptive decomposition (WAD): AFNet(FFT) swaps the learnable bank for a fixed FFT decomposition. Weighted adaptive decomposition (WAD): AFNet(FFT) swaps the learnable bank for a fixed FFT decomposition. Cross-scale attention fusion: AFNet(FixedInterp) replaces CSAF with fixed interpolation upsampling. Cross-scale attention fusion: AFNet(FixedInterp) replaces CSAF with fixed interpolation upsampling. Adaptive multi-scale feature fusion: AFNet(SingleConv) replaces AMSFF with one depthwise convolution. Adaptive multi-scale feature fusion: AFNet(SingleConv) replaces AMSFF with one depthwise convolution. Scope of ablations and internal hyperparameters Table 3 swaps each of WAD, CSAF, and AMSFF for a simpler drop-in (fixed FFT, fixed interpolation, or a single convolution). Internal hyperparameters (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$K{=}4$$\end{document} in WAD, the low-to-high CSAF loop, and the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\{3,5,7\}$$\end{document} depthwise tuple in AMSFF) are motivated in the Methodology and are not grid-searched on every dataset here. Appendix B adds an ETTh1 pilot at \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$T{=}F{=}96$$\end{document} with mean±std over two seeds (Table 5). Table 3 supports the following reading: AFNet outperforms AFNet(FFT), AFNet(FixedInterp), and AFNet(SingleConv) on most dataset–horizon pairs, so the full configuration is preferred under this protocol. AFNet(FFT) typically loses accuracy relative to WAD, especially at \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$F{=}720$$\end{document}, consistent with replacing the learnable bank by a rigid spectral split. AFNet(FixedInterp) degrades clearly relative to CSAF; gains are largest on ETTh2 and ETTm2 at long horizons, where fixed upsampling appears to bottleneck cross-scale mixing. AFNet(SingleConv) shows the largest gaps on Weather and Electricity, where the multi-branch AMSFF path seems to matter most. No single ablated variant matches the full model everywhere, so the three modules behave as complementary rather than redundant additions in these experiments.Table 3Ablation study resultsDatasetHorizonAFNetAFNet(FFT)AFNet(FixedInterp)AFNet(SingleConv)MSEMAEMSEMAEMSEMAEMSEMAEETTh1960.3820.4020.4550.4470.4340.4580.3990.4101920.4310.4290.5100.4770.4650.4740.4550.4383360.4720.4520.5450.4960.5050.4980.4900.4577200.4870.4860.5900.5180.5350.5380.5160.516Avg0.4430.4420.5250.4850.4850.4920.4650.455ETTh2960.2920.3390.3280.3700.3850.4350.3610.4011920.3970.4080.4150.4250.5030.5060.4470.4923360.4230.4310.4310.4510.6190.5710.6210.5647200.4370.4450.4520.4750.8360.6680.9850.685Avg0.3870.4060.4070.4300.5860.5450.6040.536ETTm1960.3240.3620.3710.3900.3750.3990.3640.3851920.3680.3800.4050.4050.4180.4150.4020.4033360.3990.4180.4420.4270.4490.4460.4310.4267200.4780.4550.4810.4620.4980.4720.4950.462Avg0.3920.4040.4250.4210.4350.4330.4230.419ETTm2960.1780.2640.1950.2910.2050.3040.2150.3061920.2410.3050.2640.3300.2960.3780.3120.3813360.3090.3380.3200.3650.3940.4410.4050.4517200.4020.4010.4310.4210.5610.5480.5850.551Avg0.2820.3270.3030.3520.3640.4180.3790.422Weather960.1700.2160.1950.2410.2160.2840.2150.2751920.2310.2620.2410.2750.2510.3210.2550.3143360.2780.2950.2990.3220.3050.3580.3040.3597200.3490.3410.3710.3670.3740.4010.3680.399Avg0.2570.2790.2770.3010.2870.3410.2860.337Electricity960.1400.2350.2150.3220.2020.3050.2230.3141920.1620.2520.2260.3440.2150.3160.2310.3263360.1750.2660.2340.3510.2240.3250.2460.3437200.2190.3110.2690.3780.2380.3410.2670.374Avg0.1740.2660.2360.3490.2200.3220.2420.339 Table 3Ablation study resultsDatasetHorizonAFNetAFNet(FFT)AFNet(FixedInterp)AFNet(SingleConv)MSEMAEMSEMAEMSEMAEMSEMAEETTh1960.3820.4020.4550.4470.4340.4580.3990.4101920.4310.4290.5100.4770.4650.4740.4550.4383360.4720.4520.5450.4960.5050.4980.4900.4577200.4870.4860.5900.5180.5350.5380.5160.516Avg0.4430.4420.5250.4850.4850.4920.4650.455ETTh2960.2920.3390.3280.3700.3850.4350.3610.4011920.3970.4080.4150.4250.5030.5060.4470.4923360.4230.4310.4310.4510.6190.5710.6210.5647200.4370.4450.4520.4750.8360.6680.9850.685Avg0.3870.4060.4070.4300.5860.5450.6040.536ETTm1960.3240.3620.3710.3900.3750.3990.3640.3851920.3680.3800.4050.4050.4180.4150.4020.4033360.3990.4180.4420.4270.4490.4460.4310.4267200.4780.4550.4810.4620.4980.4720.4950.462Avg0.3920.4040.4250.4210.4350.4330.4230.419ETTm2960.1780.2640.1950.2910.2050.3040.2150.3061920.2410.3050.2640.3300.2960.3780.3120.3813360.3090.3380.3200.3650.3940.4410.4050.4517200.4020.4010.4310.4210.5610.5480.5850.551Avg0.2820.3270.3030.3520.3640.4180.3790.422Weather960.1700.2160.1950.2410.2160.2840.2150.2751920.2310.2620.2410.2750.2510.3210.2550.3143360.2780.2950.2990.3220.3050.3580.3040.3597200.3490.3410.3710.3670.3740.4010.3680.399Avg0.2570.2790.2770.3010.2870.3410.2860.337Electricity960.1400.2350.2150.3220.2020.3050.2230.3141920.1620.2520.2260.3440.2150.3160.2310.3263360.1750.2660.2340.3510.2240.3250.2460.3437200.2190.3110.2690.3780.2380.3410.2670.374Avg0.1740.2660.2360.3490.2200.3220.2420.339 Ablation study results Model efficiency We compare AFNet with PatchTST20, FEDFormer14, AutoFormer13, and MICN29 in parameter count and multiply-accumulate operations (MACs) for identical (T,F)=(96,96) and batch size 32. The goal is relative cost, not a claim of Pareto optimality. Table 4 shows AFNet uses far fewer parameters and smaller MAC tallies than the listed Transformer and CNN baselines while occupying the same accuracy band as Table 2. On the six datasets the AFNet parameter count stays near \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$5.4\times 10^5$$\end{document}, roughly 6.9× smaller than the smallest PatchTST row (3.75M) and about 46.5× smaller than MICN (25.20M) on the ETT-style rows. MACs are 4.419G on the four ETT datasets in the table, below the PatchTST/FEDformer/Autoformer/MICN entries on those rows. Weather and Electricity read 13.26G and 20.26G and should be read next to their higher input widths. Relative to the same profiling setup, AFNet lowers MACs by roughly 25.5–93.8% depending on the dataset and competitor; the percentages are empirical estimates for the stated (T, F) and batch and will move if implementations or hardware change. Attention layers still scale quadratically in T, so the compact parameter count does not imply linear-time attention.Table 4Model efficiency comparison across different datasets.DatasetPatchTSTMICNFEDFormerAutoFormerAFNetParamsMACsParamsMACsParamsMACsParamsMACsParamsMACsETTh13.75M5.90G25.20M71.95G23.38M24.96G10.54M22.82G542.2K4.419GETTh210.06M17.66G25.20M71.95G23.38M24.96G10.54M22.82G542.2K4.419GETTm13.75M5.90G25.20M71.95G23.38M24.96G10.54M22.82G542.2K4.419GETTm210.06M17.66G25.20M71.95G23.38M24.96G10.54M22.82G542.2K4.419GWeather6.90M35.30G11.03M29.50G23.45M25.23G10.61M23.08G542.2K13.26GElectricity6.90M539.38G6.64M19.5G24.99M30.89G12.14M28.75G544.1K20.26G Table 4Model efficiency comparison across different datasets.DatasetPatchTSTMICNFEDFormerAutoFormerAFNetParamsMACsParamsMACsParamsMACsParamsMACsParamsMACsETTh13.75M5.90G25.20M71.95G23.38M24.96G10.54M22.82G542.2K4.419GETTh210.06M17.66G25.20M71.95G23.38M24.96G10.54M22.82G542.2K4.419GETTm13.75M5.90G25.20M71.95G23.38M24.96G10.54M22.82G542.2K4.419GETTm210.06M17.66G25.20M71.95G23.38M24.96G10.54M22.82G542.2K4.419GWeather6.90M35.30G11.03M29.50G23.45M25.23G10.61M23.08G542.2K13.26GElectricity6.90M539.38G6.64M19.5G24.99M30.89G12.14M28.75G544.1K20.26G Model efficiency comparison across different datasets. Conclusion AFNet stacks learnable depthwise filter banks with cross-scale attention so that rigid FFT splits are not required. The three trainable blocks are WAD with branch-wise mixing, AMSFF with self-attention and gated multi-scale convolutions, and CSAF with a fixed low-to-high cross-attention pass. Table 2 compares AFNet with nine baselines, including iTransformer and Pathformer rerun under the same protocol as AFNet (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$T{=}96$$\end{document}, horizons \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\{96,192,336,720\}$$\end{document}; official implementations22,21). On Electricity, iTransformer attains the lowest tabulated dataset-average MSE (0.170), with AFNet at 0.174 and a much smaller parameter footprint (Table 4), so the design trades a small average gap for compactness on this benchmark. On ETTh1, dataset-average MSE for AFNet (0.443) is close to Pathformer (0.439) and better than iTransformer (0.454) and DLinear (0.461). On ETTh2, Pathformer leads on the dataset average (0.344); AFNet (0.387) tracks iTransformer (0.383) and PatchTST (0.391) and degrades less toward \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$F{=}720$$\end{document} than unstable baselines such as MICN. On Weather, Pathformer and FreTS average lower MSE than AFNet (0.239 and 0.245 versus 0.257). Ablations in Table 3 show that removing WAD, AMSFF, or CSAF hurts accuracy under the same training recipe. Appendix A plots diagnostics from one ETTh1 checkpoint at \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$T{=}F{=}96$$\end{document} (WAD frequency responses, CSAF attention maps, and FFT–WAD overlays for selected channels). Main results notes how a shared \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$T{=}96$$\end{document} lookback differs from longer histories used elsewhere. Main results cites TimesNet, TimeMixer, and related models where their published protocols differ from ours. Claims stay at the level of empirical engineering for the benchmarks listed above.