MetaEformer: Unveiling and Leveraging Meta-patterns for Complex and Dynamic Systems Load Forecasting

Lanjun Wang; Shaoyuan Huang; Tiancheng Zhang; Xiaofei Wang; Xin Wang; Zhongtian Zhang

REVIEW 3 major objections 5 minor 1 cited by

MetaEformer: Unveiling and Leveraging Meta-patterns for Complex and Dynamic Systems Load Forecasting

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read MetaEformer claims load forecasts improve 37% by reusing a pool of fundamental waveform primitives.

desk verdict A clever architecture with released code, but the purification threshold in Eq. 5 is likely miscalibrated, so the central mechanism needs validation before the SOTA claim can be taken at face value. read the letter →

arxiv 2506.12800 v1 pith:3OG246JL submitted 2025-06-15 cs.LG

Shaoyuan Huang , Tiancheng Zhang , Zhongtian Zhang , Xiaofei Wang , Lanjun Wang , Xin Wang This is my paper

classification cs.LG

keywords timeseriesforecastingsystemloadmeta-patternpoolEchomechanismconceptdriftfew-shotlearningtransformer

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

MetaEformer is designed to settle a three-way conflict in system load forecasting: complex patterns, concept drift, and few-shot new entities each confuse current methods, and no single method handles all three at once. The paper's claim is that these problems share one cause, namely that predictors work on whole series rather than on the recurring waveform units from which series are built, and that making those units explicit fixes all three together. It introduces a Meta-pattern Pool that distills load series into a small set of fundamental waveforms and an Echo mechanism that selects and recombines the most relevant waveforms for each forecast. The paper reports a 37% relative error reduction over fifteen baselines on eight load benchmarks from cloud, power, and traffic systems. If true, one architecture covers the two most operationally painful cases, forecasting after a system entity switches behavior and forecasting for brand-new entities with little data, without sacrificing end-to-end training or interpretability.

What carries the argument

The two load-bearing objects are the Meta-pattern Pool (MPP) and the Echo mechanism. The MPP is a learnable library of P waveform primitives of length s, obtained by decomposing the load into seasonal and trend parts, slicing the seasonal part into short waveforms, merging waveforms whose alignment-boosted dot-product similarity exceeds a dynamically computed threshold, and updating the pool at regular intervals. The Echo mechanism is the counterpart that does the forecasting work: for each sliced waveform of the encoder's output, it computes similarity against the pool, selects the Top-K meta-patterns, weights them by similarity, and concatenates the reconstruction back into the representation. Together they instantiate the paper's core claim that recurring waveform primitives, rather than whole-series statistics, are the right unit for pattern recognition, concept-drift adaptation, and few-shot reconstruction.

What would settle it

Construct a test segment from waveforms deliberately absent from the training data, such as synthetic chirps or one-off transients stitched into otherwise normal loads, freeze the pool before that segment, and compare MetaEformer against the same transformer without the Echo mechanism: if Echo does not beat the no-Echo variant on these out-of-pool shapes, the central claim fails.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that load forecasting can be reorganized around meta-patterns: fundamental waveforms that are finer-grained than classical seasonal components and are meant to act like wavelet bases for system loads. The paper argues that all load patterns and their variations are compositions of these primitives, so it builds a Meta-pattern Pool that extracts seasonal components, slices them into short waveforms, merges similar ones by a similarity-weighted purification, and updates the pool as new data arrive. Forecasting then proceeds by exposing the current load to an Echo mechanism that selects the Top-K closest pool waveforms and uses them to reconstruct the load's representation, while Echo Padding fills the decoder with meta-pattern reconstructions instead of zeros. The stated consequence is that complex patterns are recognized through adaptive matching, concept drift is handled by switching to post-transition pool entries, and few-shot entities inherit precision from accumulated similar waveforms. The experimental claim is consistent state-of-the-art performance across three system scenarios, with the largest gains on the most dynamic datasets.

Load-bearing premise

The load-bearing premise is that every system load, including sudden switches and brand-new entities, is built from a finite set of recurring waveform shapes that persist across time and entities and fit inside the pool; if a load contains genuinely novel shapes outside the pool, the Echo reconstruction has nothing to fetch.

Editorial extensions

If this is right

If the reported results hold, MetaEformer becomes a single model that outperforms each of the four baseline classes in its own specialty, with roughly 20-66% lower MSE depending on class and scenario.
In dynamic systems, the advantage should widen as dynamics intensify; the paper shows the lead over DynEformer growing from 1% on ECW to 20% on CBW while baseline errors roughly double.
For few-shot and concept-drift deployments, the model should work without retraining on the new entity, since the Switch and New App test sets are never shown to the model during training.
Performance should remain stable across a wide range of Echo Top-K values and slice lengths, with optimal MPP size scaling with the scenario's dynamism rather than with dataset size.
The added computational cost should stay negligible relative to the transformer core, making the accuracy gains available without a major efficiency penalty.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Because the pool is built from a model's own training series and updated online, the framework implies that a sufficiently rich pool learned on one system class could be reused as initialization for a new entity or even a new domain; the paper does not test this cross-dataset transfer, so it is left as an open extension.
The purification threshold couples pool size P to pattern separation, suggesting an operational rule the paper leaves implicit: choose P by measuring scenario dynamism, consistent with the paper's observation that the most dynamic cloud scenarios need the largest P.
The Echo's similarity-weighted recombination of Top-K primitives is essentially dictionary-style reconstruction; a clean test would compare MetaEformer against a fixed wavelet or learned dictionary of the same capacity to isolate whether purification and adaptive pooling, rather than reconstruction alone, drive the gains.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

You should know about this paper: it has a genuinely new architecture for load forecasting, with released code and an extensive benchmark, but the central purification mechanism has a scale bug that makes it likely not work as described. The empirical SOTA claim might still hold, but the paper does not currently establish that the proposed meta-pattern purification is responsible.

The good parts: MetaEformer wraps a learnable codebook of waveform primitives (the MPP) and a reconstruction layer (Echo) around a transformer. The idea of reusing fundamental waveforms across entities and over time is intuitively appealing, and the end-to-end design is clean. The evaluation is thorough: eight datasets across cloud, power, and traffic, fifteen baselines, and the use of ADF tests to select non-stationary series is thoughtful. Code on Zenodo is a real plus.

The soft spots, in order of severity. The threshold in Eq. 5 is miscalibrated under the reported hyperparameters. With B=256, L=48, s=16, the waveform matrix has 768 rows, so the upper-triangular similarity matrix has about 294k entries. With P=650 and alpha=0.5, alpha*P/|SM| is about 0.001. So tau sits essentially at the mean of the similarity distribution. For z-scored waveforms of length 16, the row-wise maximum over ~767 similarities will almost always be far above this threshold, meaning Algorithm 1 will merge nearly all waveforms into a single weighted pattern. That is the opposite of the stated goal of separating heterogeneous patterns. The threshold is also highly sensitive to batch size, which changes |SM| quadratically, and no sensitivity analysis is reported. The De-MPP ablation removes the whole module, so it cannot distinguish a functioning purifier from a randomly initialized codebook. This is a load-bearing flaw because the paper's story and interpretability claims depend on the purification mechanism.

Secondary issues: all experiments are repeated three times but only point estimates are reported, so the 37% headline number could be partly noise. One baseline (SAN) collapses on the few-shot sets, inflating the aggregates. The standardization inside the similarity measure is underspecified, so the behavior cannot be reproduced from the text.

My take: this deserves a serious referee, but the referee should demand a sensitivity analysis of tau, error bars, and an ablation that actually isolates the purification step (e.g., varying tau or fixing it to a meaningless value). As it stands, the central mechanism is not tied to the empirical gains. If the code confirms the threshold behavior, the story needs substantial revision. Still worth engaging.

Referee Report

3 major / 5 minor

Summary. The paper proposes MetaEformer, a transformer-based load forecasting model that introduces two novel components: a Meta-pattern Pooling mechanism that decomposes input series, extracts sliced seasonal components, measures pairwise waveform similarity, and uses a threshold to purify and maintain a pool of 'meta-patterns'; and an Echo mechanism that, for each input waveform, selects the Top-K similar meta-patterns and reconstructs the representation, plus an Echo Padding module that fills decoder inputs with meta-pattern reconstructions. The authors claim consistent state-of-the-art performance across eight curated load datasets from cloud, power, and traffic scenarios, with a 37% relative improvement over fifteen baselines. The paper includes ablations, sensitivity studies, efficiency comparisons, and interpretability visualizations, and the code is publicly released.

Significance. If the empirical results hold, MetaEformer would be an interesting contribution to system load forecasting, particularly for its explicit treatment of concept drift and few-shot scenarios, and its attempt to provide an interpretable, end-to-end alternative to clustering-based methods. The paper has several strengths: the code is available, the ablation study is systematic, and the visualizations of the meta-pattern pool and echo selection are useful. However, the central mechanism's operational definition—specifically the purification threshold—is ambiguous and its numerical behavior is not verified, and the SOTA claim is presented without statistical variance, which lowers confidence in the headline improvements.

major comments (3)

[§4.1.2, Eq. (5), Algorithm 1]
[§5.1, §5.2, Table 2]
[Abstract, §5.2]

minor comments (5)

[Algorithm 1, Eq. (5)]
[§4.2.2, Eq. (12)]
[§5.4]
[Table 4]
[§5.2]

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: MetaEformer is an empirical architecture with externally evaluated forecasting benchmarks; no claim reduces to its constructor inputs or to self-citations.

full rationale

The paper's central claim is a forecasting accuracy improvement, not a derived mathematical prediction. The meta-pattern pool is constructed from the seasonal components of the first training batch (Algorithm 1) and updated from later training batches (Algorithm 2); the Echo Layer then selects Top-K similar patterns via Eq. 3 and recombines them via Eq. 12. None of these quantities is defined in terms of the forecast target Y, and the reported MSE/MAE differences are not logically entailed by the construction. P and K are tuned on held-out validation and reported as sensitivity curves (Fig. 5), so the optimal settings are not fitted quantities renamed as predictions. The only same-group citation, DynEformer [17], is used as a baseline and as the source of the ECW/Switch/New App datasets in Section 5; it does not supply any uniqueness premise or theorem that forces MetaEformer's design. The threshold in Eq. 5 is an ad hoc heuristic whose scale dependence may be a robustness concern, but that is a correctness or calibration issue, not circularity: the claimed superiority is not entailed by Eq. 5. The paper is self-contained against external benchmarks, so the appropriate finding is no significant circularity.

Assumptions & free parameters 5 free parameters · 4 assumptions · 1 invented entities

The model's effectiveness rests on the assumption of decomposable recurring waveforms, and on a handful of tuned hyperparameters (P, K, s, gamma, alpha). No new physical entities are introduced; 'meta-pattern' is an internal latent construct. The pool and echo are the load-bearing components.

free parameters (5)

Meta-pattern pool size P = 650 (CBW, ECW), 350 (Traffic); scenario-dependent
Tuned in §5.4 Fig. 5a; larger P for more dynamic scenarios.
Top-K (number of meta-patterns selected in Echo) = not specified in final config; robust across 10-330
Sensitivity analysis in Fig. 5b shows stability; exact per-dataset value is not reported.
Slice length s = 16
Fixed waveform length; robustness shown in Fig. 7.
MPP update rate gamma = 0.1
EMA coefficient in Algorithm 2; fixed without reported tuning.
Threshold scaling alpha = 0.5
Constant in Eq. 5; chosen by hand to balance P/|SM| scaling.

assumptions (4)

domain assumption All load patterns and their variations in a system are composed of fundamental waveforms (meta-patterns).
Stated in Section 1 and motivates the MPP; if loads do not share reusable waveform primitives, the pooling and Echo mechanisms cannot help.
ad hoc to paper A plain dot-product between standardized waveforms captures shape similarity and phase information.
Eq. 3 defines Sim as a dot product, while the text claims standardization and phase retention without giving the actual standardization procedure; this assumption is specific to the paper's design.
ad hoc to paper The threshold tau in Eq. 5, with alpha fixed at 0.5, correctly separates similar and distinct waveforms.
No theoretical justification is given for Eq. 5; the purification quality depends on this formula.
domain assumption ADF-based selection of datasets identifies scenarios where the proposed model will outperform generic baselines.
Datasets are chosen for non-stationarity/dynamics (Section 5); the paper assumes these selections represent the target deployment conditions.

invented entities (1)

Meta-pattern
purpose: Fundamental waveform units used by the Meta-Pattern Pool and Echo to reconstruct load patterns
Defined operationally via the pooling algorithm; no external falsifiable measurement.

how reviews work

0 comments

Cite this review

Pith. "Pith review of MetaEformer: Unveiling and Leveraging Meta-patterns for Complex and Dynamic Systems Load Forecasting." pith.science (2026). https://pith.science/paper/3OG246JL

@misc{pith2026250612800,
  author       = {Pith},
  title        = {Pith review of: MetaEformer: Unveiling and Leveraging Meta-patterns for Complex and Dynamic Systems Load Forecasting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3OG246JL}},
  note         = {Machine review of arXiv:2506.12800}
}

read the original abstract

Time series forecasting is a critical and practical problem in many real-world applications, especially for industrial scenarios, where load forecasting underpins the intelligent operation of modern systems like clouds, power grids and traffic networks.However, the inherent complexity and dynamics of these systems present significant challenges. Despite advances in methods such as pattern recognition and anti-non-stationarity have led to performance gains, current methods fail to consistently ensure effectiveness across various system scenarios due to the intertwined issues of complex patterns, concept-drift, and few-shot problems. To address these challenges simultaneously, we introduce a novel scheme centered on fundamental waveform, a.k.a., meta-pattern. Specifically, we develop a unique Meta-pattern Pooling mechanism to purify and maintain meta-patterns, capturing the nuanced nature of system loads. Complementing this, the proposed Echo mechanism adaptively leverages the meta-patterns, enabling a flexible and precise pattern reconstruction. Our Meta-pattern Echo transformer (MetaEformer) seamlessly incorporates these mechanisms with the transformer-based predictor, offering end-to-end efficiency and interpretability of core processes. Demonstrating superior performance across eight benchmarks under three system scenarios, MetaEformer marks a significant advantage in accuracy, with a 37% relative improvement on fifteen state-of-the-art baselines.

Figures

Figures reproduced from arXiv: 2506.12800 by the authors.

**Figure 2.** Framework overview of MetaEformer, consisting of two main parts: Meta-pattern Pooling (§4.1) and Forecasting (§4.2). [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Waveform similarity measurement. where 𝑤𝑥𝑖 denote the 𝑖-th value of W𝑥 . This method suggests that congruent peaks will lead to a higher Sim, facilitating pattern matching, while disparate patterns result in lower Sim due to peak and trough misalignment. Sim enhances interpretability and is computationally efficient, making it ideal for end-to-end models. Moreover, due to the incorporating of standardization, the m… view at source ↗

Figures from the paper (7 more)

**Figure 4.** Figure 4: Echo Layer. The reconstructing is to reintegrate the identified meta-patterns as new features in load series, which can be formalized as follows: 𝑂𝑖 = Softmax (𝐷𝑅(𝑊𝑖)) · (P𝑖 𝑠𝑒𝑙𝑒𝑐𝑡 ) 𝑇 ∈ R 𝐵×𝑠×𝐾 (12) O = Concat (𝐷𝐼(𝑂1, · · · ,𝑂𝑛)) ∈ R 𝐵×𝐿× 1 2 𝑑𝑚𝑜𝑑𝑒𝑙 (13) 𝑂 𝑒𝑛 = Concat…

**Figure 5.** Figure 5: Performance of MetaEformer at different MPP size [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

**Figure 6.** Figure 6: Efficiency of Transformer-based models. 6 Conclusion This paper presents MetaEformer, a model built on the novel paradigm of meta-patterns, designed to forecast complex and dynamic loads in various real-world systems, addressing the challenges of complex patterns, con…

**Figure 7.** Figure 7: Prediction performance with varying slice lengths. [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]

**Figure 8.** Figure 8: Prediction cases from the Switch, New App and CBW dataset. [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]

**Figure 9.** Figure 9: Visualization of MPP construction and updates. The first heatmap represents the original values of the MPP, and the [PITH_FULL_IMAGE:figures/full_fig_p012_9.png]

**Figure 10.** Figure 10: Displays the Echo mechanism for the input series (waveform) in every 20 batches during training on the CBW. [PITH_FULL_IMAGE:figures/full_fig_p012_10.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

External Data-Enhanced Meta-Representation for Adaptive Probabilistic Load Forecasting
cs.LG 2025-06 conditional novelty 4.0 of 10

M2oE2 uses hypernetworks and a mixture-of-experts gate to let external conditions reshape a GRU's input weights, reporting large gains in load forecasting error metrics.

Reference graph

Works this paper leans on

50 extracted references · 42 canonical work pages · cited by 1 Pith paper

[1]

Chao Chen, Karl Petty, Alexander Skabardonis, Pravin Varaiya, and Zhanfeng Jia. 2001. Freeway performance measurement system: mining loop detector data. Transportation Research Record1748, 1 (2001), 96–102

work page 2001
[2]

Mouxiang Chen, Lefei Shen, Han Fu, Zhuo Li, Jianling Sun, and Chenghao Liu

work page
[3]

Guillaume Chevillon. 2007. Direct multi-step estimation and forecasting.Journal of Economic Surveys21, 4 (2007), 746–785

work page 2007
[4]

Robert B Cleveland, William S Cleveland, Jean E McRae, and Irma Terpenning
[5]

Jovani Dalzochio, Rafael Kunst, Edison Pignaton, Alecio Binotto, Srijnan Sanyal, Jose Favilla, and Jorge Barbosa. 2020. Machine learning and reasoning for predic- tive maintenance in Industry 4.0: Current status and challenges.Computers in Industry123 (2020), 103298

work page 2020
[6]

Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xi- aohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al . 2020. An Image is Worth 16x16 Words: Trans- formers for Image Recognition at Scale. InInternational Conference on Learning Representations (ICLR)

work page 2020
[7]

Vijay Ekambaram, Arindam Jati, Nam Nguyen, Phanwadee Sinthong, and Jayant Kalagnanam. 2023. Tsmixer: Lightweight mlp-mixer model for multivariate time series forecasting. InProceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 459–469

work page 2023
[8]

Graham Elliott, Thomas J Rothenberg, and James H Stock. 1992. Efficient tests for an autoregressive unit root

work page 1992

Show all 50 references

[9]

Chen et al. 2022. Learning to Rotate: Quaternion Transformer for Complicated Periodical Time Series Forecasting. InProceedings of the 28th ACM SIGKDD (KDD ’22). 146–156

2022
[10]

Hugo et al. 2023. An embedding-based non-stationary fuzzy time series method for multiple output high-dimensional multivariate time series forecasting in IoT applications.Neural Computing and Applications35 (2023), 9407–9420

2023
[11]

Lukas et al. 2022. Few-shot forecasting of time-series with heterogeneous chan- nels. InJoint European Conference on Machine Learning and Knowledge Discovery in Databases. Springer, 3–18

2022
[12]

Liu et al. 2024. UniTime: A Language-Empowered Unified Model for Cross- Domain Time Series Forecasting. InACM WWW 2024

2024
[13]

Wei Fan, Pengyang Wang, Dongkun Wang, Dongjie Wang, Yuanchun Zhou, and Yanjie Fu. 2023. Dish-TS: A General Paradigm for Alleviating Distribution Shift in Time Series Forecasting. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 37. 7522–7529

2023
[14]

Alberto Gasparin, Slobodan Lukovic, and Cesare Alippi. 2022. Deep learning for time series forecasting: The electric load case.CAAI Transactions on Intelligence Technology7, 1 (2022), 1–25

2022
[15]

Xiao Han, Xiangyu Zhao, Liang Zhang, and Wanyu Wang. 2023. Mitigating Action Hysteresis in Traffic Signal Control with Traffic Predictive Reinforcement Learning. InProceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD ’23). 673–684

2023
[16]

Rui Hu, Jie Chen, and Li Zhou. 2022. A transformer-based deep neural network for arrhythmia detection using continuous ECG signals.Computers in biology and medicine144 (2022), 105325

2022
[17]

Shaoyuan Huang, Zheng Wang, Heng Zhang, Xiaofei Wang, Cheng Zhang, and Wenyu Wang. 2023. One for All: Unified Workload Prediction for Dynamic Multi-Tenant Edge Cloud Platforms. InProceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD ’23). 788–797

2023
[18]

Shaoyuan Huang, Heng Zhang, Xiaofei Wang, Min Chen, Jianxin Li, and Victor C. M. Leung. 2024. Fine-Grained Spatio-Temporal Distribution Prediction of Mobile Content Delivery in 5G Ultra-Dense Networks.IEEE TMC23, 1 (2024)

2024
[19]

Vinodh Kumaran Jayakumar, Jaewoo Lee, In Kee Kim, and Wei Wang. 2020. A Self-Optimized Generic Workload Prediction Framework for Cloud Computing. In2020 IEEE IPDPS. IEEE, 779–788

2020
[20]

Yilun Jin, Kai Chen, and Qiang Yang. 2022. Selective Cross-City Transfer Learning for Traffic Prediction via Source City Region Re-Weighting. InProceedings of the 28th ACM SIGKDD Conference (KDD ’22). 731–741

2022
[21]

In Kee Kim, Wei Wang, Yanjun Qi, and Marty Humphrey. 2018. CloudInsight: Utilizing a Council of Experts to Predict Future Cloud Application Workloads. In 2018 IEEE 11th International Conference on Cloud Computing (CLOUD). 41–48

2018
[22]

Taesung Kim, Jinhee Kim, Yunwon Tae, Cheonbok Park, Jang-Ho Choi, and Jaegul Choo. 2021. Reversible Instance Normalization for Accurate Time-Series Forecasting against Distribution Shift. InICLR

2021
[23]

Diederik P Kingma and Jimmy Ba. 2014. Adam: A method for stochastic opti- mization.arXiv preprint arXiv:1412.6980(2014)

2014 arXiv
[24]

Shiyang Li, Xiaoyong Jin, Yao Xuan, Xiyou Zhou, Wenhu Chen, Yu-Xiang Wang, and Xifeng Yan. 2019. Enhancing the locality and breaking the memory bottleneck of transformer on time series forecasting.Neurips32 (2019)

2019
[25]

Yaguang Li, Rose Yu, Cyrus Shahabi, and Yan Liu. 2018. Diffusion Convolutional Recurrent Neural Network: Data-Driven Traffic Forecasting. InInternational Conference on Learning Representations (ICLR)

2018
[26]

Yanan Li, Haitao Yuan, Zhe Fu, Xiao Ma, Mengwei Xu, and Shangguang Wang
[27]

Arık, Nicolas Loeff, and Tomas Pfister

Bryan Lim, Sercan Ö. Arık, Nicolas Loeff, and Tomas Pfister. 2021. Temporal Fusion Transformers for interpretable multi-horizon time series forecasting. International Journal of Forecasting37, 4 (2021), 1748–1764

2021
[28]

Yong Liu, Chenyu Li, Jianmin Wang, and Mingsheng Long. 2023. Koopa: Learn- ing Non-stationary Time Series Dynamics with Koopman Predictors. InThirty- seventh Conference on Neural Information Processing Systems (NeurIPS)

2023
[29]

Yong Liu, Haixu Wu, Jianmin Wang, and Mingsheng Long. 2022. Non-stationary transformers: Exploring the stationarity in time series forecasting.Advances in Neural Information Processing Systems (NeurIPS)35 (2022), 9881–9893

2022
[30]

Zhiding Liu, Mingyue Cheng, Zhi Li, Zhenya Huang, Qi Liu, Yanhu Xie, and Enhong Chen. 2023. Adaptive Normalization for Non-stationary Time Series Forecasting: A Temporal Slice Perspective. InThirty-seventh NeurIPS. KDD ’25, August 3–7, 2025, Toronto, ON, Canada. Shaoyuan Huang et al

2023
[31]

Jie Lu, Anjin Liu, Fan Dong, Feng Gu, João Gama, and Guangquan Zhang. 2019. Learning under Concept Drift: A Review.IEEE Transactions on Knowledge and Data Engineering (TKDE)31, 12 (2019), 2346–2363

2019
[32]

Nguyen, Phanwadee Sinthong, and Jayant Kalagnanam

Yuqi Nie, Nam H. Nguyen, Phanwadee Sinthong, and Jayant Kalagnanam. 2023. A Time Series is Worth 64 Words: Long-term Forecasting with Transformers. In International Conference on Learning Representations

2023
[33]

M Sifuzzaman, M Rafiq Islam, and Mostafa Z Ali. 2009. Application of wavelet transform and its advantages compared to Fourier transform. (2009)

2009
[34]

Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Ł ukasz Kaiser, and Illia Polosukhin. 2017. Attention is All you Need. InAdvances in Neural Information Processing Systems (NeurIPS), Vol. 30

2017
[35]

Qingsong Wen, Linxiao Yang, Tian Zhou, and Liang Sun. 2022. Robust Time Series Analysis and Applications: An Industrial Perspective. InProceedings of the 28th ACM SIGKDD (KDD ’22). 4836–4837

2022
[36]

Gerald Woo, Chenghao Liu, Doyen Sahoo, Akshat Kumar, and Steven Hoi. 2022. Etsformer: Exponential smoothing transformers for time-series forecasting.arXiv preprint arXiv:2202.01381(2022)

2022 arXiv
[37]

Haixu Wu, Tengge Hu, Yong Liu, Hang Zhou, Jianmin Wang, and Mingsheng Long. 2022. Timesnet: Temporal 2d-variation modeling for general time series analysis.arXiv preprint arXiv:2210.02186(2022)

2022 arXiv
[38]

Haixu Wu, Jiehui Xu, Jianmin Wang, and Mingsheng Long. 2021. Autoformer: De- composition transformers with auto-correlation for long-term series forecasting. NeurIPS34 (2021), 22419–22430

2021
[39]

Eunju Yang and Chan Hyun Youn. 2023. Temporal Data Pooling With Meta- Initialization for Individual Short-Term Load Forecasting.IEEE Transactions on Smart Grid14, 4 (2023), 3246–3258

2023
[40]

Kun Yi, Jingru Fei, Qi Zhang, Hui He, Shufeng Hao, Defu Lian, and Wei Fan
[41]

Kun Yi, Qi Zhang, Wei Fan, Shoujin Wang, Pengyang Wang, Hui He, Ning An, Defu Lian, Longbing Cao, and Zhendong Niu. 2024. Frequency-domain MLPs are more effective learners in time series forecasting.Neurips36 (2024)

2024
[42]

Ailing Zeng, Muxi Chen, Lei Zhang, and Qiang Xu. 2023. Are transformers effective for time series forecasting?. InProceedings of the AAAI conference on artificial intelligence, Vol. 37. 11121–11128

2023
[43]

Yuchen Zhang, Mingsheng Long, Kaiyuan Chen, Lanxiang Xing, Ronghua Jin, Michael I Jordan, and Jianmin Wang. 2023. Skilful nowcasting of extreme precip- itation with NowcastNet.Nature619, 7970 (2023), 526–532

2023
[44]

InThe Thirty-eighth Annual Conference on Neural Information Processing Systems

FilterNet: Harnessing Frequency Filters for Time Series Forecasting. InThe Thirty-eighth Annual Conference on Neural Information Processing Systems
[45]

Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wancai Zhang. 2021. Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting.AAAI 202112B (2021), 11106–11115

2021
[46]

Tian Zhou, Ziqing Ma, Qingsong Wen, Xue Wang, Liang Sun, and Rong Jin. 2022. Fedformer: Frequency enhanced decomposed transformer for long-term series forecasting. InICML. PMLR, 27268–27286. A Appendix A.1 Meta-pattern Pool Update Algorithm The process of meta-pattern pool upd...

2022
[48]

Yunhao Zhang and Junchi Yan. 2023. Crossformer: Transformer utilizing cross- dimension dependency for multivariate time series forecasting. InThe eleventh international conference on learning representations

2023
[1990]

STL: A seasonal-trend decomposition.J. Off. Stat6, 1 (1990), 3–73

1990
[2023]

InACM WWW 2023 (WWW ’23)

ELASTIC: Edge Workload Forecasting Based on Collaborative Cloud-Edge Deep Learning. InACM WWW 2023 (WWW ’23). 3056–3066

2023
[2024]

Calibration of Time-Series Forecasting: Detecting and Adapting Context- Driven Distribution Shift.ACM SIGKDD(2024), 341–352

2024

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Chao Chen, Karl Petty, Alexander Skabardonis, Pravin Varaiya, and Zhanfeng Jia. 2001. Freeway performance measurement system: mining loop detector data. Transportation Research Record1748, 1 (2001), 96–102

work page 2001

[2] [2]

Mouxiang Chen, Lefei Shen, Han Fu, Zhuo Li, Jianling Sun, and Chenghao Liu

work page

[3] [3]

Guillaume Chevillon. 2007. Direct multi-step estimation and forecasting.Journal of Economic Surveys21, 4 (2007), 746–785

work page 2007

[4] [4]

Robert B Cleveland, William S Cleveland, Jean E McRae, and Irma Terpenning

[5] [5]

Jovani Dalzochio, Rafael Kunst, Edison Pignaton, Alecio Binotto, Srijnan Sanyal, Jose Favilla, and Jorge Barbosa. 2020. Machine learning and reasoning for predic- tive maintenance in Industry 4.0: Current status and challenges.Computers in Industry123 (2020), 103298

work page 2020

[6] [6]

Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xi- aohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al . 2020. An Image is Worth 16x16 Words: Trans- formers for Image Recognition at Scale. InInternational Conference on Learning Representations (ICLR)

work page 2020

[7] [7]

Vijay Ekambaram, Arindam Jati, Nam Nguyen, Phanwadee Sinthong, and Jayant Kalagnanam. 2023. Tsmixer: Lightweight mlp-mixer model for multivariate time series forecasting. InProceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 459–469

work page 2023

[8] [8]

Graham Elliott, Thomas J Rothenberg, and James H Stock. 1992. Efficient tests for an autoregressive unit root

work page 1992