Pith. sign in

REVIEW 4 major objections 3 minor 37 references

Detection of collective and point anomalies in the presence of trend and seasonality

T0 review · 4 major / 3 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read A new detector finds anomaly blocks even when the baseline has a trend and a seasonal cycle, and proves the decomposition consistent.

desk verdict Real methodological contribution with a serious theory-algorithm mismatch: Theorem 1 is proven only for a simplified least-squares, no-point-anomaly version, not the robust STAD defined in Section 2.6. read the letter →

arxiv 2508.21128 v3 pith:3L7Z3HFH submitted 2025-08-28 stat.ME math.STstat.APstat.TH

Yiyin Zhang , Florian Pein , Idris A. Eckley This is my paper · ORCID
classification stat.MEmath.STstat.APstat.TH MSC 62M1062F35
keywords anomalydetectioncollectiveanomaliestimeseriesdecompositiontrendandseasonalityrobustM-estimationchangepointpenalizedcostseasonal-trend
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

Most anomaly detectors assume a flat baseline, so trends and seasonal cycles are mistaken for anomalies. This paper proposes STAD, which models normal behavior as a polynomial trend plus a seasonal component and detects both collective anomalies (contiguous blocks) and point anomalies on top of that baseline. The central claim is that STAD decomposes a series into anomaly, trend, seasonality, and noise, estimates the number of collective anomalies consistently, locates start and end points within a small error, and bounds each component's squared error by a constant times sigma_0^2 log n. The method differenced away seasonality, samples subsamples that avoid anomalous blocks to estimate the trend robustly, re-estimates seasonality, and then runs a penalized-cost anomaly search. If correct, STAD removes the constant-baseline restriction that has limited collective-anomaly detection on real series such as energy prices.

What carries the argument

The carrying mechanism is a four-step pipeline. Lag-differencing with lag D = C times the seasonal period removes seasonality exactly, leaving a lower-degree polynomial trend plus a noise term with doubled variance. The trend is estimated by sampling many well-separated index sets, each drawing one subsegment from alternating blocks; with probability at least 1-(1/2)^J at least one sample avoids all collective anomalies, so a robust M-estimator with Tukey's biweight loss on that clean sample gives a good trend. Seasonality is estimated per phase by a robust median/M-estimator after excluding estimated anomalies. A penalized sum-of-squares cost, minimized by dynamic programming, then detects

What would settle it

Run the full STAD versus its simplified proof version (least squares, no point anomalies) on the same simulated series with point anomalies placed near an anomaly-block boundary; if the two versions disagree on the number or start/end of the collective block, Theorem 1 does not transfer to the shipped method. Also, a single collective anomaly longer than n/4 with magnitude near 1.2 already drops the paper's own simulated accuracy from roughly 0.9 to 0.5, so the claimed total-anomaly-length-of-order-n guarantee has a visible finite-sample boundary.

Watch

Extended reading notes

Core claim

STAD's core claim is Theorem 1: under assumptions bounding anomaly length, separation, and tuning parameters, the estimated anomaly set A-hat converges in probability to the true set up to location errors kappa_k, roughly (sigma_0^2 / mu_k^2) log(n)^{1+delta}, capped below by 4Q+1, and the squared L2 errors of the anomaly, trend, and seasonal estimates are all bounded by C3 sigma_0^2 log n with probability tending to 1. The estimate of the anomaly count K is consistent. This extends the guarantee of the penalized-cost anomaly detector from a constant baseline to a polynomial trend plus seasonality, while allowing the total length of anomalies to be of order n rather than sqrt n. The proof wo

Load-bearing premise

The load-bearing premise is that the full STAD algorithm behaves like the simplified version analysed in the proof: Theorem 1 is shown for a variant with no point anomalies and least-squares trend fitting, while the implemented estimator uses Tukey's biweight loss and also returns point anomalies; if those differences change the qualitative behaviour, the consistency and L2 guarantees do not formally cover the method as described.

Editorial extensions

If this is right

  • If Theorem 1 holds, a time series with a polynomial trend and a known-period seasonal component can be split into anomaly, trend, seasonality, and noise with per-component squared error O(sigma_0^2 log n).
  • The number of collective anomalies is estimated consistently, and start and end points are located within a small error, making anomaly counts and intervals trustworthy in nonstationary baselines.
  • The method tolerates a total anomaly length of order n rather than sqrt n, so long outages or extended price spikes no longer break the baseline estimate.
  • Point anomalies are handled by the same penalized-cost objective with a separate penalty, extending the framework beyond collective blocks, though the formal theorem is proved without them.
  • On the electricity price application, STAD finds a small set of anomalies that align with renewable-generation spikes, where a constant-baseline detector produces many false positives.

Reading between the lines

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

  • Beyond the paper: the structured subsampling trick could be reused for robust estimation of other structured components, such as periodic splines or changepoint models, whenever one can guarantee an anomaly-free subdesign with high probability.
  • Beyond the paper: since the theorem is proved with least squares and no point anomalies, a direct finite-sample comparison between full STAD and its simplified proof version on simulations with point anomalies near anomaly-block boundaries would show how much of the guarantee is actually carried by the robust M-estimator.
  • Beyond the paper: the location error kappa_k scales with sigma_0^2 log^{1+delta} / mu_k^2, so weak anomalies need longer blocks; an adaptive or multiscale penalty might sharpen this trade-off in practice.
  • Beyond the paper: the guarantees require a known seasonal period and the true polynomial degree in a finite candidate set, so unknown periods or non-polynomial trends would need an extended model-selection layer.
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.

Referee Report

4 major / 3 minor

Summary. The paper proposes STAD, an iterative method for detecting collective and point anomalies in time series that may contain a polynomial trend and a seasonal component. The procedure estimates the trend on lag-differenced data, estimates the seasonality robustly, refines both estimates, and finally applies a CAPA-type penalized cost to the residuals to identify anomalies. The main theoretical result, Theorem 1, claims asymptotic consistency of the number and locations of collective anomalies and an L2 error of order σ0^2 log n for the anomaly, trend, and seasonal estimates. Finite-sample performance is studied by simulation against oracle versions of STAD, and the method is illustrated on UK electricity export price data.

Significance. If the stated guarantees covered the actual algorithm, the contribution would be substantial: STAD would extend CAPA-type anomaly detection from a constant baseline to polynomial trend plus seasonality, allow collective anomalies longer than the √n barrier of earlier work, and provide a full additive decomposition with quantified error. The proof apparatus for the simplified estimator is extensive and the simulation study is carefully organized. However, the central theorem is proven for a version of the method that differs from the one defined in Section 2.6, and there are additional mismatches between the implemented objective and the objective used in the proofs. These issues currently limit the significance of the theoretical claims to a simplified, oracle-like variant.

major comments (4)
  1. [Section 3, paragraph before Theorem 1; Settings 1–3; Lemma 9, Eq. (41)] Theorem 1 is stated for 'the STAD estimators as defined in Section 2.6', but the paragraph immediately before the theorem replaces Tukey's M-estimator by least squares, assumes no point anomalies, and replaces the seasonal M-estimator by the initial median. All subsequent settings are consistent with this simplification: Setting 1 and Setting 3 explicitly assume O=∅, and Lemma 9 uses the closed-form LS projection X_U(X_U^T X_U)^{-1} X_U^T Y_U, which does not hold for the IRLS/Tukey estimator. Consequently, the abstract's claims of consistent estimation of the number and locations of anomalies and accurate decomposition are not established for the method as defined in Section 2.6. This is a load-bearing gap: point anomalies can break the LS trend estimate, and the nonconvex Tukey loss is not a small perturbation of least squares in the proof.
  2. [Section 2.2, Eq. (12) vs. Section D.4.4, Lemmas 13, 15, 16] The algorithm minimizes the penalized cost (12) with penalty \tilde K λ_coll \hatσ_0^2 log n (and (13) uses λ_point \hatσ_0^2 log n). The proof, however, repeatedly uses a penalty of the form (1/2) λ_coll \hatσ_0^2 log(n)^{1+δ} and, in the expansion in Lemma 16, \hat K λ_coll \hatσ_0 log(n)^{1+δ}. These are not the same objective as in Section 2.2. The lower bounds in Lemmas 13 and 15 are derived for the log^{1+δ} penalty; if the actual penalty is only log n, those inequalities do not follow as written. The authors need to reconcile the penalty in the theorem with the penalty implemented in the algorithm, or explicitly state that the guarantees apply to a modified estimator.
  3. [Assumption 2, Section 3; Assumption 7, Section D.5] As printed, Assumption 2 bounds the total anomaly length by γ / max(4(Q+1),8), which is < 1/8, and Assumption 7 bounds it by 1/8. For any nonempty set of collective anomalies satisfying Assumption 1, the total length is at least of order log^{1+δ+\tildeδ} n, so these inequalities are impossible for large n. Moreover, the discussion following Assumption 2 claims that the total anomaly length can be of order n, which is inconsistent with the displayed bound. Lemma 20's proof effectively uses a bound of the form ∑(e_k-b_k) ≤ n/8. A factor n appears to be missing from the displayed assumptions; without correction, the seasonality part of the proof is vacuous.
  4. [Section 4, Tables 1–4] The simulation study evaluates only collective anomalies and only against oracle versions of STAD. There is no scenario with point anomalies, so the point-anomaly component of the method is not empirically assessed. The only comparisons with existing methods are illustrative figures for CAPA and STL. Thus the finite-sample evidence does not substantiate the claimed advantage over existing methods, nor does it support the point-anomaly part of the abstract. Adding point-anomaly settings and, where feasible, systematic benchmarks against CAPA and decomposition-based baselines would strengthen the empirical claims.
minor comments (3)
  1. [Section D.7, proof of Theorem 1] The uniform bound after Eq. (74) is written as max_i |(\hat T_0 - T_D)_i| ≤ \tilde C_2 σ_0^2 log n. Dimensionally this should be of order σ_0 √(log n / n) (or similar) to feed into Setting 2's max |r_i| ≤ C_19 √(log n/n) σ_0. Please correct the displayed bound.
  2. [Throughout Appendix D] The proofs alternate between \hatσ_0 and \hatσ_0^2 in penalty terms and sometimes use log(n)^{1+δ} where the main text uses log n. Please harmonize notation so that the reader can trace the objective from Section 2.2 through the lemmas.
  3. [Section 4.2 and Section 5] Minor typos: 'vary he seasonal period' and 'tree anomalies' should be 'vary the seasonal period' and 'three anomalies'.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular dependency found: the theory establishes bounds for a simplified least-squares/no-point-anomaly version of STAD, which is a derivation gap rather than a circular reduction.

full rationale

The paper's derivation chain does not, at any load-bearing step, define its target in terms of its inputs or rename a fitted quantity as a prediction. The main theorem is proven under an explicitly stated simplification: the paragraph before Theorem 1 says 'we assume that we have no point anomalies and use the least squares estimator instead of Tukey's estimator' and that for seasonality 'we stop after the initial median.' This is a disclosed discrepancy between the analyzed estimator and the full §2.6 algorithm, and it is a legitimate correctness/completeness concern: the full robust IRLS/Tukey estimator and point anomalies are not covered by the proof. But this is not circularity. The proof does not assume the conclusion; it starts from the model assumptions, bounds the variance estimator, analyzes the differenced series, proves lower bounds on determinants of sampled design matrices, and then uses penalized-cost comparisons to show that a good trend estimate is selected. The closed-form projection identity used in Lemma 9 is valid for the least-squares estimator, and the fact that it does not hold for Tukey's estimator means the theorem is narrower than the abstract suggests, not that the theorem is equivalent to its assumptions. The self-citations to CAPA (Fisch et al., 2020, 2022) are not load-bearing in a circular sense: CAPA is a published algorithm that the paper extends, not an unverified uniqueness theorem invoked to forbid alternatives. The variance estimator is cited to Frick et al. (2014), Dette et al. (1998), and Pein et al. (2017), which are external published results. Penalty choices such as λ_coll = 4 are justified by simulation studies, which is empirical tuning rather than circular prediction. The application sets larger penalties based on those simulations; that is a judgment call, not a reduction of the claim to its inputs. Overall, no step in the paper's derivation reduces to its own output by construction, so the appropriate circularity score is 0.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

The method introduces no new physical or statistical entities beyond the standard anomaly, trend, seasonality, and noise components. The main ledger items are hand-chosen tuning parameters and the model assumptions, especially the theorem proof's simplification to least squares without point anomalies.

free parameters (5)
  • lambda_coll = 4
    Penalty for collective anomalies in cost (12); chosen by hand after the simulation study in Section 4.4, not derived from first principles.
  • lambda_point = 3
    Penalty for point anomalies in cost (13); chosen together with lambda_coll from simulations in Section 4.4.
  • J = 20
    Number of sampling repetitions for trend estimation; default selected from Table 3 because accuracy improves only marginally beyond 20.
  • Tukey tuning constant c = 4.685
    Threshold of the biweight loss in (18); set by default according to standard robust statistics practice, not derived from the model.
  • Differencing multiplier C = C such that D = C*P is about 0.1 n; C = 1 if P >= n/10
    Choice of lag for seasonality removal in Section 2.4; affects which anomalies remain detectable after differencing via Assumption 1.
assumptions (6)
  • domain assumption Time series decomposes as Y = A + T + S + eps with i.i.d. centered Gaussian noise eps (Section 2.1)
    All theory and simulations are built on this additive model; non-Gaussian or dependent noise is outside the claimed guarantees.
  • domain assumption Seasonality has known period P and zero-sum periodic values; trend is a polynomial whose true degree is in Q (Section 2.1 and Assumption 3)
    The lag-differencing step requires P, and the regression and cost selection require a valid polynomial degree in Q.
  • domain assumption Anomalies are rectangular intervals with constant magnitudes, no overlap, and satisfy the length and separation conditions in Assumption 1 (equations 30 to 32)
    Differencing creates two anomalies per true anomaly, and detectability relies on these bounds.
  • domain assumption Total length of collective anomalies is bounded by Assumption 2: sum of lengths < gamma n / max(4(Q+1), 8)
    Needed to guarantee that an anomaly-free subsample exists for trend estimation in Lemma 2.
  • ad hoc to paper For the theorem, point anomalies are absent and the trend estimator is least squares rather than the Tukey M-estimator
    Stated immediately before Theorem 1; the proof does not cover the actual algorithm as defined in Section 2.6 for point anomalies or the robust loss.
  • domain assumption Variance estimate sigma_0 hat is within constant factors of sigma_0 (Assumption 4 and Proposition 1)
    The cost penalties and all proofs use sigma_0 hat in place of sigma_0.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Detection of collective and point anomalies in the presence of trend and seasonality." pith.science (2026). https://pith.science/paper/3L7Z3HFH

@misc{pith2026250821128,
  author       = {Pith},
  title        = {Pith review of: Detection of collective and point anomalies in the presence of trend and seasonality},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3L7Z3HFH}},
  note         = {Machine review of arXiv:2508.21128}
}
read the original abstract

Detecting anomalies in time series data is a challenging task with broad relevance in many applications. Existing methods work effectively only under idealized conditions, typically focusing on point anomalies or assuming a constant baseline. Our approach overcomes these limitations by detecting both collective and point anomalies, while allowing for polynomial trends and seasonal patterns. We establish statistical theory demonstrating that our method accurately decomposes the time series into anomaly, trend, seasonality, and remainder components. We further show that the approach provides a consistent estimate of the number of anomalies and their locations. Simulation studies confirm its strong detection performance with finite samples, and an application to energy price data illustrates its practical utility.

Figures

Figures reproduced from arXiv: 2508.21128 by the authors.

Figure 1
Figure 1. A simulated example illustrating a time series that includes a quadratic trend [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. As an illustration of its anomaly detection capabilities, CAPA is applied to the example shown in [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Decomposition of the simulated example illustrated in [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: The decomposition and anomaly detection results using STAD. It contains five subplots: the original data from [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: The anomaly detection results using STAD on the example in [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: STAD’s average detection accuracy as a function of the collective anomaly penalty 𝜆coll. It shows that a wide range of penalty values offer a good performance. and seasonality estimation. However, if the anomaly is of high magnitude then the penalty can be chosen much …
Figure 7
Figure 7. Figure 7: Comparison of CAPA and STAD applied to the Octopus Agile East Midlands export price data (April 17, 2025 – May 26, 2025). Despite the extremely large penalty CAPA detects a large number of anomalies, mostly because of the seasonal pattern. It is unlikely that these fin…
Figure 8
Figure 8. Figure 8: Decomposition of the Octopus Agile electricity price data by STAD as illustrated in [PITH_FULL_IMAGE:figures/full_fig_p021_8.png]
Figure 9
Figure 9. Figure 9: A simple quadratic trend and a long anomaly. A too small [PITH_FULL_IMAGE:figures/full_fig_p025_9.png]
Figure 10
Figure 10. Figure 10: A more complicated cubic trend and a short anomaly. A too large [PITH_FULL_IMAGE:figures/full_fig_p026_10.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

37 extracted references · 37 canonical work pages

  1. [1]

    and Keynia, F

    Amjady, N. and Keynia, F. (2010) Short-term load forecasting of power systems by combination of wavelet transform and neuro-evolutionary algorithm. Energy, 35, 1768--1774

    work page 2010

  2. [2]

    Beaton, A. E. and Tukey, J. W. (1974) The fitting of power series, meaning polynomials, illustrated on band-spectroscopic data. Technometrics, 16, 147--185

    work page 1974

  3. [3]

    and Chawla, S

    Chalapathy, R. and Chawla, S. (2019) Deep learning for anomaly detection: A review. ACM Computing Surveys (CSUR), 52, 1--38

    work page 2019

  4. [4]

    and Kumar, V

    Chandola, V., Banerjee, A. and Kumar, V. (2009) Anomaly detection: A survey. ACM Computing Surveys (CSUR), 41, 1--58

    work page 2009

  5. [5]

    B., Cleveland, W

    Cleveland, R. B., Cleveland, W. S., McRae, J. E. and Terpenning, I. (1990) Stl: A seasonal-trend decomposition procedure based on loess. Journal of Official Statistics, 6, 3--73

    work page 1990

  6. [6]

    D., Azuaje, F

    Clifford, G. D., Azuaje, F. and McSharry, P. E. (2006) Advanced methods and tools for ecg data analysis. Physiological Measurement, 27, R1--R25

    work page 2006

  7. [7]

    (1978) A practical guide to splines

    De Boor, C. (1978) A practical guide to splines. Springer-Verlag google schola, 2, 4135--4195

    work page 1978

  8. [8]

    and Wagner, T

    Dette, H., Munk, A. and Wagner, T. (1998) Estimating the variance in nonparametric regression—what is a reasonable choice? Journal of the Royal Statistical Society: Series B (Statistical Methodology), 60, 751--764

    work page 1998

Show all 37 references
  1. [9]

    and Morse, P

    Feller, W. and Morse, P. M. (1958) An introduction to probability theory and its applications

    1958

  2. [10]

    and Eckley, I

    Fisch, A., Bardwell, L. and Eckley, I. A. (2020) Real time anomaly detection and categorisation. arXiv preprint arXiv:2009.06670

    2020 arXiv

  3. [11]

    T., Eckley, I

    Fisch, A. T., Eckley, I. A. and Fearnhead, P. (2022) A linear time method for the detection of collective and point anomalies. Statistical Analysis and Data Mining: The ASA Data Science Journal, 15, 494--508

    2022

  4. [12]

    and Sieling, H

    Frick, K., Munk, A. and Sieling, H. (2014) Multiscale change point inference. Journal of the Royal Statistical Society Series B: Statistical Methodology, 76, 495--580

    2014

  5. [13]

    International Journal of Contemporary Mathematical Sciences, 6, 409--421

    Özlem Gürünlü Alma (2011) Comparison of robust regression methods in linear regression. International Journal of Contemporary Mathematical Sciences, 6, 409--421

    2011

  6. [14]

    D., Rigaill, G., Fearnhead, P

    Hocking, T. D., Rigaill, G., Fearnhead, P. and Bourque, G. (2022) Generalized functional pruning optimal partitioning (gfpop) for constrained changepoint detection in genomic data. Journal of Statistical Software, 101, 1--31

    2022

  7. [15]

    and Fan, S

    Hong, T. and Fan, S. (2016) Energy forecasting: Past, present, and future. Foresight: The International Journal of Applied Forecasting, 43, 43--48

    2016

  8. [16]

    (1981) Robust statistics

    Huber, P. (1981) Robust statistics. new york: John wiley and sons. HuberRobust statistics1981

    1981

  9. [17]

    Huber, P. J. (1964) Robust estimation of a location parameter. Annals of Mathematical Statistics, 35, 73--101

    1964

  10. [18]

    Kendrick, D. A. (2002) Stochastic control for economic models. Portal Ebook UNTAG SURABAYA, second edn

    2002

  11. [19]

    and Thenmozhi, M

    Kumar, M. and Thenmozhi, M. (2006) Financial time series forecasting with a hybrid ai model. Expert Systems with Applications, 33, 917--927

    2006

  12. [20]

    and Massart, P

    Laurent, B. and Massart, P. (2000) Adaptive estimation of a quadratic functional by model selection. Annals of statistics, 1302--1338

    2000

  13. [21]

    T., Ting, K

    Liu, F. T., Ting, K. M. and Zhou, Z.-H. (2008) Isolation forest. In 2008 eighth ieee international conference on data mining, 413--422. IEEE

    2008

  14. [22]

    Liu, S., Liu, S., Wei, S. et al. (2015) Early diagnosis of alzheimer's disease with deep learning. Proceedings of the IEEE International Conference on Bioinformatics and Biomedicine (BIBM), 534--539

    2015

  15. [23]

    A., Martin, R

    Maronna, R. A., Martin, R. D., Yohai, V. J. and Salibi \'a n-Barrera, M. (2019) Robust Statistics: Theory and Methods. John Wiley & Sons, 2nd edn

    2019

  16. [24]

    and Provost, S

    Mathai, A. and Provost, S. (1992) Quadratic Forms in Random Variables: Theory and Applications, vol. 126. Marcel Dekker inc

    1992

  17. [25]

    and Yohai, V

    Muler, N. and Yohai, V. J. (2002) Robust estimation for arch models. Annals of Statistics, 30, 449--499

    2002

  18. [26]

    Journal of Financial Econometrics, 6, 382--411

    --- (2008) Robust estimation for garch models. Journal of Financial Econometrics, 6, 382--411

    2008

  19. [27]

    Annals of Statistics, 37, 1549--1572

    --- (2009) Robust estimation for arma models. Annals of Statistics, 37, 1549--1572

    2009

  20. [28]

    and Munk, A

    Pein, F., Sieling, H. and Munk, A. (2017) Heterogeneous change point inference. Journal of the Royal Statistical Society. Series B (Statistical Methodology), 79, 1207--1227

    2017

  21. [29]

    Rousseeuw, P. J. and Yohai, V. J. (1984) Robust regression by means of s-estimators. Robust and Nonlinear Time Series Analysis, 26, 256--272

    1984

  22. [30]

    and Yairi, T

    Sakurada, M. and Yairi, T. (2014) Anomaly detection using autoencoders with nonlinear dimensionality reduction. In Proceedings of the MLSDA 2014 2nd Workshop on Machine Learning for Sensory Data Analysis, 4--11

    2014

  23. [31]

    and Clairand, J.-M

    Serrano-Guerrero, X., Escriv \'a -Escriv \'a , G., Luna-Romero, S. and Clairand, J.-M. (2020) A time-series treatment method to obtain electrical consumption patterns for anomalies detection improvement in electrical consumption profiles. Energies, 13, 1046

    2020

  24. [32]

    Taylor, S. J. and Letham, B. (2018) Forecasting at scale. The American Statistician, 72, 37--45

    2018

  25. [33]

    and Zhu, S

    Wen, Q., Gao, J., Song, X., Sun, L., Xu, H. and Zhu, S. (2019) Robuststl: A robust seasonal-trend decomposition algorithm for long time series. In Proceedings of the AAAI conference on artificial intelligence, vol. 33, 5409--5416

    2019

  26. [34]

    and Tang, H

    Wu, Y., Dai, H.-N. and Tang, H. (2021) Graph neural networks for anomaly detection in industrial internet of things. IEEE Internet of Things Journal, 9, 9214--9231

    2021

  27. [35]

    I., Pan, S., Aggarwal, C

    Zamanzadeh Darban, Z., Webb, G. I., Pan, S., Aggarwal, C. and Salehi, M. (2024) Deep learning for time series anomaly detection: A survey. ACM Computing Surveys, 57, 1--42

    2024

  28. [36]

    E., Wu, D

    Zhang, J. E., Wu, D. and Boulet, B. (2021) Time series anomaly detection for smart grids: A survey. In 2021 IEEE electrical power and energy conference (EPEC), 125--130. IEEE

    2021

  29. [37]

    Zhao, Y., Nasrullah, Z., Hryniewicki, M. K. and Li, Z. (2023) Anomaly detection in time series: A comprehensive review and benchmark. arXiv preprint arXiv:2307.03253

    2023 arXiv

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.