An Investigation into Seasonal Variations in Energy Forecasting for Student Residences

Mathumitha Sureshkumar; Muhammad Umair Danish; Tehara Fonseka; Umeshika Uthayakumar; Vinura Galwaduge

REVIEW 4 major objections 5 minor 17 references

An Investigation into Seasonal Variations in Energy Forecasting for Student Residences

T0 review · 4 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read A 13-model study of student residences finds that the best energy forecasting model changes with the season, and two adaptive designs lead overall.

desk verdict A useful two-building seasonal benchmark, but the central claim hinges on a split rule the paper never states and the rankings may be noise without error bars. read the letter →

arxiv 2501.07423 v2 pith:GCNYV4KD submitted 2025-01-13 cs.LG cs.AI

Muhammad Umair Danish , Mathumitha Sureshkumar , Tehara Fonseka , Umeshika Uthayakumar , Vinura Galwaduge This is my paper · ORCID

classification cs.LGcs.AI

keywords energyforecastingseasonalvariationsLSTMhypernetworkXGBoostMiniRocketstudentresidencesmodelselection

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

This paper tests 13 machine learning models for day-ahead energy forecasting on two student residence buildings, using hourly data from 2019 to 2023. Its central finding is that no single model is consistently the most accurate across fall, winter, spring, and summer; the ranking of models changes with the season. The paper proposes two models—a Hypernetwork-based LSTM and a MiniRocket-plus-autoencoder-plus-XGBoost ensemble—and reports that both adapt particularly well to the abrupt consumption changes of summer. The takeaway is that energy forecasters should evaluate models per season or design adaptive architectures rather than commit to one universal model.

What carries the argument

The central objects are the two proposed architectures. HyperNetLSTM uses a hypernetwork—a small fully-connected network that takes the input window and dynamically produces the LSTM's weights, letting the recurrent model change its behavior as the input regime changes. MiniAutoEncXGBoost concatenates features extracted by MiniRocket (a fast deterministic convolution-based transform) and by a convolutional autoencoder, then feeds those features to an XGBoost regressor. The seasonal evaluation procedure—splitting the test predictions by calendar season and re-ranking the models—is what carries the main argument that model performance is not season-uniform.

What would settle it

Re-run the same 13-model comparison with a strictly chronological train/validation/test split—for example, training on data up to mid-2022 and testing on the final year—and compare the per-season SMAPE rankings; if a single model then wins every season or the proposed models' edge over LSTM disappears, the central claim fails.

Watch

Extended reading notes

Core claim

The paper's core claim is that seasonality changes which model should be trusted. On Residence 1, HyperNetLSTM has the lowest overall SMAPE (8.87), but in fall the Transformer wins (8.59), in winter HyperNetLSTM wins (6.11), and in summer HyperNetLSTM wins again (8.13). On Residence 2, MiniAutoEncXGBoost leads overall (7.37), yet LSTM wins in fall, ARFFNN in winter, AttentionLSTM in spring, and MiniAutoEncXGBoost in summer. The 'best model' is therefore season-dependent, and the proposed models' strength lies in tracking the irregular summer consumption drops tied to student vacations. The paper concludes that season-aware model selection or purpose-built adaptive designs are needed.

Load-bearing premise

The claim depends on the 70/10/20 split into training, validation, and test sets being chronological rather than random, because the paper does not state that the test period follows the training period in time.

Editorial extensions

If this is right

Practitioners should report seasonal error breakdowns, not only overall errors, before selecting a forecasting model.
A season-aware ensemble that switches among the top models per season would likely beat any single fixed model on these datasets.
The HyperNetLSTM design suggests a general mechanism for adapting recurrent networks to regime shifts beyond energy.
The MiniRocket-plus-autoencoder feature combination could transfer to other load-forecasting datasets with abrupt vacation or holiday effects.
The paper's conclusion points toward human-AI hybrid forecasters that encode occupancy or event information, since human behavior is the main driver of seasonal instability.

Reading between the lines

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

Because the data come from only two university residences, the season-specific rankings may not generalize to other building types; a testable extension is to run the same 13-model protocol on public load datasets with known calendar events.
The seasonal winners appear to line up with occupancy changes (e.g., summer vacations), so adding an academic-calendar feature directly might reduce the need for different architectures per season.
The hypernetwork's dynamic weight generation can be viewed as a soft form of season-conditioning; comparing it to feeding a season indicator into the LSTM would isolate what the hypernetwork actually adds.
The paper's central claim currently rests on point estimates; statistical significance tests across multiple train/test splits would show whether the seasonal ranking differences are real or within noise.

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

Worth a look if you do building load forecasting: two new residence datasets, 13 models, seasonal breakdowns, and two new hybrids. The headline finding—no single model wins all seasons—is plausible but rests on a split rule the paper never states.

What's actually new: the datasets are genuinely new, and the two hybrids (hypernetwork-LSTM, MiniRocket-autoencoder-XGBoost) are new combinations of known pieces. The seasonal tables are helpful; they show, for example, that the Transformer wins in fall for Residence 1 but is mediocre elsewhere. The two-building comparison is a plus—many such studies use one building.

The main soft spot is real: Section IV-A4 says the data were split 70/10/20 but does not say chronological. With a stride-1 sliding window, a random split puts overlapping 24-hour windows in both train and test, turning the reported errors into interpolation scores. That would deflate the reported SMAPE/MAE/RMSE and could reshuffle the seasonal rankings. The authors must clarify or re-run with a temporal split. This is not a manufactured issue; it is the first thing any referee should check.

Minor but worth noting: no error bars or significance tests, so the differences between models in Tables III–X could be within noise. Residence 1 is missing its spring table—the text says "Tables III, IV, and V" but only fall, winter, and summer appear. No code or data are released, which makes it hard to verify the split or reproduce the numbers.

The qualitative conclusion—season-specific model selection matters—is not new; it is common in forecasting. But the evidence here is more detailed than most, and the two proposed models do appear to be competitive in several seasons. If the temporal split is confirmed, this is a solid applied benchmark for two buildings.

Send to peer review, but require the authors to state the split rule, add uncertainty estimates, and release artifacts. The paper is honest applied work, not a breakthrough, and it deserves a serious referee to insist on those fixes.

Referee Report

4 major / 5 minor

Summary. The paper compares 13 machine-learning models, including two proposed architectures (HyperNetLSTM and MiniAutoEncXGBoost), for day-ahead (24-hour) energy-consumption forecasting on two student-residence datasets at Western University spanning 2019–2023. Using SMAPE, MAE, and RMSE, it reports overall and season-wise rankings; the central claim is that no single model dominates all seasons, so season-specific model selection is needed. The two proposed models are reported to perform best overall on Residence 1 and Residence 2, respectively.

Significance. If the central empirical claim survives closer scrutiny, this is a useful applied benchmark: the comparison covers a broad model zoo, uses real building-level data, and documents seasonal heterogeneity in rankings. The paper's strength is its transparent tabular reporting of errors for every model in every season, which makes the 'no single best model' phenomenon easy to see. However, the claim currently rests on single-run point estimates from a split whose temporal ordering is never stated, and one seasonal block is missing; these issues prevent the results from being fully load-bearing as published.

major comments (4)

[IV-A4 and IV-A3b] The data-splitting description in Section IV-A4 states only that the dataset was partitioned 70/10/20 into training, validation, and test sets, without specifying whether the split is chronological. This matters because the sliding-window construction in Section IV-A3b uses a stride of one, so consecutive windows differ by a single hour; a random split would place overlapping or near-identical windows in both training and test sets, turning the evaluation into interpolation and inflating forecast skill. Please state explicitly that the split preserves temporal order (and, if it does not, re-run the evaluation with a chronological split and report the season/month boundaries of the test period).
[V-L, Tables III–V] The Residence 1 seasonal analysis in Section V-L reports Fall, Winter, and Summer (Tables III–V) but has no Spring table, even though the abstract and Section VI make claims about performance 'across all seasons' and the Residence 2 analysis in Section V-N includes a Spring table (Table IX). The absence of Residence 1 Spring results leaves the 'no single model across all seasons' claim incomplete for one of the two buildings; please add the missing table or explain why it is omitted.
[V, Tables II–X] All error metrics are single-run point estimates with no standard deviations, confidence intervals, or significance tests. Several adjacent rankings are separated by less than 0.2 SMAPE (e.g., Table II: HyperNetLSTM 8.87 vs LSTM 9.12; Table VI: MiniAutoEncXGBoost 7.37 vs AttentionLSTM 7.51), so the season-by-season 'winner' assignments and the claim of 'strong adaptability' may be within seed/initialization noise. Run multiple seeds and report mean ± std or pairwise tests, or soften the claims to rankings-of-point-estimates.
[V-H and V-K] The proposed HyperNetLSTM and MiniAutoEncXGBoost are tuned via grid search on the same data that later produces the headline test-set comparisons, with no nested validation or separate model-selection phase. The reported superiority of the proposed models is therefore subject to selection-on-the-test-set optimism; please clarify the selection protocol or evaluate the final selected configurations on a truly held-out period.

minor comments (5)

[II-A] Section II-A contains a duplicated phrase: 'Additional feature extraction techniques were integrated additional feature extraction techniques were integrated.'; please remove the repetition.
[III] Section III refers to 'Extended Short-Term Memory Networks (LSTMs)'; the standard expansion is 'Long Short-Term Memory.'
[References] Reference [1] still contains the placeholder 'accessed: [Insert Access Date Here]'; fill in the access date or remove the placeholder.
[Figure 8 caption] The caption of Figure 8 says 'Hypernetwork-based MLP and primary network-based LSTM,' but the text describes the primary network as an LSTM; align the caption with the architecture description.
[Availability] There is no data or code availability statement; providing one, even with anonymized or aggregated data, would allow readers to reproduce the tables.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's claims are empirical comparisons on held-out data, not derivations that reduce to their inputs.

full rationale

This paper is an empirical benchmark study: it defines models, trains them on a fixed dataset split (70/10/20), and reports SMAPE/MAE/RMSE on a held-out test portion. The central claim that no single model dominates across seasons is a comparative summary of the reported error tables, not a quantity derived from a fitted parameter or an imported uniqueness theorem. The proposed HyperNetLSTM and MiniAutoEncXGBoost models are described by explicit architectures (Eqs. 9-17) and are not defined in terms of the forecast errors they later claim; their success is an empirical result. References to N-BEATS, AR-Net, MiniRocket, and Transformers are external prior work used as baseline implementations, not self-citations invoked to justify the central claim. The methodological ambiguity flagged by a skeptical reader (Section IV-A4 states only that the dataset 'was partitioned into training, validation, and test sets' without specifying chronological order) is a real validity threat for forecasting conclusions, but it is a data-splitting and experimental-design concern, not a circularity: the predictions are still computed from model outputs rather than being equivalent to the inputs by construction. Likewise, choosing the best model from the test tables is a model-selection concern (winner's curse), not a reduction of the claim to its own assumptions. No circular step can be exhibited with a specific equation where an output equals an input by definition, so the appropriate finding is no significant circularity.

Assumptions & free parameters 9 free parameters · 4 assumptions · 0 invented entities

The central empirical claim rests on the data, the split, and the model configurations. The main assumptions are that the split preserves temporal order, that the data are complete, and that observed temperature is usable without forecast uncertainty. No invented physical entities are introduced; the two proposed model architectures are engineering combinations.

free parameters (9)

LSTM hidden units = 64
Selected by grid search; affects LSTM, HyperNetLSTM, and AttentionLSTM models.
GRU hidden units = 64
Selected by grid search; used in GRU baseline.
RNN hidden units = 64
Selected by grid search; used in RNN baseline.
MLP hidden units = 32
Selected by grid search; used in MLP baseline.
TCN units = 128
Selected by grid search; used in temporal convolution baseline.
N-BEATS configuration = 6 blocks, 4 hidden layers, 512 units
Fixed configuration for the N-BEATS baseline.
Transformer configuration = 2 heads, 64 FC units before and 24 after encoder, dropout 0.1, lr 0.001
Selected from the grid search space in Table I.
XGBoost hyperparameters = lr 0.05, colsample_bytree 0.5, reg_lambda 1.2, subsample 0.8, booster gbtree
Optimized for the MiniAutoEncXGBoost and MiniWXGBoost ensembles.
Hypernetwork layer sizes = 128 then 64
Chosen as the optimal configuration for the proposed HyperNetLSTM.

assumptions (4)

domain assumption The 70/10/20 data split is chronological, with no shuffling of sliding-window samples.
Section IV-A4 only says the dataset was partitioned into training, validation, and test sets. Time-series leakage would invalidate the forecast errors if the split were random.
domain assumption The datasets are complete and continuous, with missing values absent or handled by an undescribed procedure.
Section IV-A describes MinMax scaling and sliding windows but no missing-data or holiday handling.
domain assumption Observed temperature is available for the forecast day, with no distinction between forecast and actual temperature.
Section IV-A2 lists temperature as a selected feature; day-ahead forecasting normally requires forecasted temperature, which would add uncertainty.
standard math SMAPE, MAE, and RMSE as defined in Equations (1) to (3) are the accepted evaluation metrics for this comparison.
These are standard forecast error metrics; no formal proof is needed.

how reviews work

0 comments

Cite this review

Pith. "Pith review of An Investigation into Seasonal Variations in Energy Forecasting for Student Residences." pith.science (2026). https://pith.science/paper/GCNYV4KD

@misc{pith2026250107423,
  author       = {Pith},
  title        = {Pith review of: An Investigation into Seasonal Variations in Energy Forecasting for Student Residences},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GCNYV4KD}},
  note         = {Machine review of arXiv:2501.07423}
}

read the original abstract

This research provides an in-depth evaluation of various machine learning models for energy forecasting, focusing on the unique challenges of seasonal variations in student residential settings. The study assesses the performance of baseline models, such as LSTM and GRU, alongside state-of-the-art forecasting methods, including Autoregressive Feedforward Neural Networks, Transformers, and hybrid approaches. Special attention is given to predicting energy consumption amidst challenges like seasonal patterns, vacations, meteorological changes, and irregular human activities that cause sudden fluctuations in usage. The findings reveal that no single model consistently outperforms others across all seasons, emphasizing the need for season-specific model selection or tailored designs. Notably, the proposed Hyper Network based LSTM and MiniAutoEncXGBoost models exhibit strong adaptability to seasonal variations, effectively capturing abrupt changes in energy consumption during summer months. This study advances the energy forecasting field by emphasizing the critical role of seasonal dynamics and model-specific behavior in achieving accurate predictions.

Figures

Figures reproduced from arXiv: 2501.07423 by the authors.

**Figure 2.** Energy consumption over three years for Student Residence 2, [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 5.** Left: Basic AR-Net without hidden layers. Right: Advanced AR-Net [PITH_FULL_IMAGE:figures/full_fig_p004_5.png] view at source ↗

Figures from the paper (8 more)

**Figure 6.** Figure 6: The encoder-based Transformer architecture with fully connected [PITH_FULL_IMAGE:figures/full_fig_p004_6.png]

**Figure 4.** Figure 4: The Neural Basis Expansion Analysis model is an MLP-based [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]

**Figure 8.** Figure 8: The architecture of the Hypernetwork-based MLP and primary [PITH_FULL_IMAGE:figures/full_fig_p005_8.png]

**Figure 9.** Figure 9: The ensemble model with a convolutional autoencoder and [PITH_FULL_IMAGE:figures/full_fig_p005_9.png]

**Figure 10.** Figure 10: Comparison of RMSE, SMAPE, and MAE among various models [PITH_FULL_IMAGE:figures/full_fig_p007_10.png]

**Figure 11.** Figure 11: Comparison of actual energy consumption with predictions by the [PITH_FULL_IMAGE:figures/full_fig_p007_11.png]

**Figure 12.** Figure 12: Comparison of RMSE, SMAPE, and MAE among various models [PITH_FULL_IMAGE:figures/full_fig_p008_12.png]

**Figure 13.** Figure 13: Actual vs. predicted energy consumption for Residence 2, highlight [PITH_FULL_IMAGE:figures/full_fig_p009_13.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

17 extracted references · 13 canonical work pages

[1]

Year-end data: Electricity demand in ontario,

Independent Electricity System Operator (IESO), “Year-end data: Electricity demand in ontario,” https://www.ieso.ca/en/Corporate-IESO/ Media/Year-End-Data#:∼:text=Electricity%20demand%20in% 20Ontario%20grew,coming%20out%20of%20the%20pandemic, 2021, accessed: [Insert Access Date Here]

work page 2021
[2]

(2013) World energy scenarios: Composing energy futures to 2050 - executive summary

World Energy Council. (2013) World energy scenarios: Composing energy futures to 2050 - executive summary. [Online]. Available: https: //www.worldenergy.org/assets/downloads/World-Energy-Scenarios Composing-energy-futures-to-2050 Executive-summary.pdf

work page 2013
[3]

Vegh and I

G. Vegh and I. T. Thomson. (2021, October 21) How to improve ontario energy forecasting. Ontario 360. Accessed: Date. [Online]. Available: https://on360.ca/policy-papers/ how-to-improve-ontario-energy-forecasting/

work page 2021
[4]

The impact of covid-19 pandemic on electricity consumption and electricity demand forecasting accuracy: Empirical evidence from the state of qatar,

A. Abulibdeh, E. Zaidan, and R. Jabbar, “The impact of covid-19 pandemic on electricity consumption and electricity demand forecasting accuracy: Empirical evidence from the state of qatar,” Energy Strategy Reviews, vol. 44, p. 100980, 2022

work page 2022
[5]

N-beats: Neu- ral basis expansion analysis for interpretable time series forecasting,

B. N. Oreshkin, D. Carpov, N. Chapados, and Y . Bengio, “N-beats: Neu- ral basis expansion analysis for interpretable time series forecasting,” arXiv preprint arXiv:1905.10437 , 2019

arXiv 1905
[6]

Ar-net: A simple auto-regressive neural network for time-series,

O. Triebe, N. Laptev, and R. Rajagopal, “Ar-net: A simple auto-regressive neural network for time-series,” arXiv preprint arXiv:1911.12436, 2019

arXiv 1911
[7]

Temporal convolutional networks applied to energy-related time series forecasting,

P. Lara-Ben ´ıtez, M. Carranza-Garc ´ıa, J. M. Luna-Romera, and J. C. Riquelme, “Temporal convolutional networks applied to energy-related time series forecasting,” applied sciences, vol. 10, no. 7, p. 2322, 2020

work page 2020
[8]

Minirocket: A very fast (almost) deterministic transform for time series classification,

A. Dempster, D. F. Schmidt, and G. I. Webb, “Minirocket: A very fast (almost) deterministic transform for time series classification,” in Proceedings of the 27th ACM SIGKDD conference on knowledge discovery & data mining , 2021, pp. 248–257

work page 2021

Show all 17 references

[9]

Forecasting methods in energy planning models,

K. B. Debnath and M. Mourshed, “Forecasting methods in energy planning models,” Renewable and Sustainable Energy Reviews , vol. 88, pp. 297–325, 2018

2018
[10]

Survey on renewable energy forecasting using different techniques,

V . A. Natarajan and P. Karatampati, “Survey on renewable energy forecasting using different techniques,” in 2019 2nd International Con- ference on Power and Embedded Drive Control (ICPEDC) . IEEE, 2019, pp. 349–354

2019
[11]

Attention is all you need,

A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,” Advances in neural information processing systems , vol. 30, 2017

2017
[12]

Transformer-based model for electrical load forecasting,

A. L’Heureux, K. Grolinger, and M. A. Capretz, “Transformer-based model for electrical load forecasting,” Energies, vol. 15, no. 14, p. 4993, 2022

2022
[13]

A hybrid prediction model based on pattern sequence-based matching method and extreme gradient boosting for holiday load forecasting,

K. Zhu, J. Geng, and K. Wang, “A hybrid prediction model based on pattern sequence-based matching method and extreme gradient boosting for holiday load forecasting,” Electric Power Systems Research , vol. 190, p. 106841, 2021

2021
[14]

Tsfel: Time series feature extrac- tion library,

M. Barandas, D. Folgado, L. Fernandes, S. Santos, M. Abreu, P. Bota, H. Liu, T. Schultz, and H. Gamboa, “Tsfel: Time series feature extrac- tion library,” SoftwareX, vol. 11, p. 100456, 2020

2020
[15]

Time series feature extraction for data mining using dwt and dft,

F. M ¨orchen, “Time series feature extraction for data mining using dwt and dft,” 2003

2003
[16]

Residential load profile clustering via deep convolutional autoencoder,

S. Ryu, H. Choi, H. Lee, H. Kim, and V . W. Wong, “Residential load profile clustering via deep convolutional autoencoder,” in 2018 IEEE international conference on communications, control, and computing technologies for smart grids (SmartGridComm) . IEEE, 2018, pp. 1–6

2018
[17]

A novel deep learning architecture and minirocket feature extraction method for human activity recognition using ecg, ppg and inertial sensor dataset,

R. K. Bondugula, S. K. Udgata, and K. B. Sivangi, “A novel deep learning architecture and minirocket feature extraction method for human activity recognition using ecg, ppg and inertial sensor dataset,” Applied Intelligence , vol. 53, no. 11, pp. 14 400–14 425, 2023

2023

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.

[1] [1]

Year-end data: Electricity demand in ontario,

Independent Electricity System Operator (IESO), “Year-end data: Electricity demand in ontario,” https://www.ieso.ca/en/Corporate-IESO/ Media/Year-End-Data#:∼:text=Electricity%20demand%20in% 20Ontario%20grew,coming%20out%20of%20the%20pandemic, 2021, accessed: [Insert Access Date Here]

work page 2021

[2] [2]

(2013) World energy scenarios: Composing energy futures to 2050 - executive summary

World Energy Council. (2013) World energy scenarios: Composing energy futures to 2050 - executive summary. [Online]. Available: https: //www.worldenergy.org/assets/downloads/World-Energy-Scenarios Composing-energy-futures-to-2050 Executive-summary.pdf

work page 2013

[3] [3]

Vegh and I

G. Vegh and I. T. Thomson. (2021, October 21) How to improve ontario energy forecasting. Ontario 360. Accessed: Date. [Online]. Available: https://on360.ca/policy-papers/ how-to-improve-ontario-energy-forecasting/

work page 2021

[4] [4]

The impact of covid-19 pandemic on electricity consumption and electricity demand forecasting accuracy: Empirical evidence from the state of qatar,

A. Abulibdeh, E. Zaidan, and R. Jabbar, “The impact of covid-19 pandemic on electricity consumption and electricity demand forecasting accuracy: Empirical evidence from the state of qatar,” Energy Strategy Reviews, vol. 44, p. 100980, 2022

work page 2022

[5] [5]

N-beats: Neu- ral basis expansion analysis for interpretable time series forecasting,

B. N. Oreshkin, D. Carpov, N. Chapados, and Y . Bengio, “N-beats: Neu- ral basis expansion analysis for interpretable time series forecasting,” arXiv preprint arXiv:1905.10437 , 2019

arXiv 1905

[6] [6]

Ar-net: A simple auto-regressive neural network for time-series,

O. Triebe, N. Laptev, and R. Rajagopal, “Ar-net: A simple auto-regressive neural network for time-series,” arXiv preprint arXiv:1911.12436, 2019

arXiv 1911

[7] [7]

Temporal convolutional networks applied to energy-related time series forecasting,

P. Lara-Ben ´ıtez, M. Carranza-Garc ´ıa, J. M. Luna-Romera, and J. C. Riquelme, “Temporal convolutional networks applied to energy-related time series forecasting,” applied sciences, vol. 10, no. 7, p. 2322, 2020

work page 2020

[8] [8]

Minirocket: A very fast (almost) deterministic transform for time series classification,

A. Dempster, D. F. Schmidt, and G. I. Webb, “Minirocket: A very fast (almost) deterministic transform for time series classification,” in Proceedings of the 27th ACM SIGKDD conference on knowledge discovery & data mining , 2021, pp. 248–257

work page 2021