{"id":"e95dcde1-29fd-4f7d-80b7-b753631b2134","arxiv_id":"2411.15254","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"Multi-pofo is a simple shared-encoder MLP for multi-scale load forecasting, but its claimed superiority over baselines is not quantitatively demonstrated.","lead":"A proposed load-forecasting model, Multi-pofo, uses a shared MLP encoder with padding and scale tags to predict building electricity use over daily, weekly, and monthly horizons. The paper claims it beats LSTM baselines, but it shows no quantitative baseline comparison and describes a positional encoding layer that is absent from the methods.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No quantitative baseline comparison exists in the paper; the central claim of outperforming strong baselines is therefore unsupported, and the described architecture omits the 'temporal positional encoding layer' promised in the abstract.","rationale":"The paper's stated contribution is empirical (Sections I and IV): a single MLP replaces separate daily/weekly/monthly forecasters while outperforming strong baselines. For that claim to be true, the paper must (1) specify the model precisely enough to reproduce, and (2) show it beats the baselines on the same data with comparable effort. Both conditions fail. Table I has no comparison; Figure 5 is qualitative; Section III never defines the 'temporal positional encoding layer' named in the abstract; and Eq. (3) depends on L_i, which is never reported. I focused on the missing quantitative baseline because it is the minimal condition for the headline claim: if the baseline numbers were added and Multi-pofo lost, the manuscript's central assertion would fail regardless of whether the transfer assumption in Section III-C/D holds. The transfer assumption is an additional concern, but it is secondary to the absence of any measurable comparison. I do not see an internally consistent route from the reported evidence to the SOTA claim, so the reader's REJECT verdict is unchanged. A revised paper that adds per-scale baseline MAEs, error bars, hyperparameters, and either defines or removes the 'temporal positional encoding layer' could be worthy of re-review.","tokens_in":7050,"tokens_out":5485,"duration_ms":59908,"concrete_test":"Obtain the authors' code and data (or re-implement Section III exactly as specified) and run the two baselines, BiLSTM [5] and CNN-LSTM [6], on the same Suita Campus 2015-2020 split with the same normalization and evaluation protocol. Compute per-scale MAE for each baseline and add those columns to Table I, including error bars over circuits and a paired significance test. If Multi-pofo does not beat both baselines at daily, weekly, and monthly scales, the central claim in the abstract and Section I is not supported by the evidence.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim, stated in the abstract and Section I, is that Multi-pofo outperforms 'several strong baseline methods' and achieves state-of-the-art forecasting performance. The only evidence offered is Table I, which lists MAE values for Multi-pofo alone at three scales for three building groups; there are no baseline columns, no error bars, and no per-circuit spread. Section IV's comparison to BiLSTM and CNN-LSTM is purely qualitative (Figure 5), with no numerical scores, no error bars, and no statistical test. Even the hyperparameters needed to reproduce the method are missing: L_i, L_max, I, D, H, and the number of parameters are never given, and Eq. (3)'s target horizon depends on the unreported L_i. Additionally, the abstract and conclusion credit a 'temporal positional encoding layer' for indexing scales, but Section III defines only a one-hot scale embedding concatenated to the zero-padded input; no positional encoding operation is specified. Thus the method as written cannot be reconstructed, and the comparison that would substantiate the central claim is absent. This is an evidentiary and internal-consistency failure, not a matter of baseline choice or consensus.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Multi-pofo, a multi-scale power load forecasting framework that uses zero padding and a one-hot scale embedding to unify daily, weekly, monthly, and yearly inputs, a shared MLP encoder trained via reconstruction, and a frozen-encoder linear prediction head. The authors claim that Multi-pofo outperforms strong baselines such as BiLSTM and CNN-LSTM on a real-world campus dataset, and that it captures mid- and long-term dependencies in a single model. The manuscript reports only Multi-pofo's MAE values in Table I and a qualitative comparison in Figure 5, with no quantitative baseline results, no error bars, and no statistical tests. The method section omits key hyperparameters, and the abstract's promise of a 'temporal positional encoding layer' is not realized in the described architecture, which uses only a one-hot scale embedding.","tokens_in":7308,"tokens_out":3962,"duration_ms":49449,"significance":"If the claimed results were properly supported, the idea of a single simple MLP-based model that handles multiple forecast horizons via shared representation learning could be practically useful for building-level and campus-level energy management, where separate daily, weekly, and monthly models are currently common. The paper also identifies a real gap in existing deep-learning time-series methods, which generally assume a fixed input scale. However, the manuscript as written does not provide the evidence needed to assess either the performance claim or the method's novelty, since the baseline comparison is qualitative and the architecture description is incomplete. The absence of any numeric baseline comparison and the internal inconsistency about the positional encoding layer are load-bearing issues that undermine the central claims.","major_comments":[{"comment":"The central claim that Multi-pofo 'outperforms' several strong baselines is not supported by any quantitative comparison. Table I lists only Multi-pofo's MAE values; there are no baseline columns, no standard deviations, and no statistical tests. Figure 5 is purely qualitative and does not report numeric scores. The abstract and Section I make strong superiority claims, but the evidence required to substantiate them is entirely absent.","section":"Section IV, Table I"},{"comment":"The method is not reproducible because the input window lengths L_i for day, week, month, and year, the maximum length L_max, the scale embedding dimension I, the encoder hidden dimension D, the number of layers, and all training hyperparameters (learning rate, optimizer, batch size, epoch counts) are never reported. Eq. (3) defines the forecast target in terms of L_i, but L_i is unspecified, so the actual target used in the experiments cannot be determined.","section":"Section III"},{"comment":"The abstract and conclusion credit a 'temporal positional encoding layer' as a novel component of the proposed architecture, but Section III-B defines only zero padding and a one-hot scale embedding concatenated to the input. No positional encoding operation is described anywhere in the method. This internal inconsistency means the architecture actually evaluated may differ from the one claimed, and the novelty statement is not verifiable.","section":"Abstract and Section V vs. Section III-B"},{"comment":"The two-stage training procedure, in which the shared encoder is trained only for reconstruction and then frozen, is a central design choice, but its validity is not justified. The paper provides no ablation against end-to-end training, no analysis of reconstruction quality, no comparison of frozen versus fine-tuned encoders, and no evidence that the reconstruction-trained representation z is informative for predicting future period maxima. Without such validation, the claim that the shared encoder 'captures common features' across scales is an unsupported assumption.","section":"Section III-C and III-D"},{"comment":"The experimental section does not provide any quantitative comparison with the baselines, so the conclusions drawn from Figure 5—such as 'LSTM struggles to accurately capture mid- to long-term temporal dependencies'—are not backed by numerical evidence. Additionally, the MAE values in Table I are reported without context, such as relative errors or per-circuit statistics, making it impossible to judge whether the forecasts are accurate in any absolute sense.","section":"Section IV"}],"minor_comments":[{"comment":"The notation [X'_i; s_i] is used for concatenation, but the dimensions are not explicitly discussed; it would be clearer to state that [·;·] denotes concatenation along the feature dimension.","section":"Section III-B, Eq. (1)"},{"comment":"The reconstruction loss is written as a sum over I scales, but the index I is not defined; it should be the number of scales (four).","section":"Section III-C, Eq. for L_recon"},{"comment":"The target definition Yi = max(x_{t+1}, ..., x_{t+L_i-1}) appears to use L_i-1 future steps; if the intent is to predict the next period's maximum over L_i steps, the upper index should be x_{t+L_i}. This ambiguity should be resolved.","section":"Section IV, Eq. (3)"},{"comment":"The figure appears garbled with numeric sequences in the axis labels and legends (e.g., '/s48 /s53...'); this likely stems from a font or rendering issue, but the figure is not interpretable in its current form.","section":"Figure 5"},{"comment":"The paper contains several grammatical errors and awkward phrasings (e.g., 'outperforms compared to' in the abstract) that would need correction in a revised version.","section":"Throughout"}],"recommendation":"reject","confidential_remarks":"The manuscript is substantially incomplete as a research contribution: the quantitative baseline comparison promised in the abstract is missing, and the described architecture does not match the claimed 'temporal positional encoding layer.' The missing hyperparameters and experimental details further prevent reproducibility. While the core idea of a shared multi-scale encoder is not without merit, the paper in its current form does not meet the evidentiary standards for publication. I would only reconsider if the authors provide complete experimental comparisons, full architectural details, and a corrected description of the proposed method."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea is a simple shared MLP that takes zero-padded sequences plus a one-hot scale tag, learns a unified representation through autoencoder pretraining, then a frozen encoder feeds a linear head for forecasting. That combination is a legitimate and modest extension of standard conditioning tricks. The two-stage training and the use of a single model across daily/weekly/monthly horizons is practically appealing for campus energy management, and the authors use real data from 313 circuits over five years. So the paper is not without merit.\n\nBut the central claim is not supported. The abstract says the approach outperforms strong baselines, yet Table I reports only Multi-pofo's MAE with no baseline columns and no error bars. The only comparison is a qualitative plot (Figure 5) where BiLSTM and CNN-LSTM curves are eyeballed. No statistical test, no per-circuit numbers. That is a load-bearing flaw, not a cosmetic one. Additionally, the authors never report the window lengths L_i, the scale embedding dimension I, the hidden dimension D, the head size H, or training hyperparameters beyond epoch counts. Eq. (3) defines the target using L_i, so the reader cannot even reconstruct what is being predicted.\n\nThere is also an internal inconsistency: the abstract and conclusion credit a 'temporal positional encoding layer' for indexing scales, but Section III-B defines only a one-hot scale embedding. No positional encoding operation is described anywhere. This is not a circularity issue, but it is a clarity problem that needs fixing.\n\nThe underlying MLP idea may be worth revisiting, but this manuscript is not ready for publication. The missing baseline comparison alone makes the headline claim unverifiable, and the missing hyperparameters make the method unreproducible. If the authors supply quantitative baselines, error bars, the omitted architecture details, and code or data, the paper could become a legitimate incremental contribution. As it stands, I would not accept it in a peer-reviewed venue, though I would not desk-reject it out of hand—there is enough substance in the method and data that a serious referee could push it toward a usable form.","headline":"Useful multi-scale MLP idea, but the evidence as written is too thin: no baseline numbers, missing hyperparameters, and an abstract component that never appears in the methods.","tokens_in":7806,"tokens_out":2063,"would_cite":false,"duration_ms":22282,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that one shared MLP encoder, trained to reconstruct zero-padded and scale-tagged load sequences, then frozen, can forecast next-period peak load at daily, weekly, monthly, and yearly scales and outperform LSTM baselines.","keywords":["multi-timescale forecasting","peak load forecasting","power load forecasting","MLP encoder","representation learning","zero padding","scale embedding","smart grids"],"falsifier":"A concrete check: on the same campus dataset, fine-tune the encoder jointly with the prediction head and compare accuracy with the frozen-encoder version; if fine-tuning does not match or improve the reported results, then reconstruction-only pretraining is not what carries the forecasts. Reporting the exact window lengths $L_i$ and embedding dimension $D$ would also let an independent run reproduce the daily, weekly, and monthly MAE figures.","tokens_in":6841,"feed_emoji":"⚡","tokens_out":8953,"duration_ms":79285,"temperature":0.7,"pith_summary":"Multi-pofo is a unified load-forecasting framework aimed at predicting the next-period peak power from input sequences of different lengths — daily, weekly, monthly, and yearly — with one shared model. The paper argues that a shared MLP encoder, trained first to reconstruct zero-padded, scale-tagged inputs and then frozen, learns a common representation from which a simple linear layer can forecast the maximum load at every scale. On a six-year campus electricity dataset, the authors report low mean absolute error at each scale and present visual comparisons against LSTM baselines showing that the baselines lose mid- and long-term structure at the monthly horizon. If correct, the framework would let system operators replace several specialized forecasters with one lightweight model and simplify rolling optimization across scheduling horizons.","feed_headline":"One shared MLP forecasts daily, weekly, monthly power peaks","feed_subtitle":"Zero-padding and scale tags let a single frozen encoder replace separate load forecasters at every horizon.","key_machinery":"The load-bearing object is the frozen shared representation $z\\in\\mathbb{R}^D$ produced by a shared MLP encoder from zero-padded, scale-tagged inputs. Two-stage training separates representation learning from prediction: the encoder-decoder minimizes the reconstruction loss $\\mathcal{L}_{\\text{recon}}=\\sum_i\\|X'_i-\\tilde{X}_i\\|_2^2$, and after freezing, the prediction layer minimizes $\\mathcal{L}_{\\text{pred}}=\\|Y-\\tilde{Y}\\|_2^2$. The scale tag $s_i$ tells the encoder which temporal resolution a padded sequence came from, and sharing the encoder across scales is what lets one representation carry both short- and long-horizon structure.","core_discovery":"The central claim is that mid- and long-term dependencies in power load can be captured by a shared three-layer MLP encoder rather than by recurrent or convolutional architectures. For each scale $i$, the input sequence $X_i$ is zero-padded to the maximum length and concatenated with a one-hot scale tag $s_i$, forming $\\hat{X}_i$; an encoder-decoder pair is trained to reconstruct the padded input, then the encoder is frozen and a single fully connected layer $W_{\\text{pred}}z+b_{\\text{pred}}$ is trained to output the maximum load of the next period. The authors claim this produces accurate daily, weekly, and monthly forecasts on real campus circuits and that it captures both mid- and long-term dependencies in a unified model, with the monthly scale being where LSTM baselines visibly fail.","pith_inferences":["Implicit extension: if the representation $z$ is genuinely scale-invariant, the same frozen encoder should support an unseen scale, such as quarterly peaks, by training only a new linear head; this is directly testable on the same dataset.","Testable extension: replacing the one-hot scale tag with a learned continuous embedding could let the model interpolate between aggregation levels, but the paper does not explore this.","The paper does not ablate the two-stage procedure, so an independent check of whether freezing rather than fine-tuning the encoder preserves multi-scale structure would determine how much of the result rests on the transfer assumption."],"forward_implications":["A single trained model can forecast daily, weekly, monthly, and yearly peak loads from inputs of different lengths, with no per-scale architectural changes.","After the encoder is frozen, additional forecast horizons can be added by training only a new linear head, keeping deployment cost low.","Weekly forecasts are expected to be the most accurate because weekly aggregation smooths daily noise, a pattern the paper observes and attributes to the shared representation.","The comparison against BiLSTM and CNN-LSTM indicates that recurrent models lose accuracy as the horizon grows, while the MLP representation preserves monthly structure."],"supporting_citations":[{"why":"Supplies the BiLSTM-attention baseline that Multi-pofo is visually compared against in single-building forecasting.","marker":"[5]"},{"why":"Supplies the CNN-LSTM baseline used to show that recurrent models lose mid- and long-term dependencies.","marker":"[6]"},{"why":"Gives the superposition-of-scales view of building load that motivates the multi-scale input design.","marker":"[3]"},{"why":"Provides the multi-task load forecasting setting that the shared representation is meant to handle.","marker":"[4]"},{"why":"Documents the limitation of deep forecasting models on uniform-scale inputs, motivating zero-padding and scale tags.","marker":"[9]"},{"why":"Supports the need for special handling of non-stationary multi-scale series, motivating a shared representation.","marker":"[10]"}],"fun_headline_variants":["Shared MLP encoder forecasts power loads across all horizons","Scale tags and zero-padding unite daily, weekly, monthly forecasts","One MLP replaces LSTM for mid and long-term power load prediction","Frozen shared encoder captures all timescales in smart grids","Multi-timescale load forecasting via a single MLP with scale tags"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a shared encoder trained only to reconstruct zero-padded, scale-tagged load sequences learns a representation informative enough for predicting future peak loads at every scale, even though the forecasting layer is trained on a frozen encoder and never updates it.","fun_headline_variants_meta":{"raw":{"variants":["Shared MLP encoder forecasts power loads across all horizons","Scale tags and zero-padding unite daily, weekly, monthly forecasts","One MLP replaces LSTM for mid and long-term power load prediction","Frozen shared encoder captures all timescales in smart grids","Multi-timescale load forecasting via a single MLP with scale tags"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000251,"raw_usage":{"total_tokens":1503,"prompt_tokens":837,"completion_tokens":666,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":453,"completion_tokens_details":{"reasoning_tokens":578}},"tokens_in":453,"tokens_out":666,"duration_ms":6459,"temperature":1.0,"reasoning_tokens":578,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T14:52:40.181473+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete check: on the same campus dataset, fine-tune the encoder jointly with the prediction head and compare accuracy with the frozen-encoder version; if fine-tuning does not match or improve the reported results, then reconstruction-only pretraining is not what carries the forecasts. Reporting the exact window lengths $L_i$ and embedding dimension $D$ would also let an independent run reproduce the daily, weekly, and monthly MAE figures.","supporting_citations":[{"cited_title":"Power load forecasting using bilstm-attention,","cited_arxiv_id":null,"evidence_quote":"Supplies the BiLSTM-attention baseline that Multi-pofo is visually compared against in single-building forecasting."},{"cited_title":"Multi-Scale Convolutional Neural Network With Time-Cognition for Multi-Step Short-Term Load Forecasting,","cited_arxiv_id":null,"evidence_quote":"Supplies the CNN-LSTM baseline used to show that recurrent models lose mid- and long-term dependencies."},{"cited_title":"Multi-Scale Dilated Convolution Network for Long-Term Time Series Forecasting","cited_arxiv_id":"2405.05499","evidence_quote":"Gives the superposition-of-scales view of building load that motivates the multi-scale input design."},{"cited_title":"Multi-node load forecasting based on multi-task learning with modal feature extraction,","cited_arxiv_id":null,"evidence_quote":"Provides the multi-task load forecasting setting that the shared representation is meant to handle."},{"cited_title":"Fred- former: Frequency Debiased Transformer for Time Series Forecasting,","cited_arxiv_id":null,"evidence_quote":"Documents the limitation of deep forecasting models on uniform-scale inputs, motivating zero-padding and scale tags."},{"cited_title":"Frednormer: Frequency domain normalization for non-stationary time series forecast- ing,","cited_arxiv_id":null,"evidence_quote":"Supports the need for special handling of non-stationary multi-scale series, motivating a shared representation."}],"review_version":1}