{"id":"de6c16f1-f5b1-4fca-bd2b-bb250a524ea2","arxiv_id":"1909.00907","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Federated and clustered deep learning models achieved lower RMSE and communication overhead than standard ML baselines on a public EV charging dataset.","lead":"This paper applies deep learning, federated learning, and location-based clustering to predict electric vehicle charging energy demand using public charging session data from Dundee, UK. It reports lower prediction error and communication cost than several standard machine learning baselines.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline 24.63% improvement is not reproducible from Table I under any standard comparison; no error bars or protocol are provided.","rationale":"The reader's verdict is CONDITIONAL, and I agree that conditionality is appropriate. However, I do not think the weakest assumption is the non-IID convergence of Eq. (12): in this paper's setup, the local losses are defined as sums over local data (Eq. (10)), and the synchronous average of local gradients over a disjoint partition of the dataset is exactly the centralized gradient scaled by 1/J, so non-IID data alone does not break the construction. The more load-bearing problem is that the paper's headline numbers do not follow from its own Table I. The abstract and Section IV-C claim 'up to 24.63%' and '24.28%' and '24.63%' improvements, but simple arithmetic on the reported RMSEs yields a maximum of about 25.0% only for one baseline/training-split pair, and no cell gives 24.63%. This is an internal inconsistency, not merely a disagreement with external consensus. It is also connected to a reproducibility gap: no error bars, no seeds, no cross-validation, and no communication-overhead definition. These are addressable concerns, so the appropriate outcome is to maintain the CONDITIONAL verdict pending the authors' clarification and a reproducibility check. The reader's rationale already mentions the percentage mismatch but does not center it, so agreement is partial rather than full.","tokens_in":9005,"tokens_out":5475,"duration_ms":57205,"concrete_test":"Recompute the percentage improvement from Table I for every baseline at every training split using (baseline_RMSE − proposed_RMSE)/baseline_RMSE, and compare against 24.63%. If no value equals 24.63%, the abstract overstates the table. Then obtain the authors' dataset (Dundee EV charging data) and code, rerun EDL, FEDL, and FEDL+Clustering at 80% training with at least 10 random seeds, and report the mean and standard deviation of RMSE plus a paired significance test (e.g., Wilcoxon signed-rank) against RF. If the FEDL+Clustering advantage over RF is less than ~0.2 kWh and not significant, the claimed superiority is noise.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract's central quantitative claim is that the proposed approaches improve energy-demand prediction accuracy by up to 24.63%. The paper's own Table I does not support this figure under any straightforward formula. For the 80% training split, FEDL+Clustering has RMSE 5.76; the best baseline (RF) has RMSE 6.35, giving an improvement of (6.35−5.76)/6.35 ≈ 9.3%, while the worst baseline (KNR, 7.18) gives ≈ 19.8%. Across all training splits, the largest improvement over any single baseline is (7.71−5.78)/7.71 ≈ 25.0% at the 70% split, not 24.63%. The average-baseline comparisons give roughly 12–14%. The paper never defines the formula used for 'improvement', and the text in Section IV-C claims 24.28% and 24.63% at 80% training, which does not match the table. Additionally, the reported RMSE differences among EDL, FEDL, and their clustered variants are within 0.1 kWh (about 1–2%), with no error bars, seeds, or cross-validation, so it is unclear whether FEDL is even statistically distinguishable from EDL. Because the headline improvement is the paper's primary advertised contribution and it cannot be reproduced from the presented data, the central claim is not verifiable. The communication-overhead reduction of 83.4% is similarly unsupported: Fig. 4 gives no definition of what is measured (data volume? number of messages?), no units for the baseline, and no calculation. These are addressable presentation issues, but as printed they make the central results uncheckable.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes three approaches for predicting EV charging-station energy demand: a centralized energy demand learning (EDL) method, a federated energy demand learning (FEDL) method in which charging stations share only local gradients with a charging station provider, and clustering-based variants of both. The methods are evaluated on a public dataset of 65,601 charging transactions from Dundee, UK, and compared with six baseline regressors (KNR, MLPR, SGDR, DT, SVR, RF) at training splits of 80%, 70%, 60%, and 50%. The abstract reports accuracy improvements up to 24.63% and communication-overhead reductions of 83.4% relative to the baselines. The paper also describes a customized constrained K-means clustering algorithm that groups charging stations by location while enforcing cluster-size bounds.","tokens_in":9357,"tokens_out":3185,"duration_ms":33965,"significance":"If the empirical claims are reproducible, the paper makes a useful contribution: it demonstrates that a federated-learning approach can provide competitive energy-demand prediction accuracy while avoiding raw data collection, which addresses both communication cost and privacy in EV networks. The use of a public real-world dataset and the inclusion of six independent baseline algorithms are strengths, and the clustering formulation with explicit size constraints is a reasonable design choice. However, the headline quantitative claims are not currently verifiable from the manuscript: the 24.63% accuracy improvement does not match any standard formula applied to Table I, and the 83.4% communication-overhead reduction is supported only by an undefined figure. The absence of error bars or repeated-trial statistics also leaves open whether the small differences among EDL, FEDL, and their clustered variants are meaningful. These issues are central because the paper's advertised contributions are precisely these numerical improvements.","major_comments":[{"comment":"The headline claim of up to 24.63% accuracy improvement is not reproducible from Table I under any straightforward formula. For the 80% split, the best baseline (RF) has RMSE 6.35 and FEDL+Clustering has RMSE 5.76, giving an improvement of (6.35−5.76)/6.35 ≈ 9.3%; against the worst baseline (KNR, 7.18) the improvement is ≈ 19.8%. The only value close to 24.63% is at the 70% split using the worst baseline: (7.71−5.78)/7.71 ≈ 25.0%, which is not 24.63%. The text in Section IV-C states 24.28% and 24.63% for the 80% split without defining the baseline or the percentage formula. Please provide the exact computation, either as per-baseline percentages or with an explicit formula such as (RMSE_baseline − RMSE_proposed)/RMSE_baseline, and report the corresponding values for each training split.","section":"§IV-C, Table I"},{"comment":"The communication-overhead reduction of 83.4% is not supported by the presented evidence. Fig. 4 shows bars on an axis labeled 'Communication overhead (MB)' but the manuscript does not define what is being measured (e.g., total bytes of raw data versus model/gradient updates, number of messages, or per-round traffic), what the baseline value is, how the baseline is computed for methods that are not centralized (e.g., KNR, which may not involve data transmission at all), or how the 83.4% figure is derived. Without these definitions and a numerical calculation, the overhead claim cannot be checked. Please specify the measurement protocol, report the underlying quantities, and show the arithmetic for the reduction.","section":"Fig. 4"},{"comment":"The reported RMSE differences among EDL, FEDL, EDL+Clustering, and FEDL+Clustering are very small (e.g., 5.86 vs. 5.81 vs. 5.77 vs. 5.76 at the 80% split, a range of 0.10 kWh, or about 1.7% relative). The paper provides no error bars, number of random seeds, cross-validation, or statistical significance tests. As a result, the claims that clustering 'further improves prediction accuracy' and that FEDL 'achieves the average prediction with less variance and lower error' are not substantiated: the observed differences could easily arise from training noise. Please report repeated runs with different seeds and, if possible, confidence intervals or pairwise significance tests.","section":"§IV-C, Table I"},{"comment":"The FEDL method rests on the assumption that synchronously averaging local gradients, as in Eq. (12), yields a global model with accuracy comparable to centralized training. This assumption is known to be fragile when local datasets are highly non-IID, and the paper does not characterize the distribution of transactions across the charging stations or provide any convergence analysis. Since the accuracy numbers in Table I are based on a specific (and unspecified) partition of the data into J subsets, it is unclear whether the claimed 1–2% gap between FEDL and EDL would persist under more heterogeneous partitions. Please report the distribution of transactions per station, describe how the J training subsets were created, and discuss or experimentally test the effect of non-IID data on FEDL convergence.","section":"§III-B, Eq. (12)"}],"minor_comments":[{"comment":"In Eq. (12), the summand is written as ∇υ(φ)_i but the summation index is j; this should be ∇υ(φ)_j.","section":"Eq. (12)"},{"comment":"The notation 'em_i' in Algorithm 1 is not defined anywhere; please explain what it represents (presumably a per-transaction error term).","section":"Algorithm 1"},{"comment":"The experimental setup does not state the values of J, K, θlow, θhigh, the number of epochs T, or the dropout rate's relationship to the hidden-layer sizes. These hyperparameters are needed to reproduce the experiments.","section":"§IV-B"},{"comment":"The sentence 'the RMSE obtained by the centralized EDL with clustering and the FEDL with clustering are 24.28% and 24.63% lower than those of the baseline methods' is ambiguous: it does not say which baseline or baselines are used. Please clarify.","section":"§IV-C"},{"comment":"The figure would be much more informative if the numerical values were printed on the bars and if the baseline category were disaggregated (e.g., showing the overhead for each centralized method).","section":"Fig. 4"}],"recommendation":"major_revision","confidential_remarks":"The manuscript addresses a relevant application area for federated learning, but the central quantitative claims are not yet verifiable. The authors should be asked to provide the explicit improvement formula, the communication-overhead measurement definition, and error-bar statistics. If these are supplied, the paper could become a solid empirical contribution. No concerns about novelty disclosure; the work appears to be original, though closely related to existing federated-learning applications to demand forecasting."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take: this is a sincere, incremental applied ML paper that does one new thing—benchmarks federated and clustered DNNs against six standard regressors on the public Dundee EV charging dataset—and the central qualitative claim holds. But the headline numbers do not come out of the paper’s own table, and the communication-overhead claim is undefined. Fix those and it becomes a reasonable applied paper.\n\nWhat’s actually new: not federated averaging or constrained k-means; both are established. The contribution is the combination and the evaluation on real charging-session data (65,601 transactions), including the clustered variants (K=2) that group stations by location before training. That specific comparison is not in the cited prior work. The paper lays out the centralized/federated/clustered design space clearly, and the algorithms are easy to follow.\n\nWhat it does well: the problem is real, the privacy motivation for federated learning is sensible for charging-station providers, and Table I consistently shows the proposed methods below all six baselines across four training splits. So the central direction—federated and clustered deep learning beat classical regressors here—is supported. The citations look appropriate; I don’t see any missing-key-work or citation-stacking red flags.\n\nSoft spots, in order of severity. (1) The abstract’s “up to 24.63%” and Section IV-C’s “24.28% and 24.63%” do not match Table I under any standard formula. At 80% training, the best single-baseline comparison gives about 9.3% vs RF and 19.8% vs KNR; the largest cross-split improvement is about 25.0% at 70%. The paper never defines the improvement formula. That is a real reproducibility gap in the primary advertised result. (2) No error bars, seeds, or cross-validation. The differences among EDL, FEDL, and their clustered variants are roughly 0.1 kWh (1–2%), so the claimed advantage of clustering over plain EDL/FEDL is not statistically supported. (3) Figure 4’s “communication overhead (MB)” is undefined—data volume? number of messages?—so the 83.4% reduction is unverifiable. (4) One dataset, one clustering K, and several free hyperparameters; the non-IID concern about synchronous gradient averaging is plausible but not analyzed.\n\nNone of these sink the paper. They are presentation and evidence-standards issues in an otherwise straightforward empirical study.\n\nWho this is for: people working on EV charging prediction or privacy-preserving demand forecasting who want a baseline comparison and a manageable federated-learning application. It does not change the science, but it is a useful data point. I would give it a serious referee, with a clear request to fix the percentage arithmetic, define the communication metric, and add uncertainty estimates or code.\n\nRecommendation: send to peer review, expect revision, and make the authors reconcile the abstract with Table I before acceptance.","headline":"An honest incremental federated-learning benchmark for EV charging demand whose qualitative claims hold but whose headline percentages and overhead metric are not reproducible from the presented data.","tokens_in":9892,"tokens_out":2272,"would_cite":false,"duration_ms":23493,"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":"Federated learning cuts EV charging forecast error by 24.63%","keywords":["energy demand prediction","electric vehicles","charging stations","federated learning","clustering","deep learning","communication overhead","privacy"],"falsifier":"A reader could re-run the FEDL algorithm on the Dundee dataset or a similar one, but with an artificially unbalanced split (e.g., one station holds 80% of transactions) and compare the RMSE against a centralized model trained on the same data; if the federated model's error exceeds the centralized model's error by more than the claimed 24.63% margin, the synchronous-averaging assumption fails in that regime.","tokens_in":8843,"feed_emoji":"⚡","tokens_out":4284,"duration_ms":37269,"temperature":0.7,"pith_summary":"The paper claims that federated deep learning can predict the electricity demand of electric-vehicle charging stations as accurately as a central model that gathers all data, while sending only model updates over the network. On a real dataset of 65,601 charging sessions in Dundee, UK, the proposed methods reduced prediction error (RMSE) by up to 24.63% relative to six conventional machine-learning baselines, and cut communication overhead by 83.4%. This matters because charging-station providers need accurate forecasts to reserve energy in advance and avoid grid congestion, and because privacy concerns can block central collection of driving and charging data.","feed_headline":"Federated learning cuts EV charging forecast error by 24.63%","feed_subtitle":"Stations share only model updates, cutting communication overhead 83.4% and preserving user privacy.","key_machinery":"The machinery is a multilayer perceptron with two hidden layers of 64 neurons each, tanh activation, and a dropout layer, trained with the Adam optimizer. The federated variant adds a synchronous averaging step that combines local gradients from charging stations into a global model, while the clustering variant wraps the learner in a constrained K-means algorithm that partitions stations by latitude and longitude so that each cluster learns its own model. The paper's quantitative claims rest on these components working together on the Dundee charging-session dataset.","core_discovery":"The central discovery is that federated energy demand learning (FEDL) — where each charging station trains a local deep neural network and sends only its gradients to a central provider that averages them — can outperform both conventional regressors and a fully centralized deep network in forecasting transaction-level energy demand. Adding a location-based constrained K-means clustering step before learning further reduces error by grouping stations with similar features, and the federated design sends no raw data, shrinking communication overhead by 83.4% compared to central collection.","pith_inferences":["The experiments use a single city dataset with 58 stations, so the claimed gains may not transfer to networks where stations have highly skewed transaction counts, where the synchronous aggregation could slow or bias convergence.","A natural testable extension is to compare FEDL against a centrally trained model on the same data with a non-IID split that mimics real-world station heterogeneity; the paper does not report that baseline.","The privacy argument rests on gradients being hard to reverse-engineer; for small batches, gradient leakage is a known risk the paper does not address."],"forward_implications":["Charging-station providers can offer real-time demand forecasts without collecting raw transaction logs, preserving user privacy.","Communication load drops to roughly one-sixth of centralised collection, enabling frequent model updates even over bandwidth-limited links.","Location-based clustering is a cheap way to reduce bias when demand patterns differ by area, and it can be layered on any federated or central learner.","The same pipeline could be applied to other sparse, spatio-temporal demand forecasting problems such as bike-sharing or warehouse energy use."],"supporting_citations":[{"why":"Supplies the real charging-session dataset from Dundee used in all experiments.","marker":"[12]"},{"why":"Provides the Adam optimizer used for gradient updates in both EDL and FEDL.","marker":"[15]"},{"why":"Is the distributed deep-network architecture that the federated averaging design builds on.","marker":"[16]"},{"why":"Supplies the constrained K-means method adapted for location-based clustering of charging stations.","marker":"[17]"},{"why":"Identifies the six baseline regression algorithms compared against the proposed methods.","marker":"[18]"}],"fun_headline_variants":["Federated EV forecast improves 24.63%, cuts data traffic 83.4%","Privacy-first federated learning boosts EV charging forecast accuracy","EV stations federate: 24.63% better prediction, 83.4% less overhead","Federated learning slashes EV energy forecast error and communication cost","Clustered federated learning: EV demand prediction up, data sharing down"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that averaging gradients from charging stations with very different numbers of transactions and demand patterns produces a global model that predicts as well as a model trained directly on all the data together.","fun_headline_variants_meta":{"raw":{"variants":["Federated EV forecast improves 24.63%, cuts data traffic 83.4%","Privacy-first federated learning boosts EV charging forecast accuracy","EV stations federate: 24.63% better prediction, 83.4% less overhead","Federated learning slashes EV energy forecast error and communication cost","Clustered federated learning: EV demand prediction up, data sharing down"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000526,"raw_usage":{"total_tokens":2515,"prompt_tokens":895,"completion_tokens":1620,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":511,"completion_tokens_details":{"reasoning_tokens":1517}},"tokens_in":511,"tokens_out":1620,"duration_ms":13792,"temperature":1.0,"reasoning_tokens":1517,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T05:32:48.871071+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A reader could re-run the FEDL algorithm on the Dundee dataset or a similar one, but with an artificially unbalanced split (e.g., one station holds 80% of transactions) and compare the RMSE against a centralized model trained on the same data; if the federated model's error exceeds the centralized model's error by more than the claimed 24.63% margin, the synchronous-averaging assumption fails in that regime.","supporting_citations":[{"cited_title":"Available Online: https: //data.dundeecity.gov.uk/dataset/ev-charging-data","cited_arxiv_id":null,"evidence_quote":"Supplies the real charging-session dataset from Dundee used in all experiments."},{"cited_title":"Adam: a method for stochastic optimization,","cited_arxiv_id":null,"evidence_quote":"Provides the Adam optimizer used for gradient updates in both EDL and FEDL."},{"cited_title":"Large scale distributed deep networks,","cited_arxiv_id":null,"evidence_quote":"Is the distributed deep-network architecture that the federated averaging design builds on."},{"cited_title":"Constrained K-means clustering,","cited_arxiv_id":null,"evidence_quote":"Supplies the constrained K-means method adapted for location-based clustering of charging stations."},{"cited_title":"A comprehensive survey on machine learning for networking: evolution, applications and research opportunities,","cited_arxiv_id":null,"evidence_quote":"Identifies the six baseline regression algorithms compared against the proposed methods."}],"review_version":1}