{"id":"015e5e82-0678-4e65-a6df-54cc7cbf7734","arxiv_id":"2501.09713","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A group-fair trading layer using Wasserstein distances redistributes peer-to-peer electricity trades, reducing measured inequality by up to 70.1% on a 1600-peer IEEE 33-bus case study.","lead":"This paper introduces an optimization mechanism that redistributes peer-to-peer electricity trades so that different energy-poverty groups end up with similar trading volumes. It reports cuts in a Wasserstein-based unfairness measure of up to 70.1%, with near-total elimination when a non-profit community solar plant is added.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Algorithm 1 reports an unfairness value that is not the true Wasserstein distance of the output trade matrix; Tables 2–3 may not support the claimed reductions.","rationale":"The reader's weakest assumption concerned convergence of the alternating algorithm to a global optimum of the bilinear problem. The concern raised here is more fundamental and more load-bearing: the objective value reported as 'unfairness level' is not the true Wasserstein distance of the final transaction matrix. This affects every quantitative claim in the abstract and Section 5, including the headline 70.1% reduction, the 23% average reduction, and the 'full elimination' statement. The issue is not merely suboptimality; even if the fixed-π subproblems were solved to global optimality, the averaging in line 14 and the use of an upper bound for D2max mean the reported Dmax need not equal the true unfairness of the output X. This can be fixed by adding a post-processing evaluation step, but without it the central numerical results are unverified. No code or data are provided, so independent verification is not possible from the manuscript alone. The verdict should remain CONDITIONAL, but the required condition must be strengthened: the authors must recompute the true max Wasserstein distance for every reported final X and confirm that the reductions and plateaus in Tables 2 and 3 remain valid. This is a concrete, checkable correction rather than a rejection of the modeling framework, and it aligns with the reader's CONDITIONAL verdict while identifying a separate and more immediate threat to the claims.","tokens_in":14097,"tokens_out":6627,"duration_ms":67081,"concrete_test":"Run Algorithm 1 for the 08/07/2024 scenario at 18:00 with ε=50%, and record the final transaction matrix X and the reported Dmax (23.10). For this final X, compute T_g via Eq. (5), solve each pairwise Wasserstein problem (7a)–(7c) to optimality, and take the maximum over group pairs (Eq. (8)). Compare the resulting true value with 23.10. Repeat for every time slot and every ε in Tables 2 and 3. If any entry differs by more than 1%, the reported unfairness reductions and the alleged fairness limits are not supported by the output trades. Additionally, track the true max distance of the final X across iterations to confirm whether the sequence of true objective values is actually converging before the loop stops.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central quantitative claims hang on the Dmax values in Tables 2 and 3, but Algorithm 1 does not compute the true max Wasserstein distance of its final output. In each loop, D1max is the true distance of the current X (lines 7–10), then line 11 solves (9a)–(9k) with π fixed, producing a new X and D2max. Since the fixed π remains a feasible transport plan for the new X (its marginals depend only on group sizes), D2max is merely an upper bound on the true distance of the new X, not the true value. The loop stops when |D1max−D2max|≤tol, and line 14 sets the reported Dmax to (D1max+D2max)/2. The final X (from the last line-11 solve) is never evaluated for its actual unfairness. Consequently, each entry in Tables 2 and 3, and the derived claims of 70.1% and 23% reductions, are computed from a quantity that can differ materially from the true max Wasserstein distance of the trading outcome. Moreover, the termination criterion compares a true distance from one iterate with an upper bound from the next, so it does not certify that the true distances have stabilized. This is a correctness issue independent of local minima: even a globally optimal solve of the fixed-π subproblem would not make the reported Dmax the true objective.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an optimization-based mechanism for distributionally fair peer-to-peer electricity trading. Peers are partitioned into groups by energy poverty level, and the mechanism minimizes the maximum Wasserstein distance between the distributions of traded electricity across groups, subject to grid constraints, balance constraints, and a user-specified profit sacrifice level. The resulting bilinear problem is solved with an alternating algorithm. A case study on an IEEE 33-bus system with 1600 peers reports reductions in the unfairness metric of up to 70.1% at certain hours and near-elimination when a non-profit community PV plant is added. The paper also analyzes how the unfairness level varies across the day and with tariff structures.","tokens_in":14342,"tokens_out":11584,"duration_ms":111968,"significance":"The core idea of defining group-level fairness through the Wasserstein distance between transaction distributions is a useful and original contribution to P2P market design. The paper is careful to include grid constraints through the LinDistFlow approximation, and the case study is large and realistic, which strengthens the practical relevance. The authors explicitly acknowledge the NP-hard nature of their model and the local-minimum limitation of their alternating algorithm. If the technical issues below are corrected, the mechanism could be a valuable tool for community managers who want to trade off collective fairness against profits. However, the reported quantitative claims currently rest on a mis-specified profit term and on an algorithm whose output is not the true unfairness of the final trade matrix.","major_comments":[{"comment":"The buyer-side profit term is mis-specified. In Eq. (4), the profit of peer i from buying X_ji should include the utility price for that buyer, λ^us_i, not the seller's utility price λ^us_j. As written, the term X_ji(λ^us_j - (λ^s_j + λ^b_i)/2) values the purchase using the wrong utility price. In Eq. (9c), the analogous term is X_ji(λ^us_j - (λ^s_i + λ^b_j)/2), which also uses the wrong bid/ask average; the settled price for a transaction where j sells to i should be (λ^s_j + λ^b_i)/2. Because Eq. (9c) defines the per-group profit sacrifice constraint, all results that depend on ε (Tables 2-4, the 70.1% claim, and the 'fairness limit' plateaus) are not guaranteed to correspond to the stated sacrifice levels. The indices must be corrected and the case study re-run.","section":"Eq. (4) and Eq. (9c)"},{"comment":"The reported D_max is not the true maximum Wasserstein distance of the output trade matrix. At each iteration, D1_max is the true distance of the current X (computed via T, d, and the optimal π), but then line 11 solves (9a)-(9k) with π fixed, producing a new X and D2_max. For that new X, the fixed π is only a feasible transport plan, so D2_max is an upper bound on the true distance, not its value. The loop terminates when |D1_max - D2_max| ≤ tol, and line 14 reports the average of the two. The final X from the last line-11 solve is never re-evaluated for its true unfairness. Consequently, the entries in Tables 2 and 3, and the derived reduction percentages, are computed from a quantity that can differ from the true max Wasserstein distance of the reported outcome. The termination criterion also compares a true distance from one iterate with an upper bound from the next, so it does not certify stabilization of the true distances. The algorithm should be modified to compute the true D_max of the final X (by solving (7a)-(7c) for that X) and to base termination on true consecutive values.","section":"Algorithm 1, lines 7-14"},{"comment":"The definition of s^{g,g'}_{ij} is inconsistent with the index sets. The equation states s^{g,g'}_{ij} = T^{g'}_i - T^g_j for i∈g and j∈g', but T^{g'}_i is not defined because i does not belong to g'. The intended absolute-value representation should be s^{g,g'}_{ij} = T^{g'}_j - T^g_i (or equivalently the negative of t). As printed, the constraints (9h) do not enforce d^{g,g'}_{ij} = |T^g_i - T^{g'}_j|, and the model would not compute the correct Wasserstein distance. This needs to be corrected and the implementation checked against the corrected equation.","section":"Eq. (9g)"},{"comment":"The abstract claims 'even achieving a full elimination when including a non-profit community photovoltaic plant', but Table 3 shows strictly positive unfairness values in all hours, with the 20 kW case yielding 0.01 kWh at 09:00 and 0.75 kWh at 18:00. The main text (Section 5.3) correctly says 'nearly complete elimination achieved at certain hours'. The abstract and the Introduction overstate the result; they should be aligned with the actual table values.","section":"Abstract and Table 3"}],"minor_comments":[{"comment":"The claim that the fair model reduces unfairness 'by 23% on average' is not clearly defined. It does not match the reduction implied by the total row in Table 2 (from 1,117.22 to 907.16, about 18.8%), and it is ambiguous whether the average is over hours, over dates, or computed differently. The authors should specify the calculation.","section":"Section 5.2"},{"comment":"The paper uses both 'sell to the highest bidder' and a reference model that maximizes sellers' revenue, but the objective in (1a) also includes the utility buyback price λ^ub_i for each seller. The exact interpretation of the reference profit should be clarified, and the derivation of the reference model should be expanded.","section":"Throughout"},{"comment":"The tariff data for the two dates are described, but the dates are not clearly connected to the rest of the case study. The reader has to infer that 15/10/2022 and 08/07/2024 are the simulation days. A short statement in Section 4.1 would improve readability.","section":"Section 4.1 and Fig. 3"},{"comment":"The conclusion acknowledges local minima, but it does not discuss the possibility of multiple local optima or the sensitivity of the results to the initialization procedure described in Section 5.2. Since the paper already mentions warm-starting from adjacent ε, a brief sensitivity discussion would help the reader trust the reported plateaus.","section":"Section 6"},{"comment":"The computational times vary widely (e.g., 1492 s at 12:00 for ε=2% vs. 96 s at 09:00 for ε=100%), but no solver version, hardware, or MIP gap (if any) is provided. Adding these details would make the performance claims reproducible.","section":"Table 4"}],"recommendation":"major_revision","confidential_remarks":"The paper has a promising core idea, but the technical issues in the profit constraint (Eq. 9c) and in Algorithm 1's output evaluation are load-bearing: they directly affect the numbers in Tables 2 and 3 and the headline claims. I believe these are fixable within the scope of a revision, so I recommend major revision rather than rejection. I would also ask the authors to benchmark their alternating algorithm on small instances against a global optimization solver (e.g., a spatial branch-and-bound method) to quantify the gap caused by local minima, since the paper's own conclusions acknowledge that global optimality cannot be guaranteed. The abstract's overstatement relative to Table 3 should be corrected before the paper is accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Thanks for the report. I read it against the paper, and the stress-test concern holds. Algorithm 1 never evaluates the actual unfairness of the transaction matrix it outputs. In each loop, D1max is the true distance of the current X, but then line 11 fixes π and solves (9a)-(9k) to get a new X and D2max. That fixed π is feasible for any X because its marginals depend only on group sizes, so D2max is an upper bound on the true distance of the new X, not the true value. Line 14 averages the two, and the final X is never checked. Tables 2-3 and the 70.1% and 23% claims are therefore built on a quantity that can be materially larger than the actual unfairness. This is independent of the local-minima issue: even exact subproblem solves leave the bug. The fix is straightforward—evaluate the true max Wasserstein distance of the final X and report that—but without it the main quantitative results are not trustworthy.\n\nWhat the paper does well: it brings Ye et al.'s distributionally fair max-Wasserstein objective into a P2P clearing model with energy-poverty groups and a user-set profit sacrifice. The LP formulation is mostly clear, the constraints on group profit, curtailment, and utility trades are sensible, and the IEEE 33-bus case with 1600 peers is a real stress test. The authors also honestly flag the local-minima limitation in the conclusions. That is a legitimate new application, not a new theory.\n\nThe buyer-side profit term in Eq. (4) and (9c) uses λus_j where it should use λus_i for the importing peer, and the bid indices in (9c) look swapped too. This mis-specifies group profits, and since the sacrifice level ε is defined relative to those profits, the fairness/efficiency tradeoff could shift. Minor but should be fixed. The abstract's 'full elimination' also overstates the 0.01 kWh floor in Table 3; the body correctly says 'nearly complete.' No code or data are provided, so I could not check the numerics independently.\n\nOverall: the central idea is coherent and the flaws are addressable. As-is, I would not rely on the headline numbers, but the paper deserves referee time and a request for corrected tables and, ideally, code. Send it to review rather than desk reject.","headline":"A useful group-level fairness mechanism for P2P electricity markets, but the headline reductions are computed from an upper bound, not the true Wasserstein distance of the final solution.","tokens_in":721,"tokens_out":1750,"would_cite":false,"duration_ms":42232,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C26","90C90"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proposes an optimization mechanism that redistributes peer-to-peer electricity trades to minimize the largest Wasserstein distance between groups defined by energy poverty, cutting unfairness by up to 70.1% at a chosen…","keywords":["peer-to-peer electricity trading","distributional fairness","Wasserstein distance","energy poverty","alternating algorithm","bilinear optimization","IEEE 33-bus","community PV plant"],"falsifier":"Solve the same hourly instances (08/07/2024, each epsilon value) with an exact mixed-integer linear reformulation of the bilinear Wasserstein constraint and compare the optimal $D_{\\max}$ with the alternating algorithm's values; if the exact optima are materially lower, the reported reductions come from the heuristic, not from the model.","tokens_in":13827,"feed_emoji":"⚡","tokens_out":4492,"duration_ms":49596,"temperature":0.7,"pith_summary":"This paper asks whether a peer-to-peer electricity market can trade fairly across groups with different energy poverty levels, not just efficiently. It defines unfairness as the maximum Wasserstein distance between the distributions of traded electricity across groups, then proposes an optimization model that reshuffles trades to shrink that distance while limiting how much profit the community gives up. On a simulated 1600-household community on the IEEE 33-bus grid, the model cuts unfairness by up to 70.1% at one hour and by 23% on average, with a user-chosen sacrifice parameter. The paper also shows that a small non-profit community photovoltaic plant can push unfairness close to zero. If correct, the work offers a practical lever for community energy managers who want equity without fully abandoning market outcomes.","feed_headline":"P2P power trading cut group unfairness by up to 70%","feed_subtitle":"An optimization model reshuffles community electricity trades to protect poor households, sacrificing at most a chosen slice of profit.","key_machinery":"The central object is the transaction matrix $X$, where $X_{ij}$ is the electricity sold by peer $i$ to peer $j$, from which each group $g$'s transaction distribution $T^g_i = \\sum_{j\\in\\mathcal{P}}(X_{ij}+X_{ji})$ is built by summing traded electricity per peer. Unfairness is measured as $D_{\\max} = \\max\\{W(g,g')\\}$, the largest Wasserstein distance between any two group distributions, where each $W(g,g')$ is the minimum transportation cost between the two distributions. The fairness model minimizes $D_{\\max}$ subject to energy-balance, bid-matching, and LinDistFlow grid constraints, plus a group-level profit floor $(1-\\epsilon)|\\gamma^*_g|$, a utility-trade cap, and a curtailment cap. The Wasserstein distance introduces a bilinear term $\\pi^{g,g'}_{ij} d^{g,g'}_{ij}$ because both the transportation plan and the transaction distances are variables, so the paper solves the nonconvex problem with an alternating algorithm that fixes one variable set, solves the resulting LP, then swaps and iterates until the two $D_{\\max}$ values agree within tolerance.","core_discovery":"Under selfish peer-to-peer trading, electricity trades concentrate among energy-rich households, and the group that is discriminated against changes with tariff conditions. The paper's central claim is that a distributionally fair trading model can substantially reduce this group-level inequality: it minimizes the maximum Wasserstein distance among the transaction distributions of groups defined by energy poverty, subject to a user-selected profit-sacrifice level epsilon and constraints that prevent fairness from being achieved by shifting trades to the utility or by extra curtailment. Solving this bilinear problem with an alternating algorithm on the IEEE 33-bus test system with 1600 peers, the model reduces the unfairness level by up to 70.1% at 18:00 and by 23% on average, and adding a 20 kW non-profit community PV plant lowers unfairness by 51.95% on average at epsilon = 1%, with near-complete elimination at some hours. The model also identifies which group is currently discriminated against, and this target group changes over time depending on surplus and tariffs.","pith_inferences":["Editorial inference: Because profits are constrained only at the group level, two households within the same group could still end up with very different trade volumes; a within-group dispersion constraint would be a natural check that this model does not provide.","Editorial inference: The plateau behavior suggests that when total surplus is structurally smaller than total deficit, reallocating trades alone cannot eliminate inequality; changing the resource endowment, as the community PV plant does, is the lever that pushes unfairness toward zero.","Editorial inference: The same Wasserstein-minimax objective could apply to other community resource-allocation settings, such as water or data-bandwidth sharing, wherever per-agent allocations can be grouped and feasibility constraints are linear.","Editorial inference: A practical deployment could track $D_{\\max}$ in real time from the transaction matrix alone, without solving the fair model, and trigger re-optimization only when the fairness metric rises above a threshold."],"forward_implications":["Community managers can reduce group-level trading inequality by a specified amount while capping profit loss, since epsilon is a user-selected parameter.","The model identifies the discriminated group in each hour, and that target group changes with market conditions, so fairness actions can be adaptive rather than fixed.","Unfairness decreases as the allowed sacrifice grows, but only up to a per-hour fairness limit; beyond that limit, additional sacrifice yields no further improvement.","Adding a non-profit community PV plant reduces unfairness further, with a 20 kW plant cutting average unfairness by 51.95% relative to the reference at epsilon = 1%.","The same mechanism transfers to other group definitions, such as tariff type, PV ownership, or consumption patterns, because the group partition is an input to the model."],"supporting_citations":[{"why":"Supplies the semi-decentralized community-manager architecture used for the P2P market.","marker":"(Moret and Pinson, 2019)"},{"why":"Supplies the LinDistFlow branch-flow model used for voltage and line constraints.","marker":"(Farivar and Low, 2013)"},{"why":"Source of the distributionally fair Wasserstein-distance criterion and the sacrifice-level idea.","marker":"(Ye et al., 2024)"},{"why":"Provides the normalized consumption and PV production profiles used in the case study.","marker":"(IEA, 2019)"},{"why":"Provides the flat and double tariff prices used to construct peer bids.","marker":"(ENOVOS, 2021)"},{"why":"Documents that peers without generation never sell, motivating the need for group-level fairness.","marker":"(Lei et al., 2023)"}],"fun_headline_variants":["Fair P2P power trading cuts group bias by 70%","P2P power trading algorithm eliminates up to 70% unfairness","Fairness for all: P2P power trading slashes inequality by 70%","P2P power trading: up to 70% less group unfairness","Fair P2P trading: 70% less inequality between energy groups"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the alternating algorithm's output is close enough to the true optimum that the reported unfairness levels and plateaus reflect the model's real capability; the paper states in the Conclusions that the algorithm can stop at local minima and that global optimality is not guaranteed.","fun_headline_variants_meta":{"raw":{"variants":["Fair P2P power trading cuts group bias by 70%","P2P power trading algorithm eliminates up to 70% unfairness","Fairness for all: P2P power trading slashes inequality by 70%","P2P power trading: up to 70% less group unfairness","Fair P2P trading: 70% less inequality between energy groups"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000723,"raw_usage":{"total_tokens":3230,"prompt_tokens":918,"completion_tokens":2312,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":534,"completion_tokens_details":{"reasoning_tokens":2213}},"tokens_in":534,"tokens_out":2312,"duration_ms":16357,"temperature":1.0,"reasoning_tokens":2213,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T19:44:54.305589+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Solve the same hourly instances (08/07/2024, each epsilon value) with an exact mixed-integer linear reformulation of the bilinear Wasserstein constraint and compare the optimal $D_{\\max}$ with the alternating algorithm's values; if the exact optima are materially lower, the reported reductions come from the heuristic, not from the model.","supporting_citations":[],"review_version":1}