{"id":"007e4c27-8b76-4811-87fe-7be60938f0ee","arxiv_id":"2506.20401","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"EVOP-V2G models a single EV driver maximizing profit by choosing ride orders and timing charging/discharging under time-varying prices; MIP, evolutionary, and large-neighborhood-search solvers are evaluated on Melbourne data.","lead":"The paper introduces a new optimization problem for electric-vehicle ride-hailing and delivery drivers who can both charge and sell energy back to the grid, and proposes three algorithms to solve it. Experiments on Melbourne data suggest these methods can roughly double a driver's profit compared to a simple greedy baseline, with vehicle-to-grid trading contributing around 20% of earnings under default prices.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"MIP benchmark is dimensionally inconsistent: Eq. (17) and objective (1) omit the charging rate P_i and timeslot length δ, leaving the 'near-optimal' claim unanchored.","rationale":"The reader's weakest_assumption pointed to unspecified MIP parameters (δ, |T|, λ) and the two-revisit restriction. My concern is sharper: even with arbitrary values, the printed equations cannot represent the stated problem because the station rate P_i never enters the objective or the energy-balance constraint, making Eq. (17) dimensionally inconsistent and Eq. (1) independent of charging duration and rate. This internal inconsistency is more load-bearing than the weak baseline criticism, because a weak baseline would still permit a valid relative comparison, whereas an incorrect MIP invalidates the exact-benchmark interpretation of the experiments. The paper's own limitation statements—the placeholder code URL and the restriction to two station revisits for MIP—further support the need for a corrected, runnable validation. I keep the reader's CONDITIONAL verdict rather than moving to REJECT because the flaw is in the written model and can in principle be fixed by correcting the equations, instantiating δ, λ, and P_i, and re-running the small-instance comparisons. If the proposed test shows the two MIP versions differ materially, the verdict should be reconsidered as REJECT for the current version.","tokens_in":25402,"tokens_out":6759,"duration_ms":83064,"concrete_test":"Re-solve one small-scale instance with two versions of the MIP: (A) exactly as printed, and (B) with energy-correct terms P_i·δ·rc_tk_i and P_i·δ·dc_tk_i in Eq. (17) and in objective Eq. (1), using δ=0.25 h, λ=0.9, and the actual charging rates from Section 5.1 (7 kW at home and the PlugShare station rates). Compare the optimal objective values and selected routes. If version B's optimum differs materially from version A's, the published MIP is not a faithful model of the stated problem, and the near-optimality claim is unresolved.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that EA and LNS are near-optimal because their profit is close to the MIP optimum. That anchor fails at the level of the printed model. In Eq. (17), battery updates are written as Σ_{tk∈T} δ·rc_tk_i − Σ δ·dc_tk_i: δ is a time-slot length (Table 2) and rc/dc are binary, so the term has units of hours, while b_j is battery energy in kWh. The station charging rate P_i, introduced in Section 3.1, appears neither in this constraint nor in the objective Eq. (1), where each discharging timeslot contributes PD_tk_i dollars regardless of P_i or δ. The efficiency parameter λ in Eqs. (16)–(17) is never defined or assigned a value. As written, the MIP cannot know how much energy is transferred in a timeslot, so its 'optimal' objective and the V2G profit decomposition in Section 5.2.4 are not computed from the physical quantities the paper says it optimizes. Since LNS's internal MIP refinement (Section 4.3.4) inherits the same equations, the near-optimality comparison is not anchored to a well-defined optimum. This is a correctness flaw in the formal model, not merely a missing reproducibility detail.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces the Electric Vehicle Orienteering Problem with V2G (EVOP-V2G): a single-EV profit-maximization problem in which a driver selects customer orders and decides when and where to charge or discharge a battery under time-varying electricity prices, station-dependent charging rates, time windows, and range constraints. The authors propose a Mixed Integer Programming (MIP) model, an Evolutionary Algorithm (EA), and a Large Neighborhood Search (LNS) with adaptive strategy selection and a MIP-based refinement step. The experiments use real-world ride and charging-station data from Melbourne, reporting that EA and LNS achieve near-optimal profit versus the MIP on small instances, that all proposed methods more than double the profit of a greedy baseline, and that V2G contributes about 20% of profit under default settings.","tokens_in":25735,"tokens_out":10748,"duration_ms":111485,"significance":"The problem variant is timely and practically relevant, and the paper's use of realistic data from Melbourne and PlugShare is a strength. The LNS design—adaptive strategy weighting, randomized greedy insertion, and MIP post-optimization—is methodologically plausible. If the formulation and numerical claims were corrected, the paper would provide a useful benchmark problem and evidence that V2G-aware scheduling can materially increase EV driver profit. However, the printed MIP is dimensionally inconsistent, the 'near-optimal' claim rests on a restricted MIP benchmark, and several key parameters are unspecified; these issues are load-bearing for the central claims.","major_comments":[{"comment":"The MIP model as printed does not define the objective in monetary units. In Eq. (1), charging/discharging decisions are binary indicators rc_tk_i and dc_tk_i multiplied by prices in $/kWh, with no charging rate P_i (kW) or timeslot length δ (h); the resulting terms have units of $/kWh, not $. In Eq. (17), the battery update adds/subtracts δ·rc and δ·dc, which have units of hours, to b_j, which is in kWh, and Constraints (19)–(20) repeat the same unit error. The charging rate P_i introduced in Section 3.1 appears nowhere in the constraints, and the efficiency parameter λ in Eqs. (16)–(17) is never defined or assigned a value. As written, the MIP cannot determine how much energy is transferred in a charging or discharging timeslot, so the 'optimal' objective and the V2G profit decomposition in Section 5.2.4 are not computed from the physical quantities the paper says it optimizes. This must be corrected (e.g., by multiplying each charging/discharging binary by P_i·δ) and the experiments rerun.","section":"4.1, Eqs. (1) and (17)–(20)"},{"comment":"The small-instance optimality anchor is not the true EVOP-V2G optimum: the authors state that EA/LNS solutions used up to three station revisits while the MIP was restricted to at most two revisits because of solver time. The 'near-optimal' gap is therefore measured against a restricted MIP, not against the unrestricted optimum of the problem. The claim would be sound only if the heuristics never benefit from a third revisit or if the gap to the unrestricted optimum is otherwise bounded; neither is shown. Please report the distribution of the number of station revisits in heuristic solutions and quantify the profit gap between the 2-revisit MIP and the heuristics, or restrict the heuristics to two revisits to make the comparison symmetric.","section":"5.2.1"},{"comment":"Several parameters required to instantiate the MIP are never specified: the timeslot length δ, the number of slots |T|, and the efficiency λ in Eqs. (16)–(17). Section 3.1 says the planning horizon T is 'set to 24 hours' by default, but δ (and hence |T|) is not reported, and the timing constraints (11)–(13) and battery constraints (17)–(20) depend directly on δ and |T|. Without these values, the reported MIP runtimes, profits, and the comparison in Section 5.2.1 are not reproducible. Please state δ and |T| explicitly in Section 5.1 and, if δ is part of the experimental design, explain how it was chosen.","section":"5.1"},{"comment":"The LNS 'Improving Solution Using MIP Solver' step reuses the MIP model of Section 4.1 on a path graph, so it inherits the dimensional and parameter issues described above. If the implementation actually uses a corrected model (e.g., with P_i·δ factors), the paper should present that corrected model explicitly, because it is the model whose solutions are reported. As printed, the MIP refinement step is not well defined.","section":"4.3.4"},{"comment":"Equation (24) defines the EA/LNS fitness as a sum of prices over timeslots without multiplying by the energy transferred per timeslot (P_i·δ). This is consistent with the unit error in the MIP, but it is inconsistent with the paper's statement that charging rates are station-dependent. Consequently, the reported profit values (including the V2G profit share in Section 5.2.4) are not actual dollar amounts unless an implicit normalization is being used. Please clarify the exact formula used in the implementation and ensure it matches the corrected model.","section":"4.2.1, Eq. (24)"}],"minor_comments":[{"comment":"In the third bullet of the Worst Profit Removal strategy, the discharging profit is written as Σ (PC_tk_i); it should be PD_tk_i, the discharging price, rather than the charging price PC_tk_i.","section":"4.3.2, 'Worst Profit Removal' bullet"},{"comment":"The text says the home charging infrastructure has a 'charging rate of 7 kw/h'; the unit should be kW (kilowatts), not kW/h, because the charging rate is a power, not an energy rate per hour.","section":"5.1"},{"comment":"The reproducibility link is given as the placeholder 'https://will-publish-after-acceptance.com'; a working repository is needed for the experimental results to be verifiable.","section":"5.1"},{"comment":"The text mentions 'two redundant vertices in Vc' for source and destination charging, but no such dummy vertices are introduced in the formal model; please clarify that these are experimental constructs used to model station revisits.","section":"4.1"},{"comment":"Constraint (13) writes τ_d, whereas arrival time at the destination is defined as τ_i for vi∈Vd; please use consistent notation.","section":"4.1, Eq. (13)"},{"comment":"There are numerous typographical artifacts in the typeset text (e.g., 'A ffordable', 'e fficient', 'board stoke') that should be cleaned up in a final version.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The dimensional inconsistency in the MIP (Eqs. (1), (17)–(20)) is a correctness issue, not a presentation issue: as printed, the model cannot be the model whose solutions are reported, and the numerical experiments would need to be redone with a corrected formulation. The asymmetric revisit limit (2 for MIP vs. up to 3 for heuristics) and the unspecified δ, |T|, and λ further weaken the near-optimality claim. I recommend major revision rather than rejection because the problem definition and the heuristic machinery are potentially salvageable, and a corrected formulation plus new experiments could support the paper's claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The real contribution here is the problem variant: combining orienteering-style order selection with V2G discharging under time-varying prices and station-dependent rates. That combination is new, and the paper positions it well against the EVRP and EVOP literature. The EA and LNS descriptions are detailed enough to be worth studying, and the V2G profit share (about 20% under default settings) is a model output, not a fitted target. That is all real work.\n\nThe soft spot is the MIP, and the stress-test note is right. Equation (17) writes battery updates as sums of δ times binary charging/discharging indicators. δ is a timeslot length in hours; the indicators are 0/1; so the term has units of hours, not kWh. The station charging rate Pi never appears there or in the objective, and λ in (16)–(17) is never defined. As printed, the MIP cannot know how much energy transfers in a timeslot, so its “optimal” objective and the near-optimal comparison are not anchored to the physical quantities the paper says it optimizes. This is not a minor typo—it affects the objective and the battery dynamics, and since LNS inherits the same equations in its MIP refinement, the flaw propagates.\n\nAlso, the MIP was limited to two station revisits while EA/LNS used up to three. Even with a correct model, the claimed gap to true optimal would be provisional. And the baseline is extremely weak: it only discharges at the destination and otherwise greedily inserts orders or charges to full. Beating that by “two times” is a statement about the baseline, not strong evidence of quality.\n\nThat said, the core formulation is valuable and the heuristics look reasonable. The data setup from real Melbourne trip records and PlugShare is sensible. The paper deserves a serious referee, but only if the authors fix the MIP equations and either correct the revisits asymmetry or report the gap honestly. Right now it is a good problem statement with a central evaluation gap, not a demonstrated result.","headline":"Genuinely new V2G orienteering variant and plausible heuristics, but the MIP anchoring the near-optimal claim is dimensionally inconsistent and the baseline is too weak to support the headline profit gain.","tokens_in":26260,"tokens_out":2542,"would_cite":false,"duration_ms":29580,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C11","90C27","90B06"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that Vehicle-to-Grid capability roughly doubles a ride-hailing or delivery driver's profit when order selection and charge/discharge timing are optimised together, with energy arbitrage contributing about a fifth of the…","keywords":["Electric Vehicle Orienteering Problem","Vehicle-to-Grid","V2G","profit maximisation","mixed integer programming","large neighbourhood search","evolutionary algorithm","dynamic electricity pricing"],"falsifier":"Re-run the 30-order comparisons with $\\delta$ and $|T|$ fixed to explicit values, such as 15-minute slots over a 24-hour horizon, and with $\\lambda$ set to 1, then check whether the LNS and EA profits still sit within the reported margin of the MIP optimum; alternatively, benchmark the two heuristics against a stronger non-greedy planner with lookahead rather than the paper's greedy baseline, and see whether the profit advantage survives.","tokens_in":25225,"feed_emoji":"⚡","tokens_out":12201,"duration_ms":111863,"temperature":0.7,"pith_summary":"This paper introduces EVOP-V2G, a profit-maximisation problem in which a single electric-vehicle driver working in ride-hailing or delivery chooses which customer orders to accept, in which order to serve them, and when and where to charge or discharge the battery, under time-varying electricity prices and station-dependent charging rates. The paper's central argument is that Vehicle-to-Grid (V2G) capability, letting the EV sell stored energy back to the grid during high-price periods, adds a second revenue stream that earlier orienteering formulations ignore, so order selection and energy trading must be optimised together. The authors formulate the problem as a Mixed Integer Programming model and design two fast heuristics, an evolutionary algorithm and an adaptive large-neighbourhood search, then test them on real trip and price data. Their empirical claim is that both heuristics produce schedules with more than twice the profit of a greedy baseline, stay near the MIP optimum on small instances, and scale to 900 orders with 70 charging stations, with V2G contributing roughly 20% of profit under default settings. A reader should care because, if correct, the result means an EV driver's route and the grid's flexibility needs can be priced and scheduled as one optimisation rather than managed separately.","feed_headline":"EVs that sell power to the grid can double driver profit","feed_subtitle":"New algorithms pick orders plus charge and discharge times; V2G arbitrage alone adds about 20% of profit.","key_machinery":"The carrying object is the EVOP-V2G objective, the sum of accepted-order fares plus discharge revenue minus charging costs: $$\\sum_{v_i \\in V_o}\\sum_{v_j} x_{ij}p_i + \\sum_{v_i \\in V_c}\\sum_{t_k \\in T}\\left($dc^{{t_k}}$_i $P^{{D t_k}}$_i - $rc^{{t_k}}$_i $P^{{C t_k}}$_i\\right),$$ evaluated over a planning horizon discretised into time slots of length $\\delta$. The MIP encodes three constraint families, node visiting, travel time with order time windows, and battery-level tracking with a capacity bound, plus the rule that charging and discharging cannot occupy the same time slot at the same station. The scalable machinery is the large-neighbourhood search: destroy operators (random, worst-profit, closeness, and Shaw) remove blocks of actions, repair operators (max-profit, regret-$k$, closeness, and price) reinsert orders and charge or discharge actions, an adaptive weighting scheme learns which strategies work, and a small MIP re-optimises the charging schedule on the repaired path with a 5% suboptimality gap. That MIP-in-the-loop repair step is what the paper credits for LNS's advantage in discharge decisions.","core_discovery":"The paper's central claim is that bidirectional charging changes the economics of EV route planning: a driver's daily profit is not just order fares minus charging costs but includes an arbitrage term from discharging when prices are high, so route, order selection, and charge/discharge schedule must be optimised together. On instances built from real ride-hailing trip records and real charging and feed-in tariffs, the paper reports that its two heuristics achieve more than twice the profit of a greedy baseline; that V2G contributes about 20% of total profit under default settings, rising to 47.7% for the evolutionary algorithm and 66.4% for the large-neighbourhood search when charging prices and rates are tripled; and that the large-neighbourhood search solves 900-order, 70-station instances quickly while staying near the MIP optimum on 30-order instances. The paper also argues that LNS is the stronger heuristic in most settings because it re-optimises charging decisions with the MIP inside each repair step, while the evolutionary algorithm wins on some instances where random exploration helps escape local optima.","pith_inferences":["A testable extension the paper leaves implicit: re-running the same pipelines on tariff structures with larger peak/off-peak spreads, or on real-time prices, would separate how much of the profit gain is energy arbitrage versus better order selection; the 20%-to-66% V2G share range suggests arbitrage, not routing, is the dominant driver.","The LNS design of heuristic routing with an exact solver reserved for the energy subproblem is a modular recipe that could transfer to fleet-scale ride-hailing, where each vehicle's charge/discharge schedule becomes an independent small MIP once orders are assigned.","The unspecified slot length and horizon raise a reproducibility question: on markets with rapidly changing real-time prices, the granularity of $\\delta$ would directly control how much price volatility the driver can exploit, so results on time-of-use tariffs may not predict real-time-pricing performance.","The few instances where EA beats LNS share longer time windows, hinting that a hybrid that lets random exploration dominate early and switches to LNS-style exploitation later could outperform either method alone."],"forward_implications":["Accepted orders are no longer chosen by fare alone: orders that keep the vehicle near a cheap-charging, expensive-discharging station can be worth more, so route plans and energy trades have to be priced jointly.","At city scale (900 orders, 70 charging stations) a few minutes of computation suffices to produce a schedule that beats the greedy baseline by a factor of two, which makes the approach a plausible daily-scheduling tool for a single commercial driver.","Because the V2G share of profit climbs from about 20% to between 48% and 66% when charging prices and rates are tripled, the value of bidirectional scheduling depends mainly on the spread between peak and off-peak tariffs, not on absolute price levels.","The MIP allows at most two revisits to any charging station in the exact comparison, while the heuristics revisit cheap stations up to three times, so the reported near-optimality margins inherit that cap on station revisits."],"supporting_citations":[{"why":"Surveys EVRP, arc routing, and team orienteering problems in sustainable transportation, supplying the problem family that EVOP-V2G extends.","marker":"[7]"},{"why":"Presents an EVRP incorporating V2G and G2V options with time-variant prices, the closest prior setting that the paper generalises to order selection with time windows.","marker":"[18]"},{"why":"A reinforcement-learning EVRP with V2G supply that considers discharging only, cited to show bidirectional charging in routing is essentially unexplored for orienteering.","marker":"[19]"},{"why":"An EVRP with charging and discharging under time-variant electricity prices, one of the few bidirectional, price-aware models preceding this work.","marker":"[20]"},{"why":"Introduces large neighbourhood search, the destroy-and-repair paradigm on which the paper's LNS algorithm is built.","marker":"[55]"},{"why":"Provides the adaptive LNS framework, its roulette-wheel strategy weighting, and the relatedness measure reused for Shaw removal.","marker":"[57]"},{"why":"Supplies the real-world ride-sharing trip dataset from which order candidates are drawn in the experiments.","marker":"[59]"},{"why":"Provides the regulated time-of-use tariff that sets charging and discharging prices at the home source and destination locations.","marker":"[62]"}],"fun_headline_variants":["Bidirectional charging doubles EV driver profits","V2G arbitrage adds 20% and doubles EV ride profits","Route and battery optimization with V2G doubles driver profit","Large-neighborhood search doubles EV ride profits with V2G","EVs sell power to grid, drivers double profit"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The 'near-optimal' measurement is anchored to the MIP treated as the exact optimum, yet the model's time-slot length $\\delta$ and horizon size $|T|$ are never given values, and an efficiency parameter $\\lambda$ appears in the battery constraints without being defined; if that exact model is not a faithful or correctly instantiated version of the problem, the near-optimality of the heuristics is not anchored to a true optimum.","fun_headline_variants_meta":{"raw":{"variants":["Bidirectional charging doubles EV driver profits","V2G arbitrage adds 20% and doubles EV ride profits","Route and battery optimization with V2G doubles driver profit","Large-neighborhood search doubles EV ride profits with V2G","EVs sell power to grid, drivers double profit"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000213,"raw_usage":{"total_tokens":1442,"prompt_tokens":987,"completion_tokens":455,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":603,"completion_tokens_details":{"reasoning_tokens":375}},"tokens_in":603,"tokens_out":455,"duration_ms":5366,"temperature":1.0,"reasoning_tokens":375,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T22:49:21.501259+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the 30-order comparisons with $\\delta$ and $|T|$ fixed to explicit values, such as 15-minute slots over a 24-hour horizon, and with $\\lambda$ set to 1, then check whether the LNS and EA profits still sit within the reported margin of the MIP optimum; alternatively, benchmark the two heuristics against a stronger non-greedy planner with lookahead rather than the paper's greedy baseline, and see whether the profit advantage survives.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Surveys EVRP, arc routing, and team orienteering problems in sustainable transportation, supplying the problem family that EVOP-V2G extends."},{"cited_title":"Abdulaal, M","cited_arxiv_id":null,"evidence_quote":"Presents an EVRP incorporating V2G and G2V options with time-variant prices, the closest prior setting that the paper generalises to order selection with time windows."},{"cited_title":"A Reinforcement Learning Approach for Electric Vehicle Routing Problem with Vehicle-to-Grid Supply","cited_arxiv_id":"2204.05545","evidence_quote":"A reinforcement-learning EVRP with V2G supply that considers discharging only, cited to show bidirectional charging in routing is essentially unexplored for orienteering."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"An EVRP with charging and discharging under time-variant electricity prices, one of the few bidirectional, price-aware models preceding this work."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces large neighbourhood search, the destroy-and-repair paradigm on which the paper's LNS algorithm is built."},{"cited_title":"Ropke, D","cited_arxiv_id":null,"evidence_quote":"Provides the adaptive LNS framework, its roulette-wheel strategy weighting, and the relatedness measure reused for Shaw removal."},{"cited_title":"Najmi, D","cited_arxiv_id":null,"evidence_quote":"Supplies the real-world ride-sharing trip dataset from which order candidates are drawn in the experiments."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the regulated time-of-use tariff that sets charging and discharging prices at the home source and destination locations."}],"review_version":1}