{"id":"76059d24-b109-442d-901c-f535f0d08386","arxiv_id":"2412.13853","paper_version":3,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Data centers can cut carbon by pairing optimally sized PV and storage, but the gains vary from 50% in Bavaria to negative in Aargau, so location-aware sizing is essential.","lead":"This paper presents an optimization method for sizing solar panels and batteries at an existing data center so the facility can follow a day-ahead electricity plan while lowering carbon emissions and costs. A reader might care because data centers are a fast-growing share of global electricity demand, and the results show the right size depends heavily on local grid carbon intensity.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (4a) contradicts Eq. (5) by a factor W^2 = 576 in the ESS calendar-aging carbon term, so the reported carbon reductions depend on an unverified implementation detail.","rationale":"The reader identifies out-of-sample scenario representativeness as the weakest assumption. That is a legitimate validation gap, and the paper itself acknowledges that forecasting choices significantly affect results. However, the more load-bearing issue is internal: the printed Eq. (4a) is dimensionally inconsistent with Eq. (5), and the text says the two should be part of the same sum. Since the carbon objective directly determines the optimal sizes and the claimed reductions, this inconsistency is a correctness risk, not merely a missing validation. The appropriate response is to keep the CONDITIONAL verdict while requiring the authors to fix the equation, confirm which formula the implementation used, and release the code or a detailed reproducible notebook. This does not change the reader's verdict, but it sharpens the condition under which the paper should be accepted: the numerical claims must be recomputed with a consistent calendar-aging term and verified against that implementation.","tokens_in":20304,"tokens_out":9422,"duration_ms":89767,"concrete_test":"Re-solve the w=0 carbon-only sizing for canton Neuchâtel and Bayern twice: once with the calendar term as printed in Eq. (4a) (E_ess^rated * C_ess^e,LCA / (W * L_ess^calendar)) and once with Eq. (5) (W * E_ess^rated * C_ess^e,LCA / L_ess^calendar), which the text explicitly says Eq. (4a) should sum to. Compare optimal BESS capacity, PV rating, and expected daily carbon emissions. If BESS capacity or the reported 4%/49.6% reductions change by more than a few percent, the headline results are implementation-dependent; the authors should also report which formula the released code uses.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim—up to ~50% carbon-footprint reduction in Germany and 4% in Switzerland—rests on the carbon accounting in Eq. (4a). As printed, Eq. (4a) has the ESS calendar-aging term as E_ess^rated * C_ess^e,LCA / (W * L_ess^calendar), while Eq. (5) and the surrounding text define the same quantity as W * E_ess^rated * C_ess^e,LCA / L_ess^calendar. These differ by a factor W^2, which is 576 for the 24-hour horizon used in the case studies. If the code follows the printed Eq. (4a), battery embodied carbon is undercounted by roughly three orders of magnitude, which would inflate optimal storage sizes and the reported carbon reductions. If the code follows Eq. (5), then Eq. (4a) is a serious typo that must be corrected before the numbers can be audited. Either way, the paper does not currently provide a consistent, checkable statement of the objective that drives the headline results, and no code or data are released to resolve which formula was actually implemented.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a scenario-based stochastic optimization framework for jointly sizing a battery energy storage system (ESS) and a photovoltaic (PV) plant for an existing data center, with the dual objectives of reducing carbon footprint and achieving day-ahead dispatchability. The objective is a weighted sum of expected carbon costs (grid imports, ESS and PV life-cycle emissions) and financial costs (electricity bills, ESS/PV investment and operation). The formulation is a linear/continuous-relaxed optimization problem solved over 84 typical days with 20 scenarios per day, built from 2023 Swiss and German data. Case studies in three Swiss cantons and two German NUTS-1 regions show strong location dependence: optimal ESS/PV sizes vary by up to 36 times across regions, and the maximum expected carbon-footprint reduction is about 49.6% in Bayern, 14.7% in Schleswig-Holstein, 4% in Neuchâtel, 0.3% in Vaud, and -2.6% in Aargau. The paper concludes that dispatchability-oriented sizing must use geographically granular carbon-intensity and irradiance data.","tokens_in":20564,"tokens_out":13027,"duration_ms":115657,"significance":"If the inconsistencies identified below are resolved, the paper makes a useful contribution to data-center planning under carbon constraints. Its strengths are the explicit scenario-based stochastic formulation, the inclusion of both LCA-based embodied carbon and time-varying grid carbon intensity, the clear statement of working hypotheses, and the cross-regional comparison that demonstrates why location-aware sizing matters. The paper also provides useful sensitivity analyses with respect to the cost-carbon weight w and the ESS power-to-energy ratio. However, the reported quantitative reductions are in-sample expected values, no code or data is released, and there is a serious inconsistency in the printed ESS calendar-aging carbon term. These issues currently prevent the headline numerical claims from being independently audited.","major_comments":[{"comment":"The calendar-aging term for ESS carbon emissions is inconsistent by a factor of W^2. As printed, Eq. (4a) gives E_ess^rated * C_ess^{e,LCA} / (W * L_ess^calendar), whereas Eq. (5) defines the same quantity as W * E_ess^rated * C_ess^{e,LCA} / L_ess^calendar. For the W=24 h horizon used in Section 3.4, the two expressions differ by a factor of 576. Because this term enters the objective minimized in Eq. (8), it directly affects the optimal ESS sizes and the carbon-reduction percentages reported in Section 4.2.1. Please correct Eq. (4a) to match Eq. (5) and state explicitly which formula was implemented, ideally by releasing the code or data; without this, the headline numerical results cannot be audited.","section":"Section 2.3.2, Eqs. (4a) and (5)"},{"comment":"The reported reductions (up to 49.6% in Bayern, 14.7% in Schleswig-Holstein, 4% in Neuchâtel, and a 2.6% increase in Aargau) are in-sample expected values computed from 84 typical days generated from 2023 data with 20 scenarios per day. The paper itself states in Section 1.2 that forecasting methods significantly impact the results, yet no out-of-sample year or closed-loop dispatch is used to test whether these reductions materialize. The claims should be rephrased as conditional on the scenario-generation model, and an out-of-sample or sensitivity analysis should be added to support the quantitative conclusions.","section":"Sections 3.3 and 4.2.1"},{"comment":"The load model uses one month (May 2019) of Google cluster trace data, yet Section 3.3 describes seasonal clustering for the load. With a single month of data, seasonal load variability cannot be represented. Please clarify how the 84 typical days were constructed for the load variable, for instance whether the same May profile was replicated across seasons, and if so, discuss the implications for the sizing results, since cooling-related demand seasonality is relevant for data centers.","section":"Sections 3.2.1 and 3.3"},{"comment":"The continuous relaxation of the binary exclusivity variables does not by construction prevent simultaneous import/export at the PCC or simultaneous charge/discharge of the ESS. The paper should justify that the optimal solution of the relaxed problem satisfies exclusivity, for example because with p_inj = p_cons and carbon charged only on imports simultaneous flows are never strictly beneficial, or it should verify a posteriori in the reported case studies that exclusivity holds. This is an implementation-level correctness risk that should be resolved in the text.","section":"Remark 2 and constraints (10)"}],"minor_comments":[{"comment":"The abstract states a reduction of 'approximately 50% in Germany'; this is the regional maximum in Bayern, while another German region, Schleswig-Holstein, shows only a 14.7% reduction. Please clarify that the figure is a regional maximum rather than a country-level result.","section":"Abstract and Section 4.2.1"},{"comment":"The problem is described as convex, but the notation includes binary variables z_ess and z_pcc; please state explicitly that the implemented and solved problem is the continuous relaxation and that the binary variables belong to the original MILP formulation.","section":"Section 2.2 and 2.3.1"},{"comment":"The LCA values are taken from Ecoinvent entries, but the system boundaries (for example, whether recycling or end-of-life stages are included) are not reported; a sentence with the chosen boundaries would improve reproducibility.","section":"Section 3.2.5"},{"comment":"The multi-panel figure is dense; adding panel labels and a short caption describing the left/right arrangement would help the reader map the subplots to the discussion in Sections 4.2.1 and 4.2.2.","section":"Figure 3"},{"comment":"The symbol P_rated_pcc appears twice in the glossary with slightly different descriptions; please unify the nomenclature to avoid confusion.","section":"Nomenclature"}],"recommendation":"major_revision","confidential_remarks":"The main blocker is the inconsistency between Eq. (4a) and Eq. (5) in the calendar-aging carbon term. If the implemented code follows Eq. (5), the issue may be a typographical error, but the authors must confirm this and provide a corrected, unambiguous statement of the objective. The lack of out-of-sample validation is also important because the abstract and conclusions make specific quantitative claims. The paper fits the journal's scope, and I would be willing to review a revised version."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, here is my read. The genuinely new piece is combining stochastic scenario optimization, dynamic grid carbon intensity, LCA-based embodied emissions, and a day-ahead dispatchability constraint into a single PV+ESS sizing formulation. That is a useful integration. The authors are also clear about their working hypotheses, and the case-study design across Swiss cantons and German regions makes the location-dependence point concretely. The core optimization problem is convex, and the constraint logic—Big-M relaxations, fixed power-to-energy ratio for convexity—is explained well enough that a competent reader could reimplement the structure. Credit where due: the paper does not oversell the results; the conclusion notes that dispatchability raises costs in all regions and can increase emissions in low-carbon grids like Aargau.\n\nThe soft spots. First, the stress-test note is real. Eq. (4a) as printed has the ESS calendar-aging carbon term as E_ess^rated * C_ess^e,LCA / (W * L_ess^calendar), while Eq. (5) and the surrounding text define the same quantity as W * E_ess^rated * C_ess^e,LCA / L_ess^calendar. That is a factor W^2, or 576 for a 24-hour horizon. If the code follows Eq. (4a), battery embodied carbon is undercounted by roughly three orders of magnitude, which would inflate storage sizes and carbon-reduction claims. If the code follows Eq. (5), then Eq. (4a) is a serious typo. Either way, the printed objective is not a consistent statement of what was optimized, and no code or data are released to disambiguate. This is not a minor nit; it sits on the main quantitative claims—up to roughly 50% reduction in Bavaria, 4% in Switzerland. The authors need to correct the equation and ideally release an executable implementation.\n\nSecond, the scenario-validation gap. The 84 typical days and 20 scenarios per day come from a single year, 2023, and the reported reductions are in-sample values of the objective. The paper says users must apply the same forecasting methods in operation, but it does not test out-of-sample or closed-loop tracking. That is an addressable limitation, and the authors acknowledge related issues, but it means the percentages should be read as illustrative rather than expected outcomes.\n\nThird, the carbon-intensity time series are proprietary (Emissium), which limits reproducibility, though the method itself does not depend on that specific provider.\n\nOverall: the method is coherent and the paper is worth a serious referee, but the W^2 discrepancy needs to be resolved before the headline numbers can be trusted. I would send it to review with a request for correction and better reproducibility, not desk-reject it. This is primarily for power-systems researchers working on DER sizing and data-center sustainability teams.","headline":"A well-structured, location-aware PV+ESS sizing method with an honest limitations section, but a W^2 discrepancy between Eq. (4a) and Eq. (5) in the ESS calendar-aging carbon term means the headline reductions are not auditable as printed.","tokens_in":21055,"tokens_out":2419,"would_cite":false,"duration_ms":22973,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C15","90C25"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that a scenario-based multi-objective optimization can jointly size battery storage and local PV for a data center to minimize expected carbon and cost while keeping the site dispatchable day-ahead.","keywords":["data center","battery energy storage","photovoltaic sizing","dispatchability","carbon footprint","stochastic optimization","multi-objective","scenario-based"],"falsifier":"Take the same data center and rerun the sizing on out-of-sample years (e.g., 2022 or 2024) or on a hold-out period with the same optimization; if the optimal BESS and PV ratings change by more than a few percent, or if the realized carbon reduction in Bavaria falls well short of roughly 50% under the prescribed day-ahead tracking policy, the central claim about location-driven sizing is not robust.","tokens_in":20113,"feed_emoji":"🔋","tokens_out":6274,"duration_ms":54867,"temperature":0.7,"pith_summary":"This paper claims that the right size for a data center's co-located battery energy storage and photovoltaic plant is found by solving a scenario-based multi-objective optimization that minimizes expected carbon emissions and financial costs while forcing the site to track a day-ahead power profile. The framework accounts for life-cycle emissions of storage and PV, dynamically varying grid carbon intensity, day-ahead prices, and stochastic demand, irradiance, and prices through 84 typical days with 20 scenarios each. Case studies across Swiss cantons and German regions show the optimal ratings depend strongly on location: carbon reductions reach roughly 50% in Bavaria, about 4% in Neuchâtel, and are negative in Aargau, where forcing dispatchability increases emissions. A reader would care because data-center electricity demand is growing fast, and this gives operators a concrete, data-driven way to balance carbon goals, cost, and grid-friendliness.","feed_headline":"Carbon-aware storage sizing cuts data-center CO2 up to 50%","feed_subtitle":"A scenario-based optimizer picks battery and PV size by location, with 36x spread in capacity across regions.","key_machinery":"The central object is a scenario-based multi-objective convex optimization (Eq. 8) whose decision variables are the ESS rated energy and power (linked by a fixed power-to-energy ratio $r^{\\mathrm{p2e}}_{\\mathrm{ess}}$), the PV rated power, and the storage power trajectory across all scenarios. The load-bearing device is the dispatchability constraint (Eq. 3q, relaxed via tracking tolerance $\\epsilon_t$ in Eqs. 11a–11c), which forces every scenario of a typical day to share the same day-ahead PCC power plan; together with per-scenario constraints on ESS dynamics, aging (calendar plus cycling via Eq. 6), ratings, and grid connection limits, it makes the expected carbon and cost objective (Eq. 1) solvable as a convex problem with continuous relaxation of the exclusivity variables.","core_discovery":"The central claim is that the sizing problem for a dispatchable data center — how many kilowatt-hours of battery and kilowatts of PV to install — can be cast as a convex stochastic program whose objective is a weighted sum of carbon-equivalent and financial costs. The paper establishes that embedding the day-ahead dispatch constraint (a tolerance on tracking a pre-computed PCC power profile) into the sizing stage jointly determines the asset ratings and the dispatch plan, and that the solutions are location-specific: BESS capacity can vary by up to 36 times between regions, with the largest carbon reductions where grid carbon intensity is high. It also shows that a fixed power-to-energy ratio of the battery, used to keep the formulation convex, is a user-tunable parameter that materially changes the cost and carbon outcome.","pith_inferences":["Relaxing the fixed power-to-energy ratio would let the optimizer choose the ratio endogenously; the paper sweeps it as a parameter, so its reported 'optimal' ratios are a lower bound on what a full relaxation could achieve.","Because imbalance costs are excluded, the economic comparison understates the value of dispatchability; including imbalance penalties would likely increase optimal storage capacity in regions where tracking errors are costly.","The scenario-generation method is itself a design choice; the paper claims better forecasts would favor PV and reduce storage needs, which implies the 36x inter-region spread in capacity may partly reflect scenario-generation noise rather than pure locational economics.","A testable extension is to apply the same framework in a region with very high grid carbon intensity (for example, parts of the United States or India), where the model predicts carbon savings well above 50% and a larger PV share."],"forward_implications":["If the framework is correct, data-center operators can size storage and PV for a chosen tracking accuracy, carbon weight $w$, and location, and obtain a Pareto front between carbon savings and cost.","In high-carbon grid regions like Bavaria, carbon reductions of roughly 50% are possible but come with a near-doubling of operational costs; in low-carbon regions like Aargau, forcing dispatchability can raise emissions, so local renewables do not always reduce footprint.","The optimal power-to-energy ratio $r^{\\mathrm{p2e}}_{\\mathrm{ess}}$ is location- and weight-dependent; sweeping this ratio gives a better design than the default unit ratio and can lower operational costs by a few percent.","Because the sizing stage and operation must share the same forecasting methods, changing the forecaster after the sizing stage invalidates the computed ratings and reductions.","The reported reduction percentages and capacity spreads are direct outputs of the scenario-based sizing; they are not guarantees for any single future year, as they depend on the 2023 historical data used to build scenarios."],"supporting_citations":[{"why":"Supplies the dispatchability concept and storage-based day-ahead tracking approach that the sizing framework builds on.","marker":"[31]"},{"why":"Provides the calendar-plus-cycling aging model used to estimate ESS costs and carbon emissions over its lifetime.","marker":"[34]"},{"why":"An earlier BESS sizing method for data-center microgrids that this work extends with explicit carbon accounting and dispatch constraints.","marker":"[23]"},{"why":"A carbon-aware data-center design framework that ignores economic costs and 24/7 renewable coverage; the paper positions its own contribution against it.","marker":"[29]"},{"why":"Monte Carlo methods used to generate the stochastic scenarios for demand, irradiance, prices, and carbon intensity.","marker":"[50]"},{"why":"K-Means clustering used to build typical days from irradiance observations, a key step in scenario construction.","marker":"[51]"},{"why":"Source of the day-ahead electricity price data used in the Swiss and German case studies.","marker":"[40]"},{"why":"Source of the solar irradiance data for the locations studied, feeding the PV generation model.","marker":"[42]"},{"why":"Ecoinvent database supplies the life-cycle assessment values for battery and PV embodied carbon used in the objective.","marker":"[45]"},{"why":"Motivates granular carbon accounting by showing renewable energy certificates can be misleading, justifying the time- and location-resolved grid carbon intensity approach.","marker":"[12]"}],"fun_headline_variants":["Data-center storage sizing can halve CO2 emissions","Battery and PV sizing: 36x capacity gap across regions","Stochastic optimizer cuts data-center carbon up to 50%","Location-aware storage design reduces data-center CO2","Convex sizing trims data-center cost and emissions"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The optimal sizes and carbon reductions are computed from 84 typical days built from a single year (2023) of historical data; if those scenarios do not represent future weather, prices, demand, and grid carbon intensity, the computed ratings and reductions will not materialize in operation.","fun_headline_variants_meta":{"raw":{"variants":["Data-center storage sizing can halve CO2 emissions","Battery and PV sizing: 36x capacity gap across regions","Stochastic optimizer cuts data-center carbon up to 50%","Location-aware storage design reduces data-center CO2","Convex sizing trims data-center cost and emissions"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000176,"raw_usage":{"total_tokens":1312,"prompt_tokens":989,"completion_tokens":323,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":605,"completion_tokens_details":{"reasoning_tokens":243}},"tokens_in":605,"tokens_out":323,"duration_ms":3638,"temperature":1.0,"reasoning_tokens":243,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T12:43:47.481933+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the same data center and rerun the sizing on out-of-sample years (e.g., 2022 or 2024) or on a hold-out period with the same optimization; if the optimal BESS and PV ratings change by more than a few percent, or if the realized carbon reduction in Bavaria falls well short of roughly 50% under the prescribed day-ahead tracking policy, the central claim about location-driven sizing is not robust.","supporting_citations":[{"cited_title":"Sossan, E","cited_arxiv_id":null,"evidence_quote":"Supplies the dispatchability concept and storage-based day-ahead tracking approach that the sizing framework builds on."},{"cited_title":"Glasserman, Monte Carlo methods in financial engineering, softcover version of original hardcover ed","cited_arxiv_id":null,"evidence_quote":"Monte Carlo methods used to generate the stochastic scenarios for demand, irradiance, prices, and carbon intensity."},{"cited_title":"URL https://newtransparency.entsoe.eu/market/energyPrices","cited_arxiv_id":null,"evidence_quote":"Source of the day-ahead electricity price data used in the Swiss and German case studies."},{"cited_title":"URL https://ads.atmosphere.copernicus.eu/cdsapp#!/dataset/cams-solar-radiation-timeseries? tab=overview","cited_arxiv_id":null,"evidence_quote":"Source of the solar irradiance data for the locations studied, feeding the PV generation model."},{"cited_title":"Bjørn, S","cited_arxiv_id":null,"evidence_quote":"Motivates granular carbon accounting by showing renewable energy certificates can be misleading, justifying the time- and location-resolved grid carbon intensity approach."}],"review_version":1}