{"id":"ddd9d6b6-39a0-46d0-b583-66cc0008a598","arxiv_id":"2507.00628","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A single-stage framework coupling battery dispatch with internal power splitting shows LP maximizes savings and SOC balance while RL improves efficiency and thermal balance.","lead":"This paper tests two ways to control a multi-string battery system: a linear-programming optimizer that plans ahead with forecasts, and a reinforcement-learning agent trained partly by imitating that optimizer. On simulated industrial data, the planning method saved 33% more money while the learning method ran 10% more efficiently and balanced battery temperatures better.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The LP/RL trade-off is defined entirely by unreported scalarization weights (x,y,z) in Eqs. (3) and (18); without these values and a sensitivity analysis, the reported 33% savings versus 10% efficiency contrast is not yet reproducible.","rationale":"The reader's weakest_assumption identifies exactly the same scalarization issue, and I agree that it is the most load-bearing concern. The paper's central claim is not just that an integrated framework exists, but that the LP/RL comparison exhibits specific quantitative trade-offs: 33% greater savings for LP, 10% higher efficiency for RL, lower SOC imbalance for LP, and better temperature balance for RL. All of these numbers are produced by optimizing or rewarding the same weighted sum with unreported coefficients. Since the weights are not given, a reader cannot tell whether the reported trade-offs are robust properties of the two control paradigms or consequences of one particular Pareto choice. For example, a cost-only LP would likely maximize savings and ignore SOC/temperature balance, while a heavily balance-penalized LP would sacrifice savings; the reported '33% greater cumulative savings' has no meaning without knowing where the scalarization sits in that spectrum. The secondary concerns raised by the reader—best-of-ten seed selection, behavior-cloning from LP expert trajectories, and the missing code link—are real but less central: seed selection affects variance estimates, behavior cloning affects the 'model-free RL' label but not the existence of the trade-off, and the missing code is a reproducibility issue. The scalarization weights, by contrast, directly determine the numbers that constitute the paper's headline. My recommendation remains CONDITIONAL, matching the reader's verdict: the concern does not by itself falsify the framework, but it must be resolved before the quantitative trade-offs can be accepted at face value.","tokens_in":7809,"tokens_out":2575,"duration_ms":34861,"concrete_test":"Obtain the reported x, y, z values and rerun both Scenario 1 and Scenario 2 with a small weight sweep, e.g., set (y, z) to 0×, 0.5×, 1×, 2×, and 5× their reported normalized values while keeping x fixed, and also run a cost-only LP (y = z = 0). If the qualitative ordering LPp > RL on savings and RL > LPp on efficiency does not persist across this plausible range, the headline trade-off is an artifact of the unreported scalarization. As a minimal reproducibility check, the authors should also report the tax ratio k in Eq. (1) and the exact cost reduction term used in Eq. (18).","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claims—LP yields 33% greater cumulative savings, RL yields 10% higher efficiency, and the two controllers trade off SOC versus temperature balance—are all generated under a single scalarized objective whose weights are never reported. Equation (3) minimizes sum_t (x·cost_t + y·Δsoc_t + z·Δτ_t), and the RL reward in Eq. (18) is the same weighted sum. The text only says that x, y, z are 'selected such that each individual objective is normalized,' but the actual values and the normalization procedure are omitted. This matters because scalarization weights are not innocuous implementation details: they define the Pareto trade-off between cost savings and balancing objectives. If y and z are large relative to x, the LP will sacrifice savings to homogenize SOC and temperature; if they are small, the LP will behave nearly as a pure cost optimizer. The reported 33% savings advantage of LP is therefore conditional on an unreported choice. The same weights also enter the RL reward, so the comparison between controllers is not invariant to the scalarization. A different acceptable weight choice could shrink, invert, or exaggerate the headline gaps. This is a load-bearing concern for the paper's main empirical claim, not merely a missing hyperparameter, because the paper's conclusion is explicitly framed as a trade-off between savings, SOC balance, temperature balance, and efficiency under this unified objective.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a unified framework for the simultaneous optimization of external battery dispatch (price-aware charge/discharge scheduling) and internal power split among heterogeneous battery strings, with secondary objectives of SOC and temperature balancing. Two controller families are compared: a rolling-horizon linear program (LP) with perfect or persistent forecasts, and a proximal policy optimization (PPO) reinforcement-learning controller that is first pre-trained by behavior cloning from LP expert trajectories. Using the EMSx dataset and a detailed electro-thermal battery simulation, the authors run a 365-day long-term scenario and a 7-day short-term scenario with asymmetric initial conditions. The headline findings are that LP yields 33% greater cumulative savings and better SOC balance, while RL achieves 10% higher system efficiency and better temperature balance, with RL showing more adaptability under forecast uncertainty.","tokens_in":8117,"tokens_out":4309,"duration_ms":48717,"significance":"The paper addresses a real and relevant gap: most BESS EMS studies treat the system as a single homogeneous unit, whereas commercial multi-string systems require joint handling of economic dispatch and internal heterogeneity. The contribution of an integrated, open simulation framework with electro-thermal models is valuable, and comparing model-based and learning-based controllers in the same environment is timely. If the claims are confirmed, the paper would provide a useful benchmark for EMS research. However, the central trade-off results are conditioned on scalarization weights that are never reported, and the RL method is not independent of the LP due to behavior cloning. These issues are fixable with additional experiments and transparency, but as written they limit the reproducibility and robustness of the conclusions.","major_comments":[{"comment":"The scalarization weights x, y, z in the objective function are never reported. Equation (3) defines the LP objective as a weighted sum of cost, SOC imbalance, and temperature imbalance, and Equation (18) uses the same weighted sum as the RL reward. The text only says the coefficients are 'selected such that each individual objective is normalized,' but neither the values nor the normalization procedure are given. Because these weights define the Pareto trade-off between economic savings and balancing objectives, the headline results (33% greater savings for LP, 10% higher efficiency for RL, and the SOC versus temperature trade-off) are conditional on an unreported choice. A sensitivity analysis over (x,y,z) is required to establish that the reported trade-offs are robust; without it, the central empirical claim is not reproducible.","section":"Eq. (3) and Eq. (18)"},{"comment":"The RL controller is not model-free in the usual sense: it is pre-trained by behavior cloning from LP expert trajectories generated under perfect forecasts. The abstract and introduction repeatedly call the RL approach 'model-free,' but the policy is initialized from and bounded by the LP's solution structure. The comparison is therefore between LP and an LP-initialized learner, not between independent model-free and model-based methods. This conflation affects the interpretation of the 10% efficiency advantage and the adaptability claims in Scenario 2. The authors should either drop the 'model-free' characterization or include a fully model-free baseline for comparison.","section":"Section II-B"},{"comment":"In Scenario 2, only RL* is evaluated, where RL* is the instance selected from 10 training runs based on savings. This is a selection-bias problem: reporting the best seed on the evaluation metric inflates the RL results and makes the comparison with the deterministic LP controllers unfair. The paper should report the distribution over the 10 seeds (e.g., mean and interquartile range) for Scenario 2, or justify why a single selected instance is representative. Similarly, the headline long-term numbers in Fig. 2 are point estimates without confidence intervals; the authors should state the variability across seeds.","section":"Section IV-B"},{"comment":"The system efficiency metric η is defined as 1 − (Σ losses)/(Σ |pB|·Δt). This ratio depends on total absolute battery throughput: a controller that uses the battery less will tend to show higher efficiency even if its absolute losses are similar or lower. The metric is also undefined if total throughput is zero. The short-term results in Fig. 4 show that RL* incurs higher absolute losses than the LP solutions, yet the long-term results report 10% higher efficiency for RL. The authors should clarify how the efficiency metric interacts with throughput and report absolute losses and throughput separately, otherwise the efficiency advantage may be an artifact of the metric definition.","section":"Eq. (19)"}],"minor_comments":[{"comment":"Equation (1) uses a static selling tariff kSell, while Eq. (13) writes kSell_t, implying time dependence; please clarify whether the selling tariff is constant or time-varying.","section":"Eqs. (1) and (13)"},{"comment":"The notation pBESS[m]_t = Σ_m pB[m]_t is confusing; the left-hand side should be the total battery power, not an indexed quantity. Please correct the notation.","section":"Eq. (19)"},{"comment":"The numbers '33% greater savings' and '10% higher efficiency' appear only in the abstract and conclusion. The results section should state these percentages explicitly and identify the reference controller (LPp or LPf) for each comparison.","section":"Abstract and Conclusion"},{"comment":"The axis labels for savings and temperature omit units (presumably € and °C), and the caption 'Saving ( )' appears incomplete. Please include units throughout.","section":"Fig. 2"},{"comment":"The term 'model-free RL' is used despite the behavior-cloning warm start; please qualify this wording to avoid misleading readers (see major comment 2).","section":"Section II-B"},{"comment":"The paper claims an open-source framework but does not provide a repository URL or licensing information. Please add the link or specify where the code will be released.","section":"Section III-A"}],"recommendation":"major_revision","confidential_remarks":"The manuscript fits the journal's scope and the core idea is promising, but the missing scalarization weights, the behavior-cloning dependence of the RL controller, and the best-seed selection in Scenario 2 are load-bearing concerns that require additional experiments and transparent reporting. I recommend major revision rather than rejection, as these issues appear fixable within the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing to know: this paper has a genuine idea—a unified single-stage formulation for dispatch plus power split in heterogeneous BESS. But the headline LP/RL numbers rest on unreported scalarization weights. You can't reproduce the 33% savings vs 10% efficiency contrast without x, y, z in Eq. (3)/(18) and the tax ratio k in Eq. (1), and no values are given. That's a real problem.\n\nWhat's new and good: most prior work handles dispatch and internal balancing hierarchically. The integrated LP here is a modest but real step. The comparison under perfect vs persistent forecasts is clean, the EMSx dataset is public, and the paper is candid that the LP uses a simplified battery model. The authors also show RL quantiles in the long-term run, which is more than many RL papers do.\n\nWhere it's soft: the biggest issue is the missing weights. The objective is explicitly multi-objective and the weights define the Pareto trade-off. Different acceptable choices could shrink or invert the reported gaps. That's load-bearing, not a hyperparameter nitpick. Second, Scenario 2 evaluates only the single RL seed that performed best on savings—that's selection bias in favor of RL. Third, calling RL \"model-free\" is misleading; the policy is pre-trained on LP solutions via behavior cloning, so it's a hybrid. The paper describes this clearly, but the abstract overstates it. Also, no code link or repository is provided despite the open-source claim. Minor: the tax ratio and price scaling are missing.\n\nBottom line: the central idea holds up. These are fixable in revision—report the weights, add a sensitivity analysis, use a fairer seed selection, and link the code. I'd send this to peer review; it's a solid engineering contribution that needs tightening, not a reject. I wouldn't cite it until the numbers are pinned down.","headline":"Useful integrated BESS dispatch/power-split framework, but the headline LP/RL trade-off is uninterpretable without the missing objective weights.","tokens_in":8661,"tokens_out":3070,"would_cite":false,"duration_ms":40147,"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":"The paper claims that jointly optimizing price-aware dispatch and internal power splitting in one EMS yields distinct, complementary strengths for LP and RL controllers.","keywords":["battery energy storage systems","power split control","reinforcement learning","linear programming","state of charge balancing","thermal balancing","price-aware scheduling","heterogeneous multi-string BESS"],"falsifier":"Rerun the 365-day scenario with the same code but with the normalization weights explicitly varied, for example setting the temperature term to zero or making it dominate the cost term; if the LP-vs-RL ranking on savings, efficiency, SOC, and temperature does not stay qualitatively the same across those weight settings, the claimed trade-off depends on an arbitrary scalarization rather than on the controllers themselves.","tokens_in":7582,"feed_emoji":"🔋","tokens_out":5327,"duration_ms":59590,"temperature":0.7,"pith_summary":"Commercial battery storage systems usually split control into a high-level scheduler that reacts to electricity prices and a low-level power splitter that balances strings, leaving money and battery health on the table. The paper tries to close that gap with a single framework that decides, at every 15-minute step, both how much power the whole battery trades with the grid and how that power is divided between two heterogeneous strings, while tracking cost, state-of-charge imbalance, and temperature imbalance. It compares a rolling-horizon linear program (with perfect or persistent forecasts) against a reinforcement-learning controller pre-trained by cloning LP solutions, on a simulated industrial site with one 300 kWh and one 200 kWh string. The central result is a trade-off: LP achieves 33% greater cumulative savings and better SOC balance, while RL achieves roughly 10% higher system efficiency and better temperature balance, with more adaptability to forecast errors. If this holds, EMS designs need not choose between economic and physical objectives, and the choice between model-based and learning-based control can be made by the operating environment.","feed_headline":"Linear programming saves more; reinforcement learning is more efficient","feed_subtitle":"Model-based control wins on cost and SOC balance; reinforcement learning wins on efficiency and temperature.","key_machinery":"The load-bearing object is the weighted-sum objective of Eq. (3): minimize over time $x\\cdot\\text{cost}_t + y\\cdot \\Delta soc_t + z\\cdot \\Delta \\tau_t$, where the first term is electricity cost, the second is the summed absolute deviation of each string's SOC from the mean, and the third is the analogous temperature deviation. The same expression, with cost replaced by cost reduction, serves as the RL reward in Eq. (18), so both controllers optimize the identical scalarized goal. Around that objective, a coupled electro-thermal simulation chain—inverter power loss, SOC–open-circuit-voltage relation, equivalent circuit current, and a lumped thermal model—converts the two power set points into the SOC and temperature values that feed the balancing terms; the LP solves this rolling horizon with forecasts, and the RL agent observes state $s_t = [p^L_t, p^{PV}_t, soc^{B1}_{t-1}, soc^{B2}_{t-1}, \\tau^{B1}_{t-1}, \\tau^{B2}_{t-1}, k^{ToU}_t]$ and outputs $a_t = [p^{B1}_t, p^{B2}_t]$.","core_discovery":"On the paper's own terms, the discovery is that a one-stage, price-aware power split controller is achievable and that the two optimization families are complementary rather than one dominant. In 365-day simulation, the LP benchmark with perfect forecasts produced 33% larger cumulative savings than RL and kept the two strings' SOC closer to the mean, while the RL controller produced about 10% higher system efficiency and kept temperatures more uniform; under a 7-day scenario with imbalanced initial SOC and temperature, LP with perfect foresight exploited a predictable active phase and maximized savings, whereas RL matched it on most metrics and outperformed the persistent-forecast LP under uncertainty. The authors attribute the RL efficiency edge to lower throughput-related losses and the LP savings edge to sharper exploitation of price differences, and they note that even the perfect-foresight LP is not globally optimal because of its limited rolling horizon and simplified battery model.","pith_inferences":["The unreported normalization weights (x, y, z) in Eq. (3) are an unexamined degree of freedom: different weight choices would change both the LP dispatch and the RL reward, potentially inverting the reported trade-offs between savings, SOC balance, and temperature balance.","Because temperature balance improves with RL while savings improve with LP, a practical hybrid might use LP for day-ahead energy trading and RL for intra-day power split, but the paper does not test that combination.","The paper measures savings and efficiency but not battery aging directly; if temperature uniformity is a proxy for longer life, the RL advantage may translate into deferred replacement cost, a link the authors leave implicit.","The claims are established in one simulated industrial site with one price profile; scaling to more strings, different chemistries, or real market prices is a testable extension rather than a demonstrated result."],"forward_implications":["Operating a heterogeneous multi-string battery as one monolithic unit is suboptimal; dispatch and power split can live in a single optimization stage without giving up either economic or physical goals.","For operators with reliable forecasts and stable conditions, LP remains the benchmark for savings and SOC homogeneity; under forecast uncertainty, an RL policy pre-trained on LP solutions can maintain efficiency and thermal balance without a forecast model.","Savings and efficiency are not aligned: the controller that saves more money runs the batteries harder and loses more energy, so EMS design must state which metric is primary.","The behavior-cloned PPO approach makes RL training feasible for BESS control, since the policy starts from LP demonstrations rather than from random exploration."],"supporting_citations":[{"why":"Supplies the behavior-cloning-then-PPO method that produces the RL controller.","marker":"[14]"},{"why":"Supplies the industrial load and PV dataset and the simulation benchmark used in both scenarios.","marker":"[15]"},{"why":"Supplies the thermal modeling and power-split balancing formulation extended here.","marker":"[2]"},{"why":"Supplies the SOC–OCV relation used in the battery simulation.","marker":"[11]"},{"why":"Supplies the Coulomb counting method used for estimating actual SOC.","marker":"[12]"},{"why":"Supplies the persistence forecast baseline used by the LPf controller.","marker":"[13]"}],"fun_headline_variants":["LP saves 33% more, RL gains 10% efficiency in BESS","BESS split control: LP for savings, RL for efficiency","Price-aware BESS: LP for cost, RL for efficiency","Trade-off: LP cuts costs, RL balances temps in BESS"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the unstated normalization weights x, y, z in the objective and reward define the right trade-off among cost, SOC balance, and temperature balance; if different weights were chosen, the LP dispatch and the RL reward would both change, and the reported 33%-savings versus 10%-efficiency comparison could flip.","fun_headline_variants_meta":{"raw":{"variants":["LP saves 33% more, RL gains 10% efficiency in BESS","BESS split control: LP for savings, RL for efficiency","Price-aware BESS: LP for cost, RL for efficiency","Trade-off: LP cuts costs, RL balances temps in BESS"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001295,"raw_usage":{"total_tokens":5278,"prompt_tokens":927,"completion_tokens":4351,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":543,"completion_tokens_details":{"reasoning_tokens":4275}},"tokens_in":543,"tokens_out":4351,"duration_ms":30880,"temperature":1.0,"reasoning_tokens":4275,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T21:10:34.093824+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Rerun the 365-day scenario with the same code but with the normalization weights explicitly varied, for example setting the temperature term to zero or making it dominate the cost term; if the LP-vs-RL ranking on savings, efficiency, SOC, and temperature does not stay qualitatively the same across those weight settings, the claimed trade-off depends on an arbitrary scalarization rather than on the controllers themselves.","supporting_citations":[{"cited_title":"Boosting the performance of deep reinforcement learning for energy management systems using behavior cloning from linear programming solutions,","cited_arxiv_id":null,"evidence_quote":"Supplies the behavior-cloning-then-PPO method that produces the RL controller."},{"cited_title":"Emsx: a numerical benchmark for energy management systems,","cited_arxiv_id":null,"evidence_quote":"Supplies the industrial load and PV dataset and the simulation benchmark used in both scenarios."},{"cited_title":"Optimal power split control for state of charge balancing in battery systems with integrated spatial thermal analysis and aging estima- tion,","cited_arxiv_id":null,"evidence_quote":"Supplies the thermal modeling and power-split balancing formulation extended here."},{"cited_title":"Improved ocv model of a li-ion nmc battery for online soc estimation using the extended kalman filter,","cited_arxiv_id":null,"evidence_quote":"Supplies the SOC–OCV relation used in the battery simulation."},{"cited_title":"Methods for state-of-charge de- termination and their applications,","cited_arxiv_id":null,"evidence_quote":"Supplies the Coulomb counting method used for estimating actual SOC."},{"cited_title":"Analysis of the maximal possible grid relief from pv-peak-power impacts by using storage systems for increased self-consumption,","cited_arxiv_id":null,"evidence_quote":"Supplies the persistence forecast baseline used by the LPf controller."}],"review_version":1}