Pith. sign in

REVIEW 3 major objections 4 minor 18 references

Bi-level Model Predictive Control for Energy-aware Integrated Product Pricing and Production Scheduling

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A bi-level model predictive control framework that sets daily prices and hourly production schedules together can cut grid electricity costs by 49.4% and raise profit by 3.7% for a make-to-stock manufacturer with onsite solar.

desk verdict A plausible integration of pricing and scheduling MPC, but the core approximate-gradient algorithm has a sign inconsistency and no validation, so the headline gains are unsupported. read the letter →

arxiv 2507.14385 v1 pith:5OY5GHAK submitted 2025-07-18 eess.SY cs.SY

Hongliang Li , Herschel C. Pangborn , Ilya Kovalenko This is my paper · ORCID
classification eess.SYcs.SY
keywords bi-levelmodelpredictivecontrolproductionschedulingdynamicpricingrenewableenergyintegrationpriceelasticitymake-to-stockmanufacturingmixed-integerquadraticprogramenergy-aware
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper argues that product pricing and hourly production scheduling should be optimized together, not separately, when a manufacturer has onsite solar generation. It builds a bi-level model predictive control framework: a slower loop chooses the daily price to shape demand through price elasticity, and a faster loop schedules machines hour by hour against real-time electricity prices and solar availability. The higher level rewards renewable usage, so the controller lowers prices during sunny hours, pulling demand toward cheap, clean energy. In a lithium-ion battery pack case study the framework cuts grid energy costs by 49.4% and increases profit by 3.7% compared with a grid-only baseline.

What carries the argument

The load-bearing object is Algorithm 1, the component-wise approximate gradient method for bi-level MPC, combined with the affine price elasticity $\gamma = a - bp$ that couples the two levels. The algorithm replaces the intractable hypergradient $\nabla_p F = \partial F/\partial p + (\partial F/\partial u)(\partial u/\partial p) + (\partial F/\partial d)(\partial d/\partial p)$ by $\kappa_u G_u(p_k) + \kappa_d G_d(p_k)$, where $G_u$ measures renewable utilization and $G_d$ is the derivative of the revenue curve. The lower-level MIQP supplies the production schedule and energy mix for each candidate price, and the time-varying slack constraint lets the controller defer production early and converge to the target as the deadline approaches.

What would settle it

Solve the same five-day case study by an exact or exhaustive method, for example by evaluating the bi-level objective on a fine grid of admissible daily prices or by using a global MIQP solver with KKT conditions for the lower level, and compare with Algorithm 1's price path. If the exact optimum does not reproduce the 49.4% grid-cost saving and 3.7% profit gain, or if Algorithm 1's result depends on the initial price, the central claim fails.

Watch

Extended reading notes

Core claim

The central claim is that a tractable component-wise approximate gradient method can solve the bi-level pricing-and-scheduling problem, and that the resulting policy — lower prices when solar is abundant — reconciles sustainability with profitability. On the paper's own terms, the discovery is that the bi-level MPC structure with an MIQP lower level need not be solved by KKT reformulation or exact hypergradients; an approximated gradient combining a revenue component $a - 2bp_k$ and a renewable-utilization component $(1 - \iota^*/\iota(p_k))(p_{\max}-p_{\min})$ is enough to drive the daily price updates. In the five-day case study this produces a 9.4% average price reduction, 11.5% more production, 49.4% lower grid energy cost, 25.9% lower startup cost, and 3.7% higher profit while all demand targets are met.

Load-bearing premise

The entire result depends on the assumption that the simplified price-update rule used in Algorithm 1 points toward the true best price; the paper does not prove this, so the reported savings could be an artifact of the update rule if that assumption fails.

Editorial extensions

If this is right

  • A make-to-stock manufacturer with onsite solar can use dynamic pricing to shift demand into high-solar hours without sacrificing profit, because the lower price is more than offset by reduced grid costs.
  • The bi-level scheme is computable in practice: the MIQP lower level is solved with an off-the-shelf optimizer and the price update uses only two scalar gradient components, so no costly KKT reformulation is needed.
  • Grid energy costs fall roughly by half (49.4%) and startup costs by a quarter (25.9%) in the studied system, at the cost of a modest 5.6% rise in buffer holding costs.
  • Lowering the average price by 9.4% stimulates 11.5% more production, and all daily demand targets remain satisfied, so the policy does not trade away service.
  • The paper notes that comparable solar installations have approximately 24% return on investment and may qualify for U.S. tax credits, so the long-term economic case is stronger than the operational-only gain.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The 49.4% grid-cost saving is only as reliable as the approximate gradient: were the update direction systematically biased, the price path would be an artifact rather than the optimum, which is why the result should be stress-tested against a price grid search.
  • The same two-timescale bi-level template could be applied beyond make-to-stock settings, for example to demand response with storage, where the lower level would schedule charging instead of production and the price gradient would target load shifting.
  • The rise in holding costs suggests inventory is the buffer that absorbs solar variability; optimizing buffer capacity alongside price could improve the trade-off further.
  • Because the case study omits solar capital costs, the reported profit gain is operational only; including levelized cost of energy would move the break-even point and may change the optimal price level.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper proposes a bi-level model predictive control framework for a make-to-stock manufacturing system with onsite solar generation. At the higher level, a daily product price is chosen to maximize revenue and renewable-energy usage; at the lower level, an hourly MIQP scheduling problem is solved to meet the resulting demand at minimum grid-energy, startup, holding, and slack cost. The higher-level price is updated by a component-wise approximate gradient rule (Algorithm 1, Eq. (26)). In a five-day simulation of a lithium-ion battery pack production line, the authors report a 49.4% reduction in grid energy cost and a 3.7% increase in profit relative to a grid-only baseline with constant pricing. The central claim is that the proposed bi-level scheme jointly and profitably coordinates dynamic pricing with energy-aware scheduling.

Significance. If the heuristic gradient were proven to provide a reliable search direction for the nonconvex bi-level problem, the framework would be a useful step toward integrating strategic pricing with operational production control. The paper has clear strengths: the lower-level MIQP is carefully specified with realistic on/off, minimum-run, buffer, and production-tightening constraints; the case study is detailed and uses publicly available price and solar data; and the authors honestly label the gradient approximation as a heuristic. However, the load-bearing step—the component-wise approximate gradient—is not validated against any optimum, and the baseline comparison is not fully specified. The empirical claims therefore are not yet attributable to the proposed joint optimization in the way the paper states.

major comments (3)
  1. [Section 4.2, Eqs. (22)-(26), Algorithm 1] The approximate gradient is the load-bearing component of the paper, but no evidence is given that it points toward a stationary point of the upper-level objective (18). G_u in Eq. (24) is a heuristic function of the renewable fraction, not the derivative of -E_r with respect to price, and G_d in Eq. (25) is the derivative of p(a-bp), using potential demand rather than the realized lower-level outflow d*(p) that appears in the objective. The algorithm provides no convergence proof, no residual or stationarity check, and no comparison against an exact solution. I recommend adding a numerical validation on small instances (for example, comparing Algorithm 1 with a dense price grid or an exact reformulation) or a formal convergence analysis for a simplified version of the problem.
  2. [Section 5.2, Table 2 and Fig. 3] The baseline scenario is not an apples-to-apples comparison with the solar-integrated scenario. The baseline keeps price constant at 82.50 USD while the solar case uses dynamic pricing, but the paper does not explain how the baseline price 82.50 was chosen or whether it was optimized under the grid-only scenario. As a result, the reported 49.4% grid-cost reduction and 3.7% profit increase conflate the effect of solar integration with the effect of switching from a constant to a dynamic pricing policy. The paper should specify the baseline pricing policy and, ideally, decompose the outcomes into (i) the effect of dynamic pricing alone and (ii) the effect of solar integration.
  3. [Section 5, Tables 1-2] The quantitative claims rest on a single five-day simulation with no sensitivity analysis or uncertainty quantification. Several free parameters in Table 1—particularly the gradient weights kappa_u=0.4 and kappa_d=0.6, the desired renewable fraction iota*=0.5, and the L-MPC weights—are likely to influence the resulting price path and the numbers in Table 2. A sensitivity study over these parameters, or at least a small set of alternative settings, is needed to support the claim that the 49.4% and 3.7% figures are robust properties of the framework rather than artifacts of a particular parameter choice.
minor comments (4)
  1. [Algorithm 1, line 6] Line 6 of Algorithm 1 reads pk+1 = projection[pk - kappa_u G_d(pk) - kappa_d G_d(pk)], with G_d appearing twice; one of these should be G_u. The surrounding text and Eq. (26) use G_u and G_d, so this is a typo, but it makes the algorithm as printed ambiguous.
  2. [Section 4.1, Eq. (18)] The upper-level objective uses p(k) and d*(k) as time-indexed quantities, while the text states that the higher level determines a daily price. The notation should be clarified to distinguish the daily price decision from hourly time indices, and to define how the daily price enters the H-step summation.
  3. [Section 5.2, Table 2] The table entry "1,5937" appears to be a typo for "15,937" (and similarly "1,6530" for "16,530"). The formatting should be corrected.
  4. [Section 4.2, Eq. (24)] G_u does not explicitly depend on p; for a fixed renewable fraction it is a constant sign, so calling it a gradient direction is a loose use of the term. Clarifying that it is a heuristic surrogate would help readers calibrate the strength of the claim.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the price/renewable coupling is an explicit design heuristic, and the quantitative profit and grid-cost results are not imported from the inputs; the main weaknesses are correctness/validation, not circularity.

full rationale

The paper's upper and lower objectives (Eqs. 17-18) directly penalize grid energy and reward renewable use, so the case-study reduction in grid cost and rise in renewable fraction are expected consequences of the objective, not circular predictions. Eq. (24) is a sign-changing feedback toward the user-set target iota*; the paper explicitly calls the algorithm a "heuristic solution approach" with "approximated gradient information," and the low-price-during-high-solar behavior is presented as a designed capability ("allowing the system to increase demand by lowering the product price during high renewable energy generation"), not as an empirical prediction derived from an external theory. The headline quantitative figures (49.4% grid-cost savings, 3.7% profit gain) are produced by simulating the L-MPC MIQP with Gurobi and are not fitted parameters or renamed inputs. The reported renewable fraction 51.5% is close to the chosen iota*=0.5, but the paper makes no predictive claim about that fraction; it is a tracking output. Self-citations in the introduction are contextual and not load-bearing. The substantive concern, that G_u in Eq. (24) is not the derivative of the upper objective and that Eq. (26)'s sign is inconsistent with the text, is a correctness/optimality issue rather than a circularity, so it does not raise the circularity score.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

The central results are sensitive to the assumed demand curve (a, b), the hand-chosen gradient weights (kappa_u, kappa_d), the renewable target (iota*), and the heuristic slack profile. The paper does not fit these to data or test their sensitivity. No new physical entities are introduced; the algorithmic components are mathematical constructs rather than invented particles or forces.

free parameters (7)
  • a (base demand potential) = 120 units
    Assumed coefficient in the affine price elasticity model (Eq. 7); no empirical estimation. Directly sets demand scale and the revenue gradient (Eq. 25).
  • b (price sensitivity) = 0.8 unit/USD
    Assumed coefficient in Eq. (7); controls how much demand changes with price and determines the revenue-optimal price a/(2b).
  • kappa_u (renewable gradient weight) = 0.4
    Hand-chosen weight in the combined gradient (Eq. 23 and Algorithm 1); determines how strongly price responds to renewable utilization.
  • kappa_d (revenue gradient weight) = 0.6
    Hand-chosen weight in Eq. (23); balances revenue-seeking against renewable-seeking price changes.
  • iota* (desired renewable fraction) = 0.5
    Target renewable fraction in Eq. (24); by construction makes the gradient push prices down when renewable fraction exceeds 50 percent.
  • tau and xi (production tolerance) = tau=0.05, xi=0.5
    Parameters of the heuristic slack profile in Eq. (15); shape how production targets are tightened over the horizon.
  • L-MPC objective weights (omega_e, omega_r, omega_s, Qg, Qt) = omega_e=10, omega_r=5, omega_s=5000; Qg and Qt as in Sec. 5.1
    Designer-selected weights in the lower-level objective (17); directly set the trade-off between holding, energy, startup, renewable, and slack costs.
assumptions (5)
  • domain assumption Demand is an affine function of price: gamma = a - b p (Eq. 7)
    Invoked in Section 3.3 as the model of market response; no data or validation is provided, and the pricing layer depends entirely on this functional form.
  • domain assumption No raw material constraints, energy from grid and solar only, solar marginal cost zero (Section 2)
    Stated as simplifying assumptions in the problem statement; they isolate energy-driven scheduling but limit generality of the claimed benefits.
  • ad hoc to paper The component-wise approximate gradient (Eq. 23) is a valid search direction for the bi-level problem
    Algorithm 1 relies on this without proof of convergence or optimality; it replaces the exact hypergradient of Eq. (22).
  • ad hoc to paper The slack heuristic alpha(k) = tau(1 - eta(k)(1 - xi)) (Eq. 15) is a reasonable way to enforce production targets
    Introduced in Section 3.5 with a qualitative rationale but no optimality or scheduling-theoretic justification.
  • domain assumption Gurobi solves the lower-level MIQP to (near) optimality at each step
    The solution approach depends on a commercial solver for the MIQP in (17); the paper does not report optimality gaps or solve times.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Bi-level Model Predictive Control for Energy-aware Integrated Product Pricing and Production Scheduling." pith.science (2026). https://pith.science/paper/5OY5GHAK

@misc{pith2026250714385,
  author       = {Pith},
  title        = {Pith review of: Bi-level Model Predictive Control for Energy-aware Integrated Product Pricing and Production Scheduling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5OY5GHAK}},
  note         = {Machine review of arXiv:2507.14385}
}
read the original abstract

The manufacturing industry is under growing pressure to enhance sustainability while preserving economic competitiveness. As a result, manufacturers have been trying to determine how to integrate onsite renewable energy and real-time electricity pricing into manufacturing schedules without compromising profitability. To address this challenge, we propose a bi-level model predictive control framework that jointly optimizes product prices and production scheduling with explicit consideration of renewable energy availability. The higher level determines the product price to maximize revenue and renewable energy usage. The lower level controls production scheduling in runtime to minimize operational costs and respond to the product demand. Price elasticity is incorporated to model market response, allowing the system to increase demand by lowering the product price during high renewable energy generation. Results from a lithium-ion battery pack manufacturing system case study demonstrate that our approach enables manufacturers to reduce grid energy costs while increasing profit.

Figures

Figures reproduced from arXiv: 2507.14385 by the authors.

Figure 1
Figure 1. Bi-level MPC framework for integrated product pricing and production scheduling with onsite solar energy in a [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Lithium-ion battery pack manufacturing system [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Daily price, demand, and cumulative hourly prod [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Machine operation schedules (machine on-off and [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Grid and solar energy consumption over a 5-day [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

18 extracted references · 16 canonical work pages

  1. [1]

    Chen, J., Dong, M., Rong, Y., and Yang, L. (2018). Dynamic pricing for deteriorating products with menu cost. Omega, 75, 13--26

    work page 2018

  2. [2]

    Chen, T., Sun, Y., and Yin, W. (2021). Closing the gap: tighter analysis of alternating stochastic gradient methods for bilevel problems. In Proceedings of the 35th International Conference on Neural Information Processing Systems, NIPS '21

    work page 2021

  3. [3]

    Fibich, G., Gavious, A., and Lowengart, O. (2005). The dynamics of price elasticity of demand in the presence of reference price effects. Journal of the Academy of Marketing Science, 33(1), 66--78

    work page 2005

  4. [4]

    and Pecht, M

    Formica, T. and Pecht, M. (2017). Return on investment analysis and simulation of a 9.12 kilowatt (k W ) solar photovoltaic system. Solar Energy, 144, 629--634

    work page 2017

  5. [5]

    Gurobi Optimizer Reference Manual

    Gurobi Optimization, LLC (2024). Gurobi Optimizer Reference Manual . ://www.gurobi.com

    work page 2024

  6. [6]

    Ha, A.Y. (1997). Optimal dynamic scheduling policy for a make-to-stock production system. Operations Research, 45(1), 42--53

    work page 1997

  7. [7]

    Heimes, H.H., Kampker, A., Lienemann, C., Locke, M., Offermanns, C., Michaelis, S., and Rahimzei, E. (2018). Lithium-ion battery cell production process. PEM Der RWTH Aachen University

    work page 2018

  8. [8]

    Advanced energy project credit

    Internal Revenue Service (2024). Advanced energy project credit. https://www.irs.gov/credits-deductions/businesses/advanced-energy-project-credit

    work page 2024

Show all 18 references
  1. [9]

    Kovalenko, I., Li, H., and Li, Y. (2024). Harnessing controls and robotics for sustainable manufacturing. In Encyclopedia of Sustainable Technologies, volume 1, 663--671. Elsevier

    2024

  2. [10]

    Li, H., Pangborn, H.C., and Kovalenko, I. (2023). A system-level energy-efficient digital twin framework for runtime control of batch manufacturing processes. In 2023 IEEE 19th International Conference on Automation Science and Engineering (CASE), 1--6

    2023

  3. [11]

    Lu, Y., Liu, C., Kevin, I., Wang, K., Huang, H., and Xu, X. (2020). Digital twin-driven smart manufacturing: Connotation, reference model, applications and research issues. Robotics and Computer-Integrated Manufacturing, 61, 101837

    2020

  4. [12]

    Mankiw, N.G. (2021). Principles of economics. Cengage Learning

    2021

  5. [13]

    Open energy information

    National Renewable Energy Laboratory (NREL) (2023). Open energy information. ://openei.org/wiki/Main_Page

    2023

  6. [14]

    and Ames, A.D

    Olkin, Z. and Ames, A.D. (2024). Bilevel optimization for real-time control with application to locomotion gait generation. In 2024 IEEE 63rd Conference on Decision and Control (CDC), 4783--4790

    2024

  7. [15]

    PJM-Your Power, Our Grid

    PJM Interconnection, LLC (2023). PJM-Your Power, Our Grid . ://www.pjm.com/

    2023

  8. [16]

    Van de Berg , D., Shah, N., and Del Rio-Chanona , E.A. (2024). Hierarchical planning-scheduling-control — O ptimality surrogates and derivative-free optimization. Computers and Chemical Engineering, 188, 108726

    2024

  9. [17]

    , " * write output.state after.block = add.period write newline

    ENTRY address author booktitle chapter doi edition editor eid howpublished institution journal key month note number organization pages publisher school series title type url volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence afte...

  10. [18]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.