pith. sign in

arxiv: 2601.21039 · v2 · submitted 2026-01-28 · 📡 eess.SY · cs.SY

Mean-Field Learning for Storage Aggregation

Jingguan Liu , Cong Chen , Xiaomeng Ai , Jiakun Fang , Jinsong Wang , Jinyu Wen This is my paper

Pith reviewed 2026-05-16 10:05 UTC · model grok-4.3

classification 📡 eess.SY cs.SY
keywords mean-field learningstorage aggregationenergy storageconvex surrogatepower system operationsprice-responsive modelgradient-based learning
0
0 comments X

The pith

Large populations of heterogeneous storage devices converge to a unique convex mean-field limit that acts as a tractable surrogate for their aggregate behavior.

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

The paper establishes that the combined response of many storage units approaches a single convex limit as the population size increases. This limit removes the need to track every individual nonconvex device and instead supports a simple population-level model. The model is identified from price-response data through gradient-based optimization and can be inserted directly into power-system scheduling problems. Convergence also supplies an explicit bound on how much the surrogate deviates from true aggregate performance.

Core claim

As the population of storage devices grows, aggregate performance converges to a unique, convex mean-field limit. This convexity produces a price-responsive characterization of the aggregate and permits bounding the mean-field approximation error. A convex surrogate model with physically interpretable parameters is then constructed to approximate the aggregate behavior of large populations and is identified as an optimization problem solved by a gradient-based algorithm on historical price-response data.

What carries the argument

The mean-field limit of aggregate storage behavior, shown to be unique and convex for large heterogeneous populations and used to build the price-responsive surrogate model.

If this is right

  • Aggregate storage behavior can be represented by a single convex optimization problem suitable for power-system operations.
  • The approximation error between the surrogate and true aggregate response can be bounded explicitly.
  • Surrogate parameters can be learned efficiently from historical price-response data without needing detailed device-level models.
  • The learned model can be embedded directly into existing grid dispatch and market-clearing tools.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same mean-field construction could be applied to other distributed resources such as electric-vehicle fleets whose individual models are also heterogeneous and nonconvex.
  • Real-time updates to the surrogate parameters could be performed incrementally as new price-response observations arrive.
  • Convexity of the aggregate model may simplify the design of incentive mechanisms or contracts offered to storage aggregators.

Load-bearing premise

The storage devices are sufficiently heterogeneous and the population is large enough for the mean-field limit to be unique and convex.

What would settle it

An experiment or simulation in which the aggregate storage response remains visibly nonconvex or fails to converge for arbitrarily large populations of heterogeneous devices.

Figures

Figures reproduced from arXiv: 2601.21039 by Cong Chen, Jiakun Fang, Jingguan Liu, Jinsong Wang, Jinyu Wen, Xiaomeng Ai.

Figure 1
Figure 1. Figure 1: summarizes these objects and outlines how they are constructed in the subsequent sections. Specifically, we define: (i) the individual storage performance P E i and C E i (pi ) in (2)-(3); (ii) the mean-field aggregate performance P M I and C M I (p) in (5)-(6); (iii) the mean-field limit performance P L and C L(p) in (7)-(8); and (iv) the approximate mean-field limit performance P“L and C“L(p) in (13)-(14… view at source ↗
Figure 2
Figure 2. Figure 2: summarizes the proposed learning framework. Guided by Proposition 1, we first choose a convex, market￾compatible surrogate to approximate the mean-field limit, consisting of a convex parameterized cost function C“ and a convex parameterized set P“ for aggregated storage, as shown in green in the left panel of the figure and discussed in Section IV-B. Guided by Proposition 3, we use device￾level historical … view at source ↗
Figure 3
Figure 3. Figure 3: Visualization of mean-field flexibility set. (a) Impact of price number (b) Impact of device number 0 50 100 150 200 250 0 1 2 3 4 5 6 Max distance between optimal points (kW) Number of storage devices 0 2 4 6 8 10 12 14 16 18 20 0 2 4 6 8 10 12 14 Hausdorff distance (kW) Number of price signals Surrogate without (15) Surrogate with (15) [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Impact of price number and device number. with a small overestimated region in the upper-right corner. This region corresponds to relaxed operating points that allow simultaneous charging and discharging, which are inactive for the surrogate’s optimal price response. The overestimation arises because our convex surrogate is trained from historical optimal power profiles, rather than from the full flexibili… view at source ↗
Figure 6
Figure 6. Figure 6: Simulation results in the wholesale market. From [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
read the original abstract

Distributed energy storage devices can be aggregated to provide operational flexibility for power systems. This requires representing a massive device population as a single, tractable surrogate that is computationally efficient and accurate. However, surrogate identification is challenging due to heterogeneity, nonconvexity, and high dimensionality of storage devices. To address these challenges, this paper develops a mean-field learning framework for storage aggregation. We interpret aggregation as the average behavior of a large storage population and show that, as the population grows, aggregate performance converges to a unique, convex mean-field limit, enabling tractable population-level modeling. This convexity further yields a price-responsive characterization of aggregate storage behavior and allows us to bound the mean-field approximation error. We construct a convex surrogate model with physically interpretable parameters that approximates the aggregate behavior of large storage populations and can be embedded directly into power system operations. Surrogate parameter identification is formulated as an optimization problem using historical price-response data, and we adopt a gradient-based algorithm for efficient learning. Case studies validate the theoretical findings and demonstrate the effectiveness of the proposed framework in approximation accuracy and data efficiency.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The manuscript develops a mean-field learning framework for aggregating distributed energy storage devices. It interprets aggregation as the average behavior of a large population and shows that as the population grows, aggregate performance converges to a unique, convex mean-field limit. This enables a price-responsive characterization, an error bound on the mean-field approximation, and construction of a convex surrogate model with physically interpretable parameters. Surrogate parameters are identified via gradient-based optimization on historical price-response data, with case studies validating approximation accuracy and data efficiency.

Significance. If the convergence and convexity results hold, the framework would provide a tractable, convex surrogate for large-scale heterogeneous storage populations that can be directly embedded in power-system optimization, addressing nonconvexity and dimensionality issues in aggregation. The data-driven learning from external historical data and the explicit error bound are strengths that could improve upon existing heuristic aggregation methods.

major comments (2)
  1. [Abstract] Abstract: the claim that aggregate performance converges to a unique, convex mean-field limit as the population grows, enabling an error bound, is stated without an explicit list of assumptions (e.g., Lipschitz continuity of value functions or transition kernels, or a quantitative heterogeneity measure) or the dependence of the bound on N; this is load-bearing because the skeptic correctly notes that uniqueness and convexity can fail for finite N=100–1000 without sufficient heterogeneity.
  2. [Case Studies] Case studies: no quantitative error metrics, finite-N rates, or convexity checks for the learned surrogate are reported, despite the abstract asserting that the convexity yields a usable bound; without these, it is impossible to verify whether the approximation remains accurate or convex for practical population sizes.
minor comments (1)
  1. The abstract could more clearly distinguish the mean-field limit derivation (from population averaging) from the subsequent surrogate identification step that uses external data.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments on our manuscript. We address each major comment below and will revise the paper to strengthen the explicit statement of assumptions and the quantitative validation in the case studies.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the claim that aggregate performance converges to a unique, convex mean-field limit as the population grows, enabling an error bound, is stated without an explicit list of assumptions (e.g., Lipschitz continuity of value functions or transition kernels, or a quantitative heterogeneity measure) or the dependence of the bound on N; this is load-bearing because the skeptic correctly notes that uniqueness and convexity can fail for finite N=100–1000 without sufficient heterogeneity.

    Authors: We agree that the abstract should more explicitly reference the assumptions. The convergence and uniqueness results rely on Lipschitz continuity of the value functions and transition kernels together with a quantitative heterogeneity measure (detailed in Assumption 1 and Theorem 1). The error bound scales as O(1/N) modulated by the heterogeneity parameter. In the revision we will add a concise clause to the abstract listing these assumptions and noting the N-dependence, while keeping the abstract length appropriate. This clarifies that the mean-field convexity is an asymptotic property and that finite-N deviations are controlled by the heterogeneity bound. revision: yes

  2. Referee: [Case Studies] Case studies: no quantitative error metrics, finite-N rates, or convexity checks for the learned surrogate are reported, despite the abstract asserting that the convexity yields a usable bound; without these, it is impossible to verify whether the approximation remains accurate or convex for practical population sizes.

    Authors: We accept that the case-study section would be strengthened by explicit quantitative metrics. The current version emphasizes qualitative agreement and data efficiency; we will add tables reporting mean-absolute and mean-squared errors between finite-N aggregates and the mean-field limit for N ranging from 50 to 5000, empirical convergence rates, and numerical checks (e.g., Hessian eigenvalues) confirming convexity of the learned surrogate. These additions will directly demonstrate accuracy and convexity for practical population sizes and support the usability of the derived error bound. revision: yes

Circularity Check

0 steps flagged

No significant circularity; mean-field limit follows from population averaging

full rationale

The derivation chain begins with the interpretation of aggregation as the average behavior of a large heterogeneous storage population and proceeds to a standard mean-field convergence argument as N grows. This limit is shown to be unique and convex under the stated regularity and heterogeneity assumptions, after which a price-responsive characterization and error bound are obtained directly from the convexity property. Surrogate parameters are then identified by solving an optimization problem against external historical price-response data. None of these steps reduces to a self-definition, a fitted input renamed as a prediction, or a load-bearing self-citation; the central claims remain independent of the learned parameters and rest on population-level averaging rather than tautological closure within the paper's own equations.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The framework rests on standard mean-field assumptions for heterogeneous agents plus the claim that storage dynamics admit a convex aggregate limit; no new physical entities are postulated.

free parameters (1)
  • surrogate parameters
    Learned from historical price-response data via gradient descent; these are the only fitted quantities.
axioms (1)
  • domain assumption Large-population limit yields unique convex mean-field behavior
    Invoked to guarantee tractability and error bounds; appears in the convergence statement.

pith-pipeline@v0.9.0 · 5499 in / 1323 out tokens · 29081 ms · 2026-05-16T10:05:59.607992+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Reachable-Set Decomposition for Real-Time Aggregation of Multi-Zone HVAC Fleets

    eess.SY 2026-05 unverdicted novelty 6.0

    A reachable-set decomposition framework encodes remaining-horizon feasibility into per-period constraints for scalable real-time aggregation of multi-zone HVAC fleets via offline inner approximations and online parall...

Reference graph

Works this paper leans on

37 extracted references · 37 canonical work pages · cited by 1 Pith paper

  1. [1]

    Preference-oriented aggregation of heterogeneous distributed energy resources for reserve dispatch,

    J. Liu, X. Ai, S. Cui, X. Xue, S. Wang, J. Fang, W. Yao, and J. Wen, “Preference-oriented aggregation of heterogeneous distributed energy resources for reserve dispatch,”IEEE Transactions on Smart Grid, 2025

    work page 2025

  2. [2]

    Texas’ only energy provider with home backup power,

    “Texas’ only energy provider with home backup power,” [Online]. Available: https://www.basepowercompany.com/, accessed: Nov. 6, 2025

    work page 2025

  3. [3]

    Bid and self-schedule submission in California ISO markets,

    “Bid and self-schedule submission in California ISO markets,” [Online]. Available: https://www.caiso.com/documents/section-30- bid-and-self-schedule-submission-in-california-iso-markets-as-of- aug-1-2024.pdf, August 2024, accessed: Nov. 6, 2025

    work page 2024

  4. [4]

    FERC order no. 2222: Fact sheet,

    “FERC order no. 2222: Fact sheet,” [Online]. Available: https://www. ferc.gov/media/ferc-order-no-2222-fact-sheet, Sep. 2020, accessed: Nov. 6, 2025

    work page 2020

  5. [5]

    Predicting strategic energy storage behaviors,

    Y . Bian, N. Zheng, Y . Zheng, B. Xu, and Y . Shi, “Predicting strategic energy storage behaviors,”IEEE Transactions on Smart Grid, vol. 15, no. 2, pp. 1608–1619, 2023

    work page 2023

  6. [6]

    Coupling-aware aggregation of multi-zone HV AC loads under uncertainty: A two-level framework,

    J. Liu, H. Jiang, X. Ai, S. Wang, X. Xue, S. Cui, J. Hou, J. Fang, and J. Wen, “Coupling-aware aggregation of multi-zone HV AC loads under uncertainty: A two-level framework,”IEEE Transactions on Smart Grid, 2025

    work page 2025

  7. [7]

    Characterizing the flexibility of electric vehicle fleets considering battery degradation cost,

    Z. Long, N. Zhang, K. Feng, Y . Wang, and J. Wang, “Characterizing the flexibility of electric vehicle fleets considering battery degradation cost,”IEEE Transactions on Industry Applications, 2025

    work page 2025

  8. [8]

    Efficient quantification and representation of aggregate flexibility in electric vehicles,

    N. K. Panda and S. H. Tindemans, “Efficient quantification and representation of aggregate flexibility in electric vehicles,”Electric Power Systems Research, vol. 235, p. 110811, 2024

    work page 2024

  9. [9]

    Exact aggregate models for optimal management of heterogeneous fleets of storage devices,

    D. Angeli, Z. Dong, and G. Strbac, “Exact aggregate models for optimal management of heterogeneous fleets of storage devices,”IEEE Transactions on Control of Network Systems, vol. 11, no. 3, pp. 1430– 1439, 2023

    work page 2023

  10. [10]

    Continuous-time aggregation of massive flexible HV AC loads con- sidering uncertainty for reserve provision in power system dispatch,

    J. Liu, X. Ai, J. Fang, S. Cui, S. Wang, W. Yao, and J. Wen, “Continuous-time aggregation of massive flexible HV AC loads con- sidering uncertainty for reserve provision in power system dispatch,” IEEE Transactions on Smart Grid, vol. 15, no. 5, pp. 4835–4849, 2024

    work page 2024

  11. [11]

    Aggregate power flexibility in unbalanced distribution systems,

    X. Chen, E. Dall’Anese, C. Zhao, and N. Li, “Aggregate power flexibility in unbalanced distribution systems,”IEEE Transactions on Smart Grid, vol. 11, no. 1, pp. 258–269, 2019

    work page 2019

  12. [12]

    Leveraging two-stage adaptive robust optimization for power flexibility aggregation,

    X. Chen and N. Li, “Leveraging two-stage adaptive robust optimization for power flexibility aggregation,”IEEE Transactions on Smart Grid, vol. 12, no. 5, pp. 3954–3965, 2021

    work page 2021

  13. [13]

    Aggregation and disaggregation of energetic flexibility from distributed energy resources,

    F. L. M ¨uller, J. Szab ´o, O. Sundstr ¨om, and J. Lygeros, “Aggregation and disaggregation of energetic flexibility from distributed energy resources,”IEEE Transactions on Smart Grid, vol. 10, no. 2, pp. 1205– 1214, 2017

    work page 2017

  14. [14]

    An efficient method for quanti- fying the aggregate flexibility of plug-in electric vehicle populations,

    F. Al Taha, T. Vincent, and E. Bitar, “An efficient method for quanti- fying the aggregate flexibility of plug-in electric vehicle populations,” IEEE Transactions on Smart Grid, vol. 16, no. 4, pp. 3142–3154, 2024

    work page 2024

  15. [15]

    Alleviating the curse of dimensionality in minkowski sum approximations of storage flexibility,

    E. ¨Ozt¨urk, T. Faulwasser, K. Worthmann, M. Preißinger, and K. Rhein- berger, “Alleviating the curse of dimensionality in minkowski sum approximations of storage flexibility,”IEEE Transactions on Smart Grid, vol. 15, no. 6, pp. 5733–5743, 2024

    work page 2024

  16. [16]

    Model-free aggregation for virtual power plants using input convex neural networks,

    W. Lin, Y . Wang, J. Wu, and F. Feng, “Model-free aggregation for virtual power plants using input convex neural networks,”IEEE Transactions on Smart Grid, vol. 16, no. 3, pp. 2404–2415, 2025

    work page 2025

  17. [17]

    Data-driven modeling of aggregate flexibility under uncertain and non-convex device models,

    S. Taheri, V . Kekatos, S. Veeramachaneni, and B. Zhang, “Data-driven modeling of aggregate flexibility under uncertain and non-convex device models,”IEEE Transactions on Smart Grid, vol. 13, no. 6, pp. 4572–4582, 2022

    work page 2022

  18. [18]

    New insights from the shapley-folkman lemma on dispatchable demand in energy markets,

    K. Hreinsson, A. Scaglione, M. Alizadeh, and Y . Chen, “New insights from the shapley-folkman lemma on dispatchable demand in energy markets,”IEEE Transactions on Power Systems, vol. 36, no. 5, pp. 4028–4041, 2021

    work page 2021

  19. [19]

    arXiv preprint arXiv:1911.01525 , year=

    W. Han and Y . Yang, “Statistical inference in mean-field variational bayes,”arXiv preprint arXiv:1911.01525, 2019

    work page arXiv 1911

  20. [20]

    Learning mean-field games,

    X. Guo, A. Hu, R. Xu, and J. Zhang, “Learning mean-field games,” Advances in neural information processing systems, vol. 32, 2019

    work page 2019

  21. [21]

    Optimal mechanisms for demand response: An indifference set approach,

    M. Mehrabi, O. Karaduman, and S. Wager, “Optimal mechanisms for demand response: An indifference set approach,”arXiv preprint arXiv:2409.07655, 2024

    work page arXiv 2024

  22. [22]

    Perturbed decision-focused learning for modeling strategic energy storage,

    M. Yi, S. Alghumayjan, and B. Xu, “Perturbed decision-focused learning for modeling strategic energy storage,”IEEE Transactions on Smart Grid, 2025

    work page 2025

  23. [23]

    Molchanov,Theory of random sets

    I. Molchanov,Theory of random sets. Springer, 2005

    work page 2005

  24. [24]

    Integrals of set-valued functions,

    R. J. Aumann, “Integrals of set-valued functions,”Journal of mathe- matical analysis and applications, vol. 12, no. 1, pp. 1–12, 1965

    work page 1965

  25. [25]

    A strong law of large numbers for random compact sets,

    Z. Artstein and R. A. Vitale, “A strong law of large numbers for random compact sets,”The Annals of Probability, pp. 879–882, 1975

    work page 1975

  26. [26]

    Quasi-equilibria in markets with non-convex prefer- ences,

    R. M. Starr, “Quasi-equilibria in markets with non-convex prefer- ences,”Econometrica: journal of the Econometric Society, pp. 25–38, 1969

    work page 1969

  27. [27]

    IKEA Germany launches dynamic electricity tariff,

    “IKEA Germany launches dynamic electricity tariff,” [Online]. Avail- able: https://bit.ly/49B3Vxk (accessed Jan. 23, 2026), January 2026

    work page 2026

  28. [28]

    Wholesale market partic- ipation of DERAs: DSO-DERA-ISO coordination,

    C. Chen, S. Bose, T. D. Mount, and L. Tong, “Wholesale market partic- ipation of DERAs: DSO-DERA-ISO coordination,”IEEE Transactions on Power Systems, vol. 39, no. 5, pp. 6605–6614, 2024

    work page 2024

  29. [29]

    Optnet: Differentiable optimization as a layer in neural networks,

    B. Amos and J. Z. Kolter, “Optnet: Differentiable optimization as a layer in neural networks,” inInternational conference on machine learning. PMLR, 2017, pp. 136–145

    work page 2017

  30. [30]

    Differentiable convex optimization layers,

    A. Agrawal, B. Amos, S. Barratt, S. Boyd, S. Diamond, and J. Z. Kolter, “Differentiable convex optimization layers,”Advances in neural information processing systems, vol. 32, 2019

    work page 2019

  31. [31]

    Dataset of electricity prices,

    “Dataset of electricity prices,” [Online]. Available: https://bit.ly/ 4r4w4mg (accessed Sep. 27, 2025), September 2025. 11 TABLE II:Random variables versus random sets. Concept Random variablexRandom setX Object (value space)x: Ω→R n, a measurable map taking values inR n X: Ω→2 Rn , a measurable map taking values in nonempty, compact subsets ofR n (Defini...

    work page 2025

  32. [32]

    In particular, aggregation is captured by Minkowski addition, and conver- gence is measured via the Hausdorff distance

    Side-by-side comparison.Table II summarizes the key correspondences between random variables and random sets that are directly used in our mean-field analysis. In particular, aggregation is captured by Minkowski addition, and conver- gence is measured via the Hausdorff distance. These correspondences enable a random-set analog of the strong law of large n...

    work page

  33. [33]

    Accordingly, the appropriate notion of expectation is the Aumann expectation, defined through (measurable) selections of the random set

    From points to sets.The key distinction is that a random variable realizes apointinR n, whereas a random set realizes acompact setinR n for each outcomeω. Accordingly, the appropriate notion of expectation is the Aumann expectation, defined through (measurable) selections of the random set. Lemma 1 further implies that Minkowski averaging of ran- dom sets...

    work page

  34. [34]

    Letξbe a nonnegative integrable random variable and define the random set X(ω) := [0, ξ(ω)] = x∈R 0≤x≤ξ(ω)

    A toy example.To illustrate the analogy, consider a one-dimensional example. Letξbe a nonnegative integrable random variable and define the random set X(ω) := [0, ξ(ω)] = x∈R 0≤x≤ξ(ω) . A single realization ofXis thus an interval[0, ξ(ω)]. Now considerIi.i.d. random sets{X i}I i=1, whereX i = [0, ξi]and{ξ i}I i=1 are i.i.d. random variables. The mean-fiel...

    work page

  35. [35]

    random sets defined on a non-atomic probability space

    Aggregate flexibility set.Under Assumption 1, the aug- mented sets{ ‹PE i }I i=1 are integrable i.i.d. random sets defined on a non-atomic probability space. Projecting{ ‹PE i }I i=1 onto the power coordinates shows that the flexibility sets{P E i }I i=1 12 are also integrable i.i.d. random sets defined on a non-atomic probability space. Therefore, Lemma ...

    work page

  36. [36]

    Applying Lemma 1 to{ ‹PE i }I i=1 yields lim I→∞ dH Ä‹PM I ,E î Conv(‹PE i ) óä =a.s

    Augmented power-cost set.Define the mean-field aug- mented set by ‹PM I := 1 I IM i=1 ‹PE i ⊂R 2T+1 . Applying Lemma 1 to{ ‹PE i }I i=1 yields lim I→∞ dH Ä‹PM I ,E î Conv(‹PE i ) óä =a.s. 0

    work page

  37. [37]

    Identification of the cost limit.From the definition ofC M I in (6) and of the augmented sets in Section III-B, one verifies that the truncated epigraph of the mean-field aggregate cost coincides with the mean-field augmented set, i.e.,epiC M I = ‹PM I .Combining this identity with the convergence in Step 2, we obtain lim I→∞ dH Ä epiC M I ,E î Conv(‹PE i...

    work page