Mean-Field Learning for Storage Aggregation
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.
Rosa: We're looking at the paper: "Mean-Field Learning for Storage Aggregation".
Paper summary: Rosa: So this paper, "Mean-Field Learning for Storage Aggregation," is about taking tons of individual batteries and making them work together as one big system for power grids <ref:2601.21039#pg1>. The main idea is they're building a way to find a computationally efficient and accurate model for these storage populations, even though the devices are all different <ref:2601.21039#pg1>.
Dev: It tackles that problem by interpreting aggregation as just the average behavior of a massive storage population, and they show that as this population gets bigger, the aggregate performance settles down to one unique, convex mean-field limit <ref:2601.21039#pg2>. That convexity is key because it gives them a price-responsive way to describe what the aggregate storage actually does <ref:2601.21039#pg1>.
Taro: So, if you're thinking about how a huge fleet of devices coordinates, this paper suggests that all that messiness—the individual differences and the high number of devices—smooth out when you look at the average behavior <ref:2601.21039#pg3>. It moves away from trying to model every single device individually <ref:2601.21039#pg4>.
Rosa: Exactly, they're not trying to build a perfect model for every battery; they are building a tractable surrogate that mimics the big picture behavior <ref:2601.21039#pg1>. They construct this convex surrogate model using parameters that make physical sense <ref:2601.21039#pg4>.
Dev: That surrogate model is designed to approximate the mean-field limit, which they call a convex parameterized flexibility set and a quadratic cost function <ref:2601.21039#pg4>. This lets them use standard optimization techniques to find the right parameters <ref:2601.21039#pg4>.
Taro: And they train this surrogate by looking at historical price-response data to minimize a loss function that measures how much the surrogate misses what actually happened in the past <ref:2601.21039#pg4>. That's their learning signal, minimizing the discrepancy between what the model predicts and real-world observations <ref:2601.21039#pg4>.
Rosa: What really stands out is how they use that convexity—the mean-field limit having it—to bound the error of their approximation <ref:2601.21039#pg4>. This allows them to get an estimate of how far off their surrogate model might be from the true aggregate behavior <ref:2601.21039#pg4>.
Dev: And they show that this structural insight enables a convex surrogate model with high data efficiency and approximation accuracy, even when trained on less data <ref:2601.21039#pg5>. They also demonstrate that the learned surrogate is less sensitive to how different the individual devices are <ref:2601.21039#pg4>.
Taro: So, for someone running real-world tests, what does this mean for robustness? If you use this framework, you get a structure-preserving virtual battery surrogate that satisfies the standard storage operational model <ref:2601.21039#pg5>. It's not just some black box <ref:2601.21039#pg5>.
Rosa: It means that when you schedule these aggregators, you can use this learned aggregate representation to participate in the power system operations directly <ref:2601.21039#pg5>. The operator enforces the exclusivity constraint, which helps remove those invalid regions at the top right of the feasibility set <ref:2601.21039#pg5>.
Dev: And they show that this approach is efficient because it learns the mean-field limit directly at the population level instead of modeling every single device <ref:2601.21039#pg5>. The total training time scales linearly with the number of training days, which is good for practical use <ref:2601.21039#pg5>.
Taro: But I wonder about the limitations they mention: they note that available historical data are typically insufficient to train such high-capacity models reliably <ref:2601.21039#pg2>. That’s a real constraint for deploying this in practice <ref:2601.21039#pg5>.
Rosa: So, while the framework is very accurate and robust across different levels of heterogeneity, the reliance on high-quality historical data remains a hurdle for immediate deployment <ref:2601.21039#pg5>. But they show that even with less data, the performance gap stays low <ref:2601.21039#pg5>.
Dev: Overall, this paper introduces a framework to derive an operational model for an aggregator that coordinates massive heterogeneous storage resources <ref:2601.21039#pg5>. It uses the convexity of mean-field performance to schedule these aggregators efficiently and accurately <ref:2601.21039#pg5>.
Taro: The implication is that power system operators can finally schedule these storage aggregators in a way that respects the collective behavior of the fleet, not just individual devices <ref:2601.21039#pg5>. It smooths out those device-level nonconvexities through coordinated operation <ref:2601.21039#pg5>.
Rosa: So, to wrap up this summary of "Mean-Field Learning for Storage Aggregation," it’s about using the mathematical structure of mean-field limits to create a practical, data-efficient model for huge storage fleets <ref:2601.21039#pg1>. It shows how aggregation leads to a unique convex limit that can be approximated by a physically interpretable surrogate <ref:2601.21039#pg4>.
Dev: And the results show that this learned surrogate captures both the economic value and the price-responsive behavior of the aggregator well, even under price-taking optimization settings <ref:2601.21039#pg5>. It provides a structure-preserving virtual battery model <ref:2601.21039#pg5>.
Taro: It’s about showing that coordinating massive storage resources leads to a unique convex mean-field limit, which simplifies the problem for anyone trying to manage the fleet <ref:2601.21039#pg5>. That smoothing effect is what makes it useful for real operations <ref:2601.21039#pg5>.
Rosa: And while they acknowledge that having good historical data is still a challenge, the framework itself provides strong guarantees on approximation accuracy <ref:2601.21039#pg4>. It’s a way to build an operational model that works well with the actual constraints of power system scheduling <ref:2601.21039#pg5>.
Dev: So, the paper by Jingguan Liu and colleagues proposes this mean-field learning framework to help power system operators schedule storage aggregators efficiently <ref:2601.21039#pg5>. It’s a way to handle the complexity of heterogeneous storage populations with a computationally efficient surrogate model <ref:2601.21039#pg1>.
Conclusion: Rosa: So, we’re wrapping up this look at "Mean-Field Learning for Storage Aggregation," and what you really need to know is that this paper shows how you can use a massive fleet of different batteries to work together really well <ref:2601.21039#pg5>.
Dev: Yeah, the authors are Jingguan Liu, Cong Chen, Xiaomeng Ai, Jiakun Fang, Jinsong Wang and Jinyu Wen <ref:2601.21039#pg5>. They developed this framework to find a way to model those huge populations that are all different <ref:2601.21039#pg5>.
Taro: It boils down to them using a mean-field learning method to build a surrogate model for storage aggregations <ref:2601.21039#pg5>. They’re not modeling every single device individually <ref:2601.21039#pg4>.
Rosa: Right, so the big idea is that by treating the whole population as one average behavior, you get this unique, convex limit for the aggregate performance <ref:2601.21039#pg5>.
Dev: That convexity lets them build a surrogate cost function that's physically interpretable and quadratic in nature <ref:2601.21039#pg4>. It’s about making the math fit what actually happens on the grid <ref:2601.21039#pg4>.
Taro: And they validate this by showing that even when you use less historical data, this learned surrogate still gets a really close match to what the true aggregate set is <ref:2601.21039#pg5>.
Rosa: So, the implication is that power system operators can actually schedule storage aggregators much better using this model <ref:2601.21039#pg5>.
Dev: It means they can leverage that structural insight to create a structure-preserving virtual battery model <ref:2601.21039#pg5>. That's something we need to talk about next on how this actually runs in real time <ref:2601.21039#pg5>.
State Key Laboratory of Advanced Electromagnetic Technology, Huazhong University of Science and Technology
eess.SY, cs.SY
Submitted: 2026-01-28
Updated: 2026-11-17
Comments: 15 pages, 7 figures
Code: https://github.com/JingguanLiu/Dataset-of-electricity-prices-for-meanfield-learning-paper
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 78/100
The gist: Distributed energy storage devices can be aggregated to provide operational flexibility for power systems by developing a mean-field learning framework that identifies a computationally efficient and…
Key concepts
- Mean-Field Limit
- This concept describes what happens when a very large population of independent storage devices is aggregated. The paper shows that as the number of devices grows infinitely, their collective performance converges to a single, unique convex limit. This limit is easier to model than tracking every individual device.
- Surrogate Model
- A surrogate model is a simplified mathematical representation used to approximate a complex system. Here, it's designed as a convex function with physically meaningful parameters that mimics the true mean-field limit. This simplification makes the learning process computationally efficient and accurate for real-world applications.
- Convexity
- Convexity is a mathematical property where the 'middle' of a set lies below any line connecting two points in that set. The paper leverages this property because it allows electricity prices to be interpreted as directions for bounding errors, leading to a more robust and accurate surrogate model.
Terminology
Summary
Distributed energy storage devices can be aggregated to provide operational flexibility for power systems by developing a mean-field learning framework that identifies a computationally efficient and accurate surrogate model for heterogeneous storage populations The paper proposes a mean-field learning framework to establish an accurate and computationally efficient aggregation model for heterogeneous, nonconvex, and high-dimensional storage populations
Theoretical Foundation of Aggregation
The paper interprets aggregation as the average behavior of a large storage population and shows that aggregate performance converges to a unique, convex mean-field limit as the population grows This convexity further yields a price-responsive characterization of aggregate storage behavior and allows for bounding the mean-field approximation error The framework is motivated by studying how to construct a tractable aggregation model that faithfully approximates the collective performance of distributed storage populations
Mean-Field Limit Analysis
The analysis models individual devices as i.i.d. random sets and shows that as the fleet size I → ∞, there exists a unique pair of convex mean-field limits: a convex aggregate flexibility set P L in the Hausdorff metric and a convex cost function C L(p) Proposition 1 yields two aggregation insights, namely uniqueness and convexity, which enables population-level identification of aggregate performance that accommodates heterogeneous storage parameters
Surrogate Model Construction
A convex surrogate model is designed with physically interpretable parameters to approximate the mean-field limit, consisting of a convex parameterized flexibility set PbL and a quadratic cost function CbL(p) The surrogate is formulated as a convex program by omitting the mutually exclusive charging–discharging constraint p C t p D t = 0, ∀t ∈ T, which is essential for the gradient-based learning procedure
Learning and Validation
The framework uses historical price-response data to identify surrogate parameters by minimizing a loss function F measuring the discrepancy between the surrogate-based mean-field responses and those observed in the historical dataset The learning algorithm is solved iteratively via gradient descent, where the update rule involves computing gradients of the loss function with respect to both the mean-field response and surrogate parameters Case studies validate that the learned surrogate closely matches the exact aggregate set, demonstrating high data efficiency and approximation accuracy The framework achieves superior approximation accuracy compared to existing machine learning baselines, even when trained on less data
Operational Insights
The learned surrogate faithfully captures both the aggregate economic value and the price-responsive behavior of the storage aggregator, as shown in simulation results under a price-taking optimization setting The method is less sensitive to heterogeneity than model-based methods because it is learned directly at the population level This approach allows power system operators to schedule storage aggregators efficiently and accurately
Conclusion
The paper concludes that the proposed mean-field learning framework enables power system operators to schedule storage aggregators efficiently and accurately by leveraging the convexity of mean-field performance The framework successfully derives bounds on the mean-field approximation error via a price-response analysis and identifies an accurate, convex surrogate model from population-level price-response data This results in a structure-preserving virtual-battery surrogate that satisfies the standard storage operational model The coordinated and complementary operation of a large storage fleet smooths device-level nonconvexities, yielding a unique convex mean-field limit for the storage aggregator The framework demonstrates both data efficiency and approximation accuracy in practice
Pruning redundant reasoning tokens
The paper proposes a mean-field learning framework to establish an accurate and computationally efficient aggregation model for heterogeneous, nonconvex, and high-dimensional storage populations The main contributions are threefold: 1) Theoretical Analysis: To our knowledge, this is the first work that interprets large-scale storage aggregation as the expectation of random sets in a mean-field setting Under this viewpoint, we show that as the storage population grows, aggregate performance converges to a unique, convex mean-field limit 2) Aggregation Framework: Building on the theoretical results, we design a convex surrogate model with physically interpretable parameters for storage aggregation 3) Numerical Analysis: Through numerical case studies, we demonstrate that the storage aggregator can coordinate large populations of devices, yielding a smoothed aggregate flexibility set that mitigates device-level nonconvexities
How the training signal is built
The learning objective minimizes the discrepancy between the surrogate-based mean-field responses and those observed in historical data over the dataset The loss function F measures this discrepancy, and it is minimized by selecting surrogate parameters Θ that make the surrogate model B closely match the intractable mean-field limit set A The optimization problem (16) selects Θ to minimize F(Θ) subject to constraints derived from historical price-response observations
Where the gains come from
The gains stem from the convexity of the mean-field limit, which allows electricity prices to be interpreted as supporting directions for bounding the set distance in the Hausdorff metric This structural insight enables a convex surrogate model with high data efficiency and approximation accuracy The resulting surrogate is less sensitive to heterogeneity because it is learned directly at the population level rather than modeling individual devices The framework achieves an average value gap of about 3% in price-taking optimization settings, significantly outperforming model-based methods
Loss calculation
The surrogate cost function is defined as CbL(p, Θ) = λ T p C - p D + U L(p), where UbL denotes a surrogate disutility function that approximates U L and is convex, continuous, and bounded on its power domain The parameter identification problem (16) minimizes the loss F(Θ) which measures the discrepancy between the surrogate-based mean-field responses and those observed in the historical dataset The solution algorithm involves solving a bilevel program iteratively via gradient descent to obtain surrogate parameters Θ that minimize this loss function
Parameter update
The parameter update is performed using the formula Θ(o+1) ← Θ(o) − κ X D d=1 ∂F(o)/∂z star(o) ∂z star(o) B(πd) ∂Θ(o), where κ is the learning rate This update captures the sensitivity of the response to the surrogate parameters by differentiating the KKT conditions of the inner problem The overall algorithm procedure involves initialization, solving (18) for each d, evaluating loss F(Θ(o)), computing gradients, and updating parameters iteratively
Participation in power system operations
The resulting surrogate model enables storage aggregators to participate in power system operations using a learned aggregate-level representation The operator typically enforces the exclusivity constraint (15) during scheduling, which removes invalid relaxed regions at the top right of the feasibility set, leading to an almost coinciding flexibility set This validates the effectiveness of the mean-field learning framework in approximating aggregate performance The method is efficient because it approximates the mean-field limit directly at the population level rather than modeling individual devices The total training time grows linearly with the number of training days but remains modest due to the compact physical parameterization of the surrogate model The average normalized RMSE decreases steadily with more training data, consistent with Proposition 2 The framework is robust across different heterogeneity levels, maintaining a consistently low value gap The overall results confirm that the learned surrogate faithfully captures both the aggregate economic value and the price-responsive behavior of the storage aggregator The framework is suitable for power system operators to schedule storage aggregators efficiently and accurately The paper proposes a mean-field learning framework to derive an operational model for an aggregator that coordinates massive heterogeneous storage resources The paper is 2601.21039 EVERY sentence must end with a reference token. p. 2
--- Page 1 ---
UNDER REVIEW 1
Mean-Field Learning for Storage Aggregation Jingguan Liu, Cong Chen, Xiaomeng Ai, Jiakun Fang, Jinsong Wang, and Jinyu Wen Abstract—Distributed energy storage devices can be aggregated to provide operational flexibility for power systems by developing a mean-field learning framework that identifies a computationally efficient and accurate surrogate model for heterogeneous storage populations
The learning objective minimizes the discrepancy between the surrogate-based mean-field responses and those observed in historical data over the dataset The loss function F measures this discrepancy, and it is minimized by selecting surrogate parameters Θ that make the surrogate model B closely match the intractable mean-field limit set A
The gains stem from the convexity of the mean-field limit, which allows electricity prices to be interpreted as supporting directions for bounding the set distance in the Hausdorff metric This structural insight enables a convex surrogate model with high data efficiency and approximation accuracy The resulting surrogate is less sensitive to heterogeneity because it is learned directly at the population level rather than modeling individual devices
The surrogate cost function is defined as CbL(p, Θ) = λ T p C - p D + U L(p), where UbL denotes a surrogate disutility function that approximates U L and is convex, continuous, and bounded on its power domain The parameter identification problem (16) minimizes the loss F(Θ) which measures the discrepancy between the surrogate-based mean-field responses and those observed in the historical dataset
The resulting surrogate model enables storage aggregators to participate in power system operations using a learned aggregate-level representation The operator typically enforces the exclusivity constraint (15) during scheduling, which removes invalid relaxed regions at the top right of the feasibility set, leading to an almost coinciding flexibility set This validates the effectiveness of the mean-field learning framework in approximating aggregate performance The method is efficient because it approximates the mean-field limit directly at the population level rather than modeling individual devices The total training time grows linearly with the number of training days but remains modest due to the compact physical parameterization of the surrogate model The average normalized RMSE decreases steadily with more training data, consistent with Proposition 2 The framework is robust across different heterogeneity levels, maintaining a consistently low value gap The overall results confirm that the learned surrogate faithfully captures both the aggregate economic value and the price-responsive behavior of the storage aggregator The framework is suitable for power system operators to schedule storage aggregators efficiently and accurately The paper proposes a mean-field learning framework to derive an operational model for an aggregator that coordinates massive heterogeneous storage resources The paper is 2601.21039
--- Page 5 ---
UNDER REVIEW 5
C. Mean-Field Approximation with Price-Response Bounds
Recall from Proposition 1 that the aggregate cost function of the mean-field limit admits the decomposition C L(p) = λ T p C − p D + U L(p) Similarly, we design our surrogate cost function as CbL(p) = λ T p C - p D + UbL(p), where UbL denotes a surrogate disutility function that approximates U L and is convex, continuous, and bounded on its power domain We then represent the mean-field limit performance in the joint power–disutility space by defining the augmented mean-field limit set A:= n p T, u T p ∈ P L, p T, u T ∈ epi U L o ⊂ R 2T +1
The paper is 2601.21039
--- Page 4 ---
UNDER REVIEW 4
III. MEAN-FIELD PERFORMANCE ANALYSIS As introduced at the end of Section II-A, we treat the storage population as i. p. 2i.d. random sets for our analysis (formalized in Assumption 1 below) In this section, we first review preliminaries on random sets, then theoretically establish the existence and convexity of the mean-field limits for large storage aggregations Finally, we provide a bound on the mean-field approximation error, which supports the mean-field learning method in Section IV
The paper is 2601.
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Improve surrogate model construction by utilizing a convex relaxation for computational tractability while ensuring it approximates mean-field limits, as noted in Section IV-B:
We intentionally formulate the surrogate model (13)–(14) as a convex program by omitting the mutually exclusive charging–discharging constraint pC t pD t = 0, for all t in 1 to T.
-
Enhance price-response learning efficiency by formulating the parameter identification problem to minimize discrepancy between surrogate mean-field responses and historical data, as specified in Section IV-C:
We formulate an optimization problem (16), which selects Theta to minimize the loss function F measuring the discrepancy between the surrogate-based mean-field responses and those observed in the historical dataset.
-
Improve system scheduling by using a learned aggregate representation that is less sensitive to heterogeneity, as shown in Section V.D:
It achieves an average value gap of about 3%, with the maximum remaining below 6%,
compared to model-based methods which exhibit large performance gaps under high heterogeneity.
Sources
- Statistical Inference in Mean-Field Variational Bayes
- Optimal Mechanisms for Demand Response: An Indifference Set Approach
Related papers
- Industrial Load Modeling and Optimization for Market-Based Interaction with Power Systems
- Exact and Approximate Convex Reformulation of Linear Stochastic Optimal Control with Chance Constraints
- Certified Set Convergence for Piecewise Affine Systems
- Symbolic Control for Stochastic Systems via Finite Parity Games
- Multiobjective optimization-based design and dispatch of islanded, hybrid microgrids for remote, off-grid communities in sub-Saharan Africa
- Physiology-informed layered sensing for intelligent human-exoskeleton interaction