Incentive-Based Load Curtailment with Limited Information: A Bilevel Zeroth-Order Learning Approach
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.
Rosa: I'm Rosa.
Dev: And I'm Dev, here with Taro.
Rosa: We're looking at the paper: "Incentive-Based Load Curtailment with Limited Information".
Dev: Here's the gist of the paper. Incentive-based load curtailment unlocks critical demand-side flexibility but is hindered by the limited knowledge of private user parameters and the inherent nonsmoothness of responses due to physical device constraints.
Rosa: Problem Formulation The paper formulates incentive-based load curtailment as a constrained bilevel optimization problem where the System Operator (SO) acts as the leader determining incentives and end-users act as followers minimizing their net cost.
Dev: The SO's upper-level problem aims to minimize total incentive payments while driving the total system load toward a target P target.
Rosa: The lower-level problem models end-users minimizing their net cost by adjusting flexible loads subject to device limits, formulated as (4a) and (4b).
Paper summary: Rosa: We talked about how this paper, Incentive-Based Load Curtailment with Limited Information: A Bilevel Zeroth-Order Learning Approach, sets up the problem as a bilevel optimization framework where the system operator tries to minimize incentive payments while steering the total load toward a target, and end-users try to minimize their own net cost subject to their device limitations >
Dev: The main challenge they tackle is that we don't have perfect knowledge of what those private user parameters are, and we also know that the actual response from users isn't smooth because of the physical constraints on their devices >
Rosa: To solve this, they introduce the Bi-ZOL algorithm, which exploits the bilevel structure to decompose the hypergradient by using a zeroth-order gradient estimator on feedback from end-users >
Dev: What they claim is that this decomposition allows them to learn this aggregate response sensitivity directly from user feedback, which provides a smoothed approximation of that response landscape >
Taro: So it’s essentially building a way to handle the nonsmoothness by focusing the learning on just the part that causes the problem, which is that user-side decision-making term >
Rosa: Right, they do this by incorporating the exact analytical gradient of the system operator's cost function alongside this zeroth-order estimate of user sensitivity >
Dev: By doing this, they claim to enhance optimality while rigorously handling those nonsmooth behaviors inherent in how users actually behave >
Taro: I wonder if that reliance on a zeroth-order estimator means we are still missing some detail about the true underlying physics of the device constraints, or are we just getting lucky with the approximation >
Rosa: The paper is focused on showing that this structural decomposition method mathematically smooths out those response landscapes and reduces the error in hypergradient estimation >
Dev: And they provide theoretical convergence guarantees to an approximate stationary point, meaning we know where it’s heading even if we can't find the absolute perfect solution immediately >
Conclusion: Rosa: So, thinking about Incentive-Based Load Curtailment with Limited Information: A Bilevel Zeroth-Order Learning Approach, the title really captures the essence of the research: we’re trying to manage load curtailment when we only have limited information about users and deal with those rough response curves >
Dev: And by proposing this Bi-ZOL algorithm, they are suggesting that we can get good results even when we don't know every single user setting perfectly, because they’ve found a way to handle the nonsmoothness structurally >
Taro: What this means for us is that we don't need perfect information from every household to implement effective demand-side flexibility programs on a large scale, as long as we can build something that learns the necessary response sensitivity >
Rosa: Exactly. It’s about making the system smarter by focusing our learning efforts on what actually drives the non-smoothness in the aggregate response, instead of trying to model every single user device detail >
Dev: The implication is that this framework gives us a path toward deploying more effective incentive-based load curtailment systems that can work in real time loops where perfect parameter knowledge isn't available >
Zhisen Jiang, Florian Dorfler, Saverio Bolognani
Automatic Control Laboratory, ETH Zurich
eess.SY, cs.SY
Submitted: 2026-05-26
Updated: 2026-05-26
Comments: 9 pages, 5 figures, submitted to PowerUP conference
Journal ref: PowerUp Conference, 2026
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 91/100
The gist: Incentive-based load curtailment unlocks critical demand-side flexibility but is hindered by the limited knowledge of private user parameters and the inherent nonsmoothness of responses due to…
Key concepts
- Bilevel Optimization
- This mathematical framework models two interacting decision-making levels: the System Operator (leader) setting incentives, and end-users (followers) adjusting their loads. The goal is to find a solution where both parties optimize their respective objectives simultaneously under these constraints.
- Zeroth-Order Gradient Estimator
- Because user responses are often non-smooth, standard gradient methods fail. This technique estimates the sensitivity of the system by observing outcomes at slightly perturbed incentive levels. It provides a smoothed approximation of the complex response landscape without needing an exact analytical formula for every user.
- Bi-ZOL Algorithm
- This is a learning method designed to handle the non-smooth nature of load curtailment responses. It uses the bilevel structure to decompose the required learning signal, effectively combining known system information with learned sensitivities from end-user feedback to find an approximate solution.
- Frank-Wolfe Stationary Point (FWSP)
- This is a specific mathematical definition used to find a stable solution for the optimization problem. It allows the algorithm to stop searching when the change in the objective function becomes very small, even when dealing with constrained and non-smooth landscapes.
Terminology
Summary
Incentive-based load curtailment unlocks critical demand-side flexibility but is hindered by the limited knowledge of private user parameters and the inherent nonsmoothness of responses due to physical device constraints. p. 0
Problem Formulation
The paper formulates incentive-based load curtailment as a constrained bilevel optimization problem where the System Operator (SO) acts as the leader determining incentives and end-users act as followers minimizing their net cost. The SO's upper-level problem aims to minimize total incentive payments while driving the total system load toward a target P target. The lower-level problem models end-users minimizing their net cost by adjusting flexible loads subject to device limits, formulated as (4a) and (4b). p. 1
How the training signal is built
The Bi-ZOL algorithm addresses the nonsmoothness by exploiting the bilevel structure to decompose the hypergradient via the chain rule. This decomposition focuses on learning the aggregate response sensitivity directly from end-user feedback via a zeroth-order gradient estimator, which provides a smoothed approximation of the landscape. p. 1 The algorithm estimates the smoothed Jacobian JR b using a two-point estimator involving observations at perturbed incentives λk + δw and λk - δw. p. 3 This estimation circumvents the SO’s limited knowledge regarding end-user parameters by utilizing a randomized smoothing of the nonsmooth response landscape. p. 3
Where the gains come from
The gains in performance stem from isolating the learning target, which is the response sensitivity term, while incorporating the exact analytical gradient of the SO’s known cost function. The algorithm defines an approximate stationary point using a (δ, ϵ)-Frank-Wolfe Stationary Point (FWSP) definition to handle the constrained, nonsmooth landscape. p. 3 The theoretical convergence guarantee shows that the expected Frank-Wolfe gap converges to a bound dependent on stepsize and smoothing radius. p. 4
Convergence Analysis
The analysis establishes the Lipschitz continuity of the ancillary function ∇φ¯δ(λ) with a derived constant L∇φ¯δ. The bias error bound shows that the difference between the approximate hypergradient and the gradient of the ancillary function is bounded by (1 + 2ρNLR) δLR. p. 7 The final proof of Theorem 1 demonstrates that with a constant stepsize γ ∈ (0, 1], Bi-ZOL guarantees convergence to a (δ, ϵ)-FWSP under Assumption 1. p. 4 Numerical experiments validate this by showing the objective value stabilizes near the global optimum with a relative optimality gap of approximately 5. p. 566%.
Conclusion
This paper presents a bilevel optimization framework for incentive-based load curtailment and proposes the Bi-ZOL algorithm which synergizes analytical model information with learning response sensitivities. The method rigorously handles nonsmooth points in aggregate responses through structural hypergradient decomposition and a two-point estimator for sensitivity. p. 5 Future work will extend this framework to incorporate grid constraints and investigate methods that reduce active probing for real-time operations. p. 5
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
-
Bilevel Zeroth-Order Learning Algorithm for Incentive-Based Load Curtailment: The improved system can solve complex load curtailment problems by exploiting the bilevel structure to decompose hypergradient, which
integrates the exact analytical information of the SO’s objective with a zeroth-order estimate of the unknown response sensitivity.
-
Robust Handling of Nonsmooth Responses: The system can manage physically constrained device responses by learning
the aggregate response sensitivity directly from end-user feedback via a zeroth-order gradient estimator,
therebymathematically smoothes the nonsmooth response landscape and reduces hypergradient estimation error.
-
Theoretical Convergence Guarantees: The algorithm provides a guarantee to reach an approximate stationary point, as stated in Theorem 1, ensuring that the learning process converges under mild assumptions even when dealing with non-smooth objectives.
-
Exploitation of Analytical Information: Unlike conventional black-box methods, the system
fully utilizes the SO’s available analytical information of the upperlevel cost function to enhance the precision of the hypergradient estimate.
-
Efficient Gradient Estimation: The use of a two-point estimator allows for learning
the aggregate behavior of end-users without access to their private parameters,
circumventing privacy concerns while maintaining accuracy through randomized smoothing.
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