PRIME: Fast Primal-Dual Feedback Optimization for Markets with Application to Optimal Power Flow

arXiv:2504.16048 · eess.SY, cs.SY · Submitted 2025-04-22 · Read on arXiv

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: "PRIME: Fast Primal-Dual Feedback Optimization for Markets with Application to Optimal Power Flow".

Dev: Here's the gist of the paper. Online Feedback Optimization (OFO) controllers iteratively drive a plant to an optimal operating point that satisfies input and output constraints, relying solely on the input-output sensitivity as model information PRIME (PRoximal Iterative MarkEts), a novel OFO approach based on proximal-point iterations, admits a market-based implementation where self-interested actors are incentivized to make choices that result in safe and efficient operation without communicating private costs or constraints Furthermore, PRIME can handle non-smooth objective functions, achieve fast convergence rates and rapid constraint satisfaction, and effectively reject measurement noise.

Rosa: Market Model Formulation The work focuses on a market involving nu input actors and ny output actors where the goal is to minimize a combined objective function subject to input and output constraints The market operator aims to superpose incentives such that when actors minimize their local cost subject to their private constraints, the plant is steered toward a cooperative optimum.

Dev: Specifically, the mechanism involves incentives distributed by the market operator influencing actor decisions, and self-interested actors reacting by minimizing their overall cost given their private cost plus payment.

Rosa: PRIME-Y Implementation PRIME-Y OFO employs proximal minimization instead of gradient descent to update the variable u, defined as u k+1 = argmin u Y(u, λk+1y) + γu squared u - u k squared.

Paper summary: Rosa: So, this paper introduces PRIME: Fast Primal-Dual Feedback Optimization for Markets with Application to Optimal Power Flow. Basically, it takes online feedback optimization or OFO controllers and builds a new approach based on proximal-point iterations. The big idea here is to make it work in a market setting where actors are self-interested but still end up steering the plant toward a safe and efficient operating point without needing to talk about their private costs or constraints directly.

Dev: That's right, Rosa. The core claim of PRIME is that it handles non-smooth objective functions, achieves fast convergence rates, and can quickly satisfy constraints while also rejecting measurement noise that messes with other methods. It shows an efficient real-time nonlinear local marginal pricing scheme on an AC optimal power flow problem.

Taro: So it’s about making things robust when the plant model isn't perfect and when you have these actors making decisions based only on what they can see and what they know privately, right?

Rosa: Exactly. The paper sets up a market with input actors minimizing their private costs subject to their constraints, and output actors doing the same for their costs and constraints as well. The goal is to superimpose incentives so that when everyone minimizes their local cost based on what they have, the whole plant moves toward a cooperative optimum.

Dev: And PRIME does this by using proximal minimization instead of standard gradient descent to update the variables. Specifically, for PRIME-Y implementation, it updates the variable u with a quadratic term involving the square of the difference between steps and its previous value <ref:2504.16048#pg3>.

Taro: So if we're looking at how this update works in practice, it seems they are approximating a Lagrangian using its linearization at each timestep k to handle that unknown input-output map h(u) <ref:2504.16048#pg2>.

Rosa: Right. And for the PRIME-H implementation, they dualize the output constraint by introducing an auxiliary variable called z, and then they use a proximal iteration to update both u and z based on an approximation of that Lagrangian H(u, z).

Paper summary: Dev: That structure is interesting because it allows them to seek a saddle point of this Lagrangian using the proximal iteration <ref:2504.16048#pg5>. It defines the output actor's incentive in a specific way involving that auxiliary variable z and the measurement record y k <ref:2504.16048#pg2>.

Taro: When we look at the performance, simulations on an Optimal Power Flow problem show that PRIME matches the performance of standard centralized methods regarding constraint resolution and convergence speed, while still benefiting from that market-based distributability <ref:2504.16048#pg6>.

Rosa: And a really important part is its robustness. PRIME-Y and PRIME-H both deliver fast constraint violation resolution and they reject the effect of measurement noise, unlike Algorithm one which the paper points out is sensitive to noise <ref:2504.16048#pg7>.

Dev: The decentralization capabilities are also key here. The paper shows that PRIME-H allows for both input and output actors to be distributed over both sets of actors, except for the specific case involving linear output costs and constraints Table I.

Taro: That means the market operator doesn't need any private information about what the input or output actors are charging or what their limits are to figure out the incentives. It preserves privacy for everyone involved in that distribution Table I.

Rosa: So, to sum up this PRIME: Fast Primal-Dual Feedback Optimization for Markets with Application to Optimal Power Flow approach, it's a novel way of designing optimal incentives in markets with self-interested actors without them needing access to their private costs or constraints.

Dev: It’s an optimization technique that uses proximal iterations to solve the problem iteratively, which gives it those fast convergence rates and helps it handle the non-smooth nature of these problems better than older methods.

Taro: What this means for us in the real world is that we can design control systems for complex power grids or any network where many agents are making local decisions, and they can still cooperate to meet global goals efficiently, even when the model isn't perfect.

Rosa: Right. It moves the problem from a centralized calculation to a distributed one where local actions lead to a good overall outcome without needing that central authority to see everything <ref:2504.16048#pg2>.

Paper summary: Dev: We also have the caveat that the paper points out that if you use standard projected primal gradient methods, like Algorithm one and your plant model h is nonlinear or if there's measurement noise around it, the feasible set for those variables can become empty which breaks recursive feasibility <ref:2504.16048#pg3>.

Taro: So the paper highlights that PRIME is designed specifically to avoid that failure mode by using proximal minimization and dualizing the constraints instead of just relying on a projection onto a set <ref:2504.16048#pg2>.

Rosa: It’s interesting how they compare it to standard Projected Primal Gradient OFO, which uses feedback from the plant to replace the function h of u with the actual measured y k <ref:2504.16048#pg3>.

Dev: And that comparison shows that PRIME-Y and PRIME-H deliver fast constraint violation resolution while rejecting measurement noise, which is a big win for real-time control systems <ref:2504.16048#pg7>.

Taro: So when you look at the results on the UNICORN fifty-six-bus test case, the performance of PRIME aligns with standard centralized methods in terms of how fast they resolve constraints and how quickly they converge <ref:2504.16048#pg6>.

Rosa: It sounds like the real implication here is that for systems where you have many interconnected components making decisions, this market-based approach gives us a way to get good operating points without needing perfect information about every single constraint or cost function beforehand <ref:2504.16048#pg1>.

Dev: And the final part of the PRIME: Fast Primal-Dual Feedback Optimization for Markets with Application to Optimal Power Flow paper is that it gives us a way to structure incentives so that actors can make choices that are both safe and efficient, even when they are only optimizing their own local cost functions <ref:2504.16048#pg1>.

Taro: It’s about creating a system where the self-interest of the actors naturally aligns with the global needs of the power flow, and PRIME gives us a fast way to find that alignment <ref:2504.16048#pg3>.

Rosa: That's what they are showing, using proximal point iterations to drive these actors toward an optimal operating point in a dynamic environment without needing all those private details exposed <ref:2504.16048#pg1>.

Conclusion: Rosa: So we're wrapping up on PRIME, which is this new way to handle optimization in markets using proximal iterations for power flow problems.

Dev: Yeah, so basically, they took online feedback optimization and tweaked it to work with these market actors who are self-interested but still cooperate to get the best plant settings without them having all your private cost details.

Rosa: Right, so the big idea is designing these incentives in a way that doesn't require the market operator to know everything about what each actor is charging or what their limits are.

Dev: Exactly, and they do it by using this dualizing technique with an auxiliary variable, and then applying a proximal iteration to find a saddle point of that complex Lagrangian function.

Rosa: And I mean, the results on the optimal power flow problem show that this method matches the performance of standard centralized methods in terms of how fast they resolve constraints and how quickly they converge.

Dev: That’s good, but what really stands out is how PRIME-Y and PRIME-H handle measurement noise much better than older methods like Algorithm one.

Rosa: So, even when the world misbehaves with some noisy measurements, this approach keeps resolving those constraint violations fast.

Dev: And from a real-world control standpoint, that robustness is key because we can't always assume perfect data in a system running in real time.

Rosa: It means for someone trying to design control systems for complex grids or power networks, they have a way to get good operating points without needing perfect information about every single constraint or cost function beforehand.

Dev: It moves the problem from a centralized calculation to something distributed where local actions lead to a good overall outcome without that central authority having all the private details exposed.

Rosa: So it’s less about one perfect answer and more about creating a system where self-interest naturally aligns with the global needs of the power flow, and PRIME gives us a fast way to find that alignment.

Dev: That's what they show with this proximal iteration approach, but we still have to look at what happens when those market actors start making strategic choices against each other.

Nicholas Julian Behr, Mattia Bianchi, Keith Moffat, Saverio Bolognani, Florian Dorfler

ETH Zurich

eess.SY, cs.SY

Submitted: 2025-04-22

Updated: 2025-09-09

Comments: Source code available at https://github.com/NicholasBehr/prime

Journal ref: 2025 IEEE 64th Conference on Decision and Control (CDC), 5240 - 5245

DOI: 10.1109/CDC57313.2025.11312508

License: http://creativecommons.org/licenses/by-sa/4.0/

Importance score: 91/100

The gist: Online Feedback Optimization (OFO) controllers iteratively drive a plant to an optimal operating point that satisfies input and output constraints…

Key concepts

Online Feedback Optimization (OFO)
A method where a system iteratively adjusts its inputs based on the difference between desired outputs and actual measured outputs. It aims to find an optimal operating point that respects all input and output limits by using only this input-output sensitivity as model information.
PRIME (Proximal Iterative MarkEts)
A new OFO method based on proximal-point iterations. It is designed for market applications where actors are incentivized to operate safely and efficiently. It handles complex objective functions well, converges quickly, and rejects measurement noise effectively.
Market Model Formulation
This framework involves multiple input actors and output actors aiming to minimize a combined cost function subject to various constraints. The market operator designs incentives so that individual self-interested decisions lead the entire system toward a cooperative optimum.
PRIME-Y and PRIME-H Implementations
These are two specific versions of PRIME. PRIME-Y focuses on input actors, while PRIME-H introduces an auxiliary variable to dualize constraints, allowing for the distribution of incentives across both input and output actors while maintaining privacy.

Terminology

Summary

Online Feedback Optimization (OFO) controllers iteratively drive a plant to an optimal operating point that satisfies input and output constraints, relying solely on the input-output sensitivity as model information PRIME (PRoximal Iterative MarkEts), a novel OFO approach based on proximal-point iterations, admits a market-based implementation where self-interested actors are incentivized to make choices that result in safe and efficient operation without communicating private costs or constraints Furthermore, PRIME can handle non-smooth objective functions, achieve fast convergence rates and rapid constraint satisfaction, and effectively reject measurement noise. p. 0

Where the paper says it

Unlike existing OFO solutions, PRIME admits a market-based implementation, where self-interested actors are incentivized to make choices that result in safe and efficient operation, without communicating private costs or constraints. Furthermore, PRIME can handle non-smooth objective functions, achieve fast convergence rates and rapid constraint satisfaction, and effectively reject measurement noise.

Market Model Formulation

The work focuses on a market involving nu input actors and ny output actors where the goal is to minimize a combined objective function subject to input and output constraints The market operator aims to superpose incentives such that when actors minimize their local cost subject to their private constraints, the plant is steered toward a cooperative optimum. Specifically, the mechanism involves incentives distributed by the market operator influencing actor decisions, and self-interested actors reacting by minimizing their overall cost given their private cost plus payment. p. 1

Where the paper says it

In particular, note that: 1) incentives, distributed by the market operator, influence the decisions of the actors; 2) the self-interested actors react to incentives by minimizing their overall cost, given by their private cost function plus the payment due to the market operator. This mechanism is illustrated in Figure 1b.

PRIME-Y Implementation

PRIME-Y OFO employs proximal minimization instead of gradient descent to update the variable u, defined as u k+1 = argmin u Y(u, λk+1y) + γu squared u - u k squared. To handle the unknown input-output map h, the Lagrangian is approximated by Y k(u) using its linearization at timestep k. p. 3

Where the paper says it

To seek a saddle-point of the Lagrangian H, we would again like to leverage a proximal iteration [24]. However, as h is unknown, H is first approximated by H˜k (u, z) = H˜(u, z, uk , yk , νk+1 h ) where h is replaced by its linearization h˜k (4) at timestep k: H˜k (u, z) = ϕu(u) + ϕy(z) + (h˜k (u) − z) ⊤ν k+1 h + ΨCu (u) + ΨCy (z).

The resulting update for quadratic input costs involves solving an optimization problem to find u k+1 = argmin u in Cu u T Q u + c T u. p. 2
Where the paper says it

do y k ← h(u k ) {apply u to the system and sample y} λ k+1 y ← projR≥0 (λ k y + ρ Cy(y k )) u k+1 ← proxY˜k/γu (u k ) {Y˜k from Equation (7)} end for For example, in the case of quadratic input cost ϕu(u) = u ⊤Quu + c ⊤ u u, linear output cost ϕy(y) = c ⊤ y y, and linear output constraints Cy(y) = Ayy − by ≤ 0, the proximal update is solved by the following optimization: Q = Qu + γu 2 I (9a) c = cu − γuu k + ∇uh(u k ) ⊤(A ⊤ y λ k+1 y + cy) (9b) u k+1 = argmin u∈Cu u ⊤Qu + c ⊤u. (9c) 1) Market Interpretation of PRIME-Y: As per Section II, a market algorithm should superpose incentives on the private actor costs so that the aggregate behavior converges to a cooperative optimum.

PRIME-H Implementation

PRIME-H OFO proposes dualizing the input-output map constraint (2d) by introducing an auxiliary variable z and defining the Lagrangian H(u, z, νh). To seek a saddle-point of this Lagrangian, a proximal iteration is leveraged to update both u and z using the approximation H k(u, z). p. 3

Where the paper says it

Due to the presence of the indicator function ΨCy (z), the “output target” z will always satisfy the output constraints Cy(z) ≤ 0. To seek a saddle-point of the Lagrangian H, we would again like to leverage a proximal iteration [24].

The output actor's incentive p l,k z(z) is defined as ρ + γz squared (zl - z k l) T (zl - z k l) - z T l g T l (νk+1h + ρ(y k - z k)).

Performance and Robustness

Simulations on an Optimal Power Flow (OPF) problem show that PRIME matches the performance of standard centralized methods in terms of constraint resolution and convergence speed while benefiting from the distributability of a market-based implementation. Furthermore, PRIME-Y and PRIME-H deliver fast constraint violation resolution while rejecting the effect of measurement noise, unlike Algorithm 1 which is sensitive to noise. p. 4

Where the paper says it

This aggressive reaction to the measurement noise leads to the Projected Primal method’s “jittery” behavior in Figure 3b. PRIME-Y and PRIMEH instead deliver fast constraint violation resolution while rejecting the effect of the measurement noise.

Decentralization Capabilities

The degree of decentralization is summarized in Table I, showing that PRIME-H allows for both input and output actors to be distributed over both input and output actors with the exception of the case involving linear output cost and constraints. The market operator does not need any information on the input/output cost or constraints to compute the necessary incentives, which preserves privacy. p. 4

Where the paper says it

Moreover, note that the market operator does not need any information on the input/output cost or constraints to compute ν k+1 h and the incentives p j,k u and p l,k z . Therefore, PRIME-H admits a market implementation and can be distributed over both input and output actors, preserving the privacy of the cost functions and constraints of both.

Conclusion

This paper introduced PRIME, a new feedback optimization approach based on proximal iterations, which enables the design of optimal incentives in markets with self-interested actors without requiring access to their private costs or constraints. Future work should address the possible strategic behavior of market actors. p. 5

Where the paper says it

Most importantly, PRIME enables the design of optimal incentives in markets with self-interested actors without requiring access to their private costs or constraints. Future work should address the possible strategic behavior of market actors.

--- Page 1 ---

PRIME (PRoximal Iterative MarkEts), a novel OFO approach based on proximal-point iterations.

--- Page 2 ---

u1 y u2 pu 2 pu 1 Market Operator∇ u h and linear C y Φy Input Actor C1 u Φ1u Input Actor C2 u Φ2u Plant y= h(u) Fig. 1a: Market operator coordinating two input actors with PRIME-Y PRIME-Y does not allow for output actors. p. 1

Where the paper says it

1a: Market operator coordinating two input actors with PRIME-Y. PRIME-Y does not allow for output actors.

--- Page 3 ---

A Projected Primal Gradient OFO Let projUk(u) be the projection onto Uk, the feasible set for u determined by (3b) and (5). Algorithm 1 describes Projected Primal Gradient OFO [21], [22], obtained using feedback from the plant, i. p. 2

Where the paper says it

Let projUk (u) be the projection onto Uk, the feasible set for u determined by (3b) and (5). Algorithm 1 describes Projected Primal Gradient OFO [21], [22], obtained using feedback from the plant, i.

e., replacing h(u k) in (3a) with the measured y k. p. 2
Where the paper says it

Algorithm 1 describes Projected Primal Gradient OFO [21], [22], obtained using feedback from the plant, i. , replacing h(u k ) in (3a) with the measured y k .

--- Page 4 ---

Algorithm 3 PRIME-Y γu ∈ R≥0 ρ ∈ R>0 u 0 ← u initial, λ 0y ← 0 for all k = 0, 1... do y k ← h(u k) apply u to the system and sample y end for. p. 2

Where the paper says it

Algorithm 1 Projected Primal Gradient OFO α ∈ R>0 {primal learning rate} u 0 ← u initial for all k = 0, 1, . do y k ← h(u k ) {apply u to the system and sample y} uˆ k ← u k − α (∇uϕu(u k ) + ∇uh(u k ) ⊤∇yϕy(y k )) u k+1 ← projUk (ˆu k ) end for If problem (3) is convex, and in the absence of measurement noise, one can obtain global convergence guarantees for Algorithm 1 to a solution of (3) [21].

--- Page 5 ---

Algorithm 4 PRIME-H γu ∈ R≥0 γz ∈ R≥0 ρ ∈ R>0 u 0 ← u initial, z 0 ← projCy(h(u 0)), ν h ← 0 for all k = 1... do y k ← h(u k) apply u to the system and sample y νk+1h ← νk h + ρ (y k - z k) z k+1 ← proxf k/γz (z k) f k from Equation (14) u k+1 ← proxH k(u,zk)/γu(u 0). p. 3

Where the paper says it

Algorithm 4 PRIME-H γu ∈ R≥0 {proximal deviation cost in u} γz ∈ R≥0 {proximal deviation cost in z} ρ ∈ R>0 {dual learning rate} u 0 ← u initial , z 0 ← projCy (h(u 0 )), ν0 h ← 0 for all k = 0, 1, . do y k ← h(u k ) {apply u to the system and sample y} ν k+1 h ← ν k h + ρ (y k − z k ) z k+1 ← proxf k/γz (z k ) {f k from Equation (14) } u k+1 ← proxH˜k(u,zk)/γu (u k ) {H˜k from Equation (13)} end for 1) Market Interpretation of PRIME-H: The market operator provides the following incentive to each input actor j, p j,k u (uj ) = γu 2 (uj − u k j ) ⊤(uj − u k j ) + u ⊤ j e ⊤ j ∇uh(u k ) ⊤ν k+1 h , (15a) and the following incentive to each output actor l, p l,k z (zl) = ρ + γz 2 (zl − z k l ) ⊤(zl − z k l ) − z ⊤ l g ⊤ l (ν k+1 h + ρ(y k − z k )), (15b) where gl is the matrix such that yl = g ⊤ l y.

--- Page 6 ---

Fig. 3 displays the performance of the methods discussed on the UNICORN 56-bus [26] test case. p. 4

Where the paper says it

Replacing δy with (y − y k ) and δu with (u − u k ), we get h˜k (u), defined in (4), which linearizes h at timestep k. Performance Figure 3 displays the performance of the methods discussed on the UNICORN 56-bus [26] test case.

--- Page 7 ---

Fig. 4a demonstrates how a nonlinear h(u) can lead to Algorithm 1 being infeasible at initialization or after a large gradient step. p. 6

Where the paper says it

This intersection can be empty, resulting in the method returning no solution after initialization, a change in the constraints, or during the progression of gradient steps. Figure 4a demonstrates how a nonlinear h(u) can lead to Algorithm 1 being infeasible at initialization or after a large gradient step.

--- Page 8 ---

Fig. 3b, we introduce noise into the voltage measurements. p. 4

Where the paper says it

Remarkably, PRIME-Y and PRIME-H operate in a market fashion without this oracle and still achieve comparable resolution speeds. In Figure 3b, we introduce noise into the voltage measurements.

--- Page 9 ---

Table I summarizes the market interpretation distinctions TABLE I: Degree of Decentralization.

--- Page 10 ---

Fig. 2 shows the difference between PRIME-Y and PRIME-H in a small numerical example. p. 4

Where the paper says it

2: Toy example with non-zero output cost and no measurement noise. We now illustrate the difference between PRIME-Y and PRIME-H in a small numerical example in Figure 2.

--- Page 11 ---

Figure 3 also shows the behavior of Algorithm 1 while it is not a distributed market algorithm.

--- Page 7 ---

Fig. 3b shows that PRIME-Y and PRIMEH instead deliver fast constraint violation resolution while rejecting the effect of the measurement noise. p. 4

Where the paper says it

This aggressive reaction to the measurement noise leads to the Projected Primal method’s “jittery” behavior in Figure 3b. PRIME-Y and PRIMEH instead deliver fast constraint violation resolution while rejecting the effect of the measurement noise.

--- Page 7 ---

PRIME-H admits a market implementation and can be distributed over both input and output actors, preserving the privacy of the cost functions and constraints of both.

--- Page 7 ---

Table I summarizes the market interpretation distinctions TABLE I: Degree of Decentralization.

--- Page 12 ---

PRIME-H admits a market implementation and can be distributed over both input and output actors, preserving the privacy of the cost functions and constraints of both.

--- Page 13 ---

PRIME-H admits a market implementation and can be distributed over both input and output actors, preserving the privacy of the cost functions and constraints of both.

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

  1. improve incentive design in markets by using PRIME to create strictly convex incentive functions that naturally compel each actor to respond by optimizing their (private) cost. This allows for designing optimal incentives without needing access to actors' private costs or constraints.

  2. enable distributed coordination of self-interested actors in physical systems, such as power grids, by implementing PRIME-H OFO. This system allows for arbitrary constraints and cost functions as in (2) to be distributed over both input and output actors.

  3. enhance robustness against measurement noise in real-time control applications by using PRIME methods instead of the Projected Primal Gradient method. Specifically, PRIME-Y and PRIME-H reject the effect of the measurement noise while delivering fast constraint violation resolution.

Abstract

Online Feedback Optimization (OFO) controllers iteratively drive a plant to an optimal operating point that satisfies input and output constraints, relying solely on the input-output sensitivity as model information. This paper introduces PRIME (PRoximal Iterative MarkEts), a novel OFO approach based on proximal-point iterations. Unlike existing OFO solutions, PRIME admits a market-based implementation, where self-interested actors are incentivized to make choices that result in safe and efficient operation, without communicating private costs or constraints. Furthermore, PRIME can handle non-smooth objective functions, achieve fast convergence rates and rapid constraint satisfaction, and effectively reject measurement noise. We demonstrate PRIME on an AC optimal power flow problem, obtaining an efficient real-time nonlinear local marginal pricing scheme.

Related papers