Cross-layer Design for Real-Time Grid Operation: Estimation, Optimization and Power Flow

arXiv:2109.13842 · eess.SY, cs.SY · Submitted 2021-09-28 · 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: "Cross-layer Design for Real-Time Grid Operation".

Dev: Here's the gist of the paper. In this paper, we propose a combined Online Feedback Optimization (OFO) and dynamic estimation approach for a real-time power grid operation under time-varying conditions A dynamic estimation uses grid measurements to generate the information required by an OFO controller, that incrementally steers the controllable power injections set-points towards the solutions of a time-varying AC Optimal Power Flow (AC-OPF) problem Problem Formulation and Context The operation of power systems is experiencing an extensive transformation due to the introduction of controllable elements, pervasive sensing, and communication networks that enable fast sampling and control-loop rates Grid operation can be optimized by solving an AC optimal power flow (AC-OPF) problem

1: to determine controllable power injection set-points that minimize operation costs and satisfy grid operational constraints The AC-OPF (3) is not a static problem, it changes with time due to the temporal variations of the physical grid quantities θt =

d T t,¯u T t, u T t: T ∈ Θ caused by, e.g., a sudden change in the power demand d Standard AC-OPF (3) requires monitoring the grid to obtain full measurements of d, unfortunately these are not available in most low-voltage grids Proposed Control Architecture The authors propose a feedback architecture combining a measurement-driven dynamic estimation, e.g., a Kalman filter

26: , and an OFO approach

2: –

5: This approach allows us to monitor the grid state, and use it as feedback to update the input set-points at sub-second sampling and control-loop rates The architecture consists of three main blocks: Online Feedback Optimization (OFO), Dynamic estimation, and Online power flow solver Online Power Flow Solver Design Since a closed-form expression for the high-voltage solution map χ(·) is not available, the dynamic estimation (6) requires evaluating χ(ut, ˆdt−1 + µt), and χ(ut, ˆdt) using a power flow solver The authors design an efficient online power flow solver based on the recently proposed sensitivity-conditioning

17: This solver generates a power flow approximation zt ≈ xˆt, and updates it at every step using an operator F: R nx × U squared × D squared → R nx: zt+1 = F(zt, ut+1, ut, ˆdt+1, ˆdt) Convergence and Stability Guarantees The paper certifies the stability and convergence of this combined approach under time-varying conditions Proposition 1 establishes that if z0 is initialized sufficiently close to the power flow solution χ(u0, ˆd0), and the number of steps N of the online power flow solver FN (·) is sufficiently large, then Algorithm 1 is well-defined Theorem 1 establishes that if the initial set-points u0 are close to a strict local minimum u∗(ˆθ0), and the values kˆθt+1 − ˆθtk2 and kxˆt−ztk2 are sufficiently small, then the set-points ut produced by Algorithm 1 converge to a neighborhood of the time-varying solution u∗(ˆθt) with finite tracking error Theorem 2 shows that under Assumptions 1 and 2, there exists a KL-function β(·) and a K-function γ(·), such that utU∗(ˆθt)≤ β(u0U∗(ˆθ0), t) + γ supk<t kˆθk+1− ˆθkk2 + kxˆk − zkk2 Simulation Validation In the test case, the authors simulate and validate the proposed combined OFO and estimation approach in Algorithm 1 on a benchmark test feeder They compare it against a projected gradient descent operator based on (8) used in

5: during a 1-hour simulation with 1-second sampling period The results show that with both operators the optimization errors decrease to close to 0, but not exactly 0, due to the effects of the time-varying conditions in θ and the estimation error kdt − ˆdtk2 However, while the inputs uQP of quadratic operator (9) need only around 30 seconds (hence time steps) to converge, the projected gradient descent uP G requires about 30 minutes The quadratic programming operator (9) can handle constraints directly and efficiently In this paper, we proposed an algorithm for real-time grid operation that combines Online Feedback Optimization (OFO) and dynamic estimation with an online power flow solver First, we have shown that our online power flow solver is stable and converges despite having control inputs and disturbances that change simultaneously Then, we have proved that a quadratic-programming-based OFO converges to a neighborhood of the AC-OPF solutions, with a quantifiable small tracking error depending on the size of exogenous disturbances, and the power flow approximation Finally, we have observed in a simulated test case how our approach succeeds in tracking the solution of the AC-OPF, and adapts very quickly to time-varying conditions The paper is structured as follows References

Paper summary: Rosa: Okay, so to recap this Cross-layer Design for Real-Time Grid Operation: Estimation, Optimization and Power Flow paper. They propose a combined approach using online feedback optimization with dynamic estimation to manage real-time power grid operation when things change over time.

Dev: The core thesis is that they use measurements from the grid to create information that feeds into the optimization controller, which then incrementally steers controllable power injections toward the solution of a time-varying AC optimal power flow problem.

Taro: So, it’s not just one thing; it’s cross-layer design. They’re linking estimation, optimization, and the actual power flow solver together in this feedback architecture so that everything operates at sub-second sampling rates.

Rosa: Right. The paper claims they propose a quadratic programming based of online feedback optimization that guarantees satisfying the grid operational constraints, like those admissible voltage limits. This is important because it makes sure the solution we get is physically possible for the grid.

Dev: Within the estimation block, they design an online power flow solver that computes power flow approximations in real time. This approximation is what feeds into the optimization controller to make its decisions without having to solve a massive system equation repeatedly.

Taro: So, the main claim here is that this combined setup provides stability and convergence guarantees under time-varying conditions, which proves it’s reliable when the grid environment isn't static.

Rosa: That reliability is backed by their simulation validation on an IEEE one hundred twenty-three-bus test feeder <ref:2109.13842#pg2,IEEE 123-bus test feeder>. They show superior performance compared to a gradient based OFO operator that was used in another paper during a one hour simulation with a one second sampling period.

Dev: The key result there is the speed comparison, where the quadratic programming operator needed only around thirty seconds to converge, while the projected gradient descent version took about thirty minutes. This shows their method can find solutions much faster for real-time needs.

Taro: It’s about computational efficiency in a system that has to react fast. If you need a solution in thirty seconds, that’s very different from needing it in thirty minutes when the system is fluctuating rapidly.

Rosa: So, the paper is essentially showing a way to combine estimation and optimization so that we get real-time set-points for power injections that track the changing optimal power flow solutions effectively. This is what matters for real-time control.

Conclusion: Dev: So, looking at Cross-layer Design for Real-Time Grid Operation: Estimation, Optimization and Power Flow by Picallo et al., the authors have put forward a system that integrates estimation, optimization, and power flow into a single feedback architecture.

Rosa: The title highlights how they are doing this cross-layer design—linking these three distinct parts to achieve real-time operation under time-varying conditions. It’s about making the whole process work together seamlessly when the grid is dynamic.

Taro: What this means in simpler terms is that we have an algorithm that takes noisy, changing measurements and uses them immediately to calculate optimal control set-points for the power grid.

Dev: Precisely. It's not just about finding a solution; it’s about tracking the time-varying optimal solutions with a quantifiable small tracking error dependent on how big your disturbances and your initial guess are.

Rosa: So, the real implication is that we have a system for power systems that can react very quickly to sudden shifts in demand or generation without needing perfect measurements of everything upfront.

Dev: The authors have shown that their quadratic programming based OFO converges to a neighborhood of the AC-OPF solutions, and this error is bounded by terms related to the size of your initial disturbances and the approximation from the power flow solver.

Taro: It means for deployment, you don't need perfect inputs; you just need an estimation block that runs fast enough so it doesn't introduce too much lag into the control loop.

Rosa: Exactly. So, this work gives us a practical way to implement adaptive control for power grids that can handle the mess of real-world physics in a dynamic environment.

Miguel Picallo, Dominic Liao-McPherson, Saverio Bolognani, Florian Dorfler

Automatic Control Laboratory at ETH Zurich

eess.SY, cs.SY

Submitted: 2021-09-28

Updated: 2022-05-16

Journal ref: Electric Power Systems Research, 2022 212, 108378

DOI: 10.1016/j.epsr.2022.108378

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

Importance score: 85/100

The gist: In this paper, we propose a combined Online Feedback Optimization (OFO) and dynamic estimation approach for a real-time power grid operation under time-varying conditions A dynamic estimation uses…

Key concepts

AC Optimal Power Flow (AC-OPF)
This is the mathematical problem used to find the best way to set power injections in a grid. It minimizes operational costs while respecting all physical limits of the system. Because grid conditions change over time, this problem is not static, requiring continuous recalculation.
Online Feedback Optimization (OFO)
OFO is a control strategy that continuously optimizes decisions based on real-time feedback from the system. In this paper, it uses measurements to update the desired power injection set-points incrementally at high speeds, allowing the system to react quickly to changes.
Dynamic Estimation
This component uses grid measurements (like voltage and current) to estimate unmeasured variables needed for optimization. It functions like a filter that tracks how the grid is actually behaving in real-time, providing accurate information to guide the optimization controller.

Terminology

Summary

In this paper, we propose a combined Online Feedback Optimization (OFO) and dynamic estimation approach for a real-time power grid operation under time-varying conditions A dynamic estimation uses grid measurements to generate the information required by an OFO controller, that incrementally steers the controllable power injections set-points towards the solutions of a time-varying AC Optimal Power Flow (AC-OPF) problem

Problem Formulation and Context

The operation of power systems is experiencing an extensive transformation due to the introduction of controllable elements, pervasive sensing, and communication networks that enable fast sampling and control-loop rates Grid operation can be optimized by solving an AC optimal power flow (AC-OPF) problem [1] to determine controllable power injection set-points that minimize operation costs and satisfy grid operational constraints The AC-OPF (3) is not a static problem, it changes with time due to the temporal variations of the physical grid quantities θt = [d T t,¯u T t, u T t] T ∈ Θ caused by, e.g., a sudden change in the power demand d Standard AC-OPF (3) requires monitoring the grid to obtain full measurements of d, unfortunately these are not available in most low-voltage grids

Proposed Control Architecture

The authors propose a feedback architecture combining a measurement-driven dynamic estimation, e.g., a Kalman filter [26], and an OFO approach [2]–[5] This approach allows us to monitor the grid state, and use it as feedback to update the input set-points at sub-second sampling and control-loop rates The architecture consists of three main blocks: Online Feedback Optimization (OFO), Dynamic estimation, and Online power flow solver

Online Power Flow Solver Design

Since a closed-form expression for the high-voltage solution map χ(·) is not available, the dynamic estimation (6) requires evaluating χ(ut, ˆdt−1 + µt), and χ(ut, ˆdt) using a power flow solver The authors design an efficient online power flow solver based on the recently proposed sensitivity-conditioning [17] This solver generates a power flow approximation zt ≈ xˆt, and updates it at every step using an operator F: R nx × U squared × D squared → R nx: zt+1 = F(zt, ut+1, ut, ˆdt+1, ˆdt)

Convergence and Stability Guarantees

The paper certifies the stability and convergence of this combined approach under time-varying conditions Proposition 1 establishes that if z0 is initialized sufficiently close to the power flow solution χ(u0, ˆd0), and the number of steps N of the online power flow solver FN (·) is sufficiently large, then Algorithm 1 is well-defined Theorem 1 establishes that if the initial set-points u0 are close to a strict local minimum u∗(ˆθ0), and the values kˆθt+1 − ˆθtk2 and kxˆt−ztk2 are sufficiently small, then the set-points ut produced by Algorithm 1 converge to a neighborhood of the time-varying solution u∗(ˆθt) with finite tracking error Theorem 2 shows that under Assumptions 1 and 2, there exists a KL-function β(·) and a K-function γ(·), such that utU∗(ˆθt)≤ β(u0U∗(ˆθ0), t) + γ supk<t kˆθk+1− ˆθkk2 + kxˆk − zkk2

Simulation Validation

In the test case, the authors simulate and validate the proposed combined OFO and estimation approach in Algorithm 1 on a benchmark test feeder They compare it against a projected gradient descent operator based on (8) used in [5] during a 1-hour simulation with 1-second sampling period The results show that with both operators the optimization errors decrease to close to 0, but not exactly 0, due to the effects of the time-varying conditions in θ and the estimation error kdt − ˆdtk2 However, while the inputs uQP of quadratic operator (9) need only around 30 seconds (hence time steps) to converge, the projected gradient descent uP G requires about 30 minutes The quadratic programming operator (9) can handle constraints directly and efficiently

Conclusion

In this paper, we proposed an algorithm for real-time grid operation that combines Online Feedback Optimization (OFO) and dynamic estimation with an online power flow solver First, we have shown that our online power flow solver is stable and converges despite having control inputs and disturbances that change simultaneously Then, we have proved that a quadratic-programming-based OFO converges to a neighborhood of the AC-OPF solutions, with a quantifiable small tracking error depending on the size of exogenous disturbances, and the power flow approximation Finally, we have observed in a simulated test case how our approach succeeds in tracking the solution of the AC-OPF, and adapts very quickly to time-varying conditions The paper is structured as follows

References

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. Verification of stability under fast time variation: The system can be designed to guarantee that the online power flow solver (7) is stable and converges despite having control inputs and disturbances that change simultaneously. This ensures real-time feasibility even when grid conditions shift rapidly.

  2. Tracking capability for optimal set-points: The improved AI system can continuously steer controllable power injections towards the minimizer of a time-varying AC Optimal Power Flow problem by using the quadratic programming based OFO approach, ensuring it tracks solutions despite changes in exogenous disturbances like load variations or renewable generation.

  3. Robustness against model mismatch: By incorporating a dynamic estimation that includes a nonlinear power flow model, the system gains robustness against model mismatch and unmeasured disturbances, which is crucial for distribution grids with scarce measurements.

  4. Efficient real-time computation: The system can achieve fast control-loop rates by using an approximate online power flow solver, such as operator F(cdot) in (7), which is designed to be computationally efficient and suitable for embedded hardware, avoiding the high computational cost of solving the full nonlinear power flow equation at every iteration.

  5. Quantifiable tracking error: The system can provide a quantifiable measure of performance by showing that the set-points converge to a neighborhood of the time-varying solution with a tracking error determined by the power flow approximation error zt minus x hat t and the temporal variation hat theta t plus one minus hat theta t.

Abstract

In this paper, we propose a combined Online Feedback Optimization (OFO) and dynamic estimation approach for a real-time power grid operation under time-varying conditions. A dynamic estimation uses grid measurements to generate the information required by an OFO controller, that incrementally steers the controllable power injections set-points towards the solutions of a time-varying AC Optimal Power Flow (AC-OPF) problem. More concretely, we propose a quadratic programming-based OFO that guarantees satisfying the grid operational constraints, like admissible voltage limits. Within the estimation, we design an online power flow solver that efficiently computes power flow approximations in real time. Finally, we certify the stability and convergence of this combined approach under time-varying conditions, and we validate its effectiveness on a simulation with a test feeder and high resolution consumption data.

Sources

Related papers