Learning to Pursue AC Optimal Power Flow Solutions with Feasibility Guarantees

1. Problem Statement

The paper addresses the AC Optimal Power Flow (AC OPF) problem in distribution systems equipped with controllable Distributed Energy Resources (DERs). The goal is to determine optimal active and reactive power setpoints for DERs to minimize costs (e.g., curtailment, losses) while satisfying:

• Nonlinear AC power flow constraints: Voltage magnitudes and line currents must adhere to physical laws.
• Operational limits: Voltage magnitudes ($V_{min} \le |v| \le V_{max}$) and line current limits ($|i| \le I_{max}$).
• DER constraints: Inverter capacity and power factor limits.

Key Challenges:

• Nonconvexity: The AC OPF is nonconvex, potentially admitting multiple local and global optima, making the solution map set-valued.
• Real-time Requirements: Traditional iterative solvers (like IPOPT) are too slow for real-time control (seconds/minutes) given the variability of renewable generation and loads.
• Feasibility of Learning: Existing neural network (NN) approaches that directly predict OPF solutions often fail to guarantee feasibility. Post-processing to fix violations is computationally expensive and lacks theoretical guarantees. Furthermore, approximating a set-valued mapping (multiple possible solutions) with a single-valued NN is theoretically problematic.

2. Methodology

The authors propose a Neural Network-based Safe Gradient Flow (NN-SGF) framework. Instead of training an NN to predict the final optimal solution directly, they train it to approximate the update direction of a constrained optimization algorithm.

A. The Safe Gradient Flow (SGF)

The core algorithm is based on the Safe Gradient Flow, a continuous-time dynamical system derived from Control Barrier Functions (CBFs).

• Dynamics: $\dot{u} = \eta F(u, \xi, \theta)$, where $u$ are DER setpoints, $\xi$ are measurements (voltages/currents), and $\theta$ are system parameters.
• Mechanism: At each step, the algorithm solves a Convex Quadratic Program (QP) to find a direction $z$ that:
  1. Minimizes the cost function gradient (descent direction).
  2. Satisfies linearized constraints to ensure the system state remains within a "safe" region (forward invariance).
• Inputs: The QP uses the current setpoints $u$, real-time measurements (or power flow solutions) $\xi$, and problem parameters $\theta$.

B. Neural Network Approximation

To accelerate computation, the authors replace the iterative QP solver with a Feedforward Neural Network (FNN).

• Learning Task: The NN learns the mapping $(u, \xi, \theta) \mapsto z^*$, where $z^*$ is the unique optimal solution of the convex QP defined in the SGF.
• Why this works: Unlike the AC OPF solution map (which is set-valued and discontinuous), the optimal solution map of a convex QP is single-valued and continuous (under mild assumptions), making it well-suited for NN approximation.
• Training Data: Generated offline by simulating the SGF dynamics (using linearized power flow approximations for efficiency) across various operating conditions.

C. Implementation Modes

The framework supports two distinct implementations:

1. Online Feedback-Based: Uses real-time measurements of voltages and currents from the grid. No knowledge of load profiles is required.
2. Offline Model-Based: Uses a power flow solver (e.g., Newton-Raphson) to compute voltages/currents based on known load profiles.

3. Key Contributions

1. Feasibility Guarantees without Post-processing: The method ensures practical feasibility at any time step. Even if the algorithm is terminated early (online) or before convergence (offline), the DER setpoints satisfy constraints (within a bounded error margin). This is achieved by training the NN on the "safe" update direction rather than the final solution.
2. Convergence Guarantees: The authors prove that the NN-SGF converges exponentially fast to a neighborhood of a strict local optimizer of the AC OPF.
3. Novel Learning Framework: By approximating the gradient flow update (a single-valued function) rather than the OPF solution map (a set-valued function), the method avoids the theoretical pitfalls of learning nonconvex optimization solutions directly.
4. Constraint Tightening Strategy: The paper provides a theoretical bound on constraint violations caused by linearization and NN approximation errors. This allows for constraint tightening during training (making the NN learn to be more conservative), ensuring that the final output strictly satisfies the original physical constraints.

4. Experimental Results

The method was tested on a 93-node distribution feeder with realistic time-series data for loads and solar PV generation.

• Comparison Strategies:

  • BO (Batch Optimization): Solves the full nonlinear AC OPF to convergence every 10s (using IPOPT).
  • NN-BO: Trains an NN to directly emulate the BO solution (standard approach).
  • NN-SGF: The proposed method.

• Performance Metrics:

  • Voltage Regulation: NN-SGF maintained voltages within the 0.95–1.05 p.u. limits effectively. In contrast, NN-BO exhibited significant overvoltage violations because it could not guarantee feasibility.
  • Computational Speed:
    • NN-SGF online step: 0.0026 seconds.
    • BO (IPOPT): 0.771 seconds.
    • Speedup: NN-SGF is approximately 297x faster than the batch solver and 45x faster than the online Safe Gradient Flow (without NN).
  • Cost: While NN-BO achieved slightly lower cumulative costs (by ignoring constraints), it did so at the expense of safety. NN-SGF and BO incurred higher costs due to necessary curtailment to maintain voltage safety, but NN-SGF did so much faster.

5. Significance

This paper represents a significant advancement in feedback optimization for power systems:

• Safety-Critical AI: It bridges the gap between the speed of deep learning and the rigorous safety requirements of power grids. It proves that learning-based control can be both fast and safe.
• Scalability: The ability to run in milliseconds enables the integration of high-penetration renewables and fast-acting DERs, which is critical for modernizing distribution grids.
• Theoretical Rigor: It moves beyond heuristic "black-box" learning by providing formal proofs of stability, convergence, and feasibility, addressing a major criticism of applying ML to critical infrastructure.

In summary, the proposed NN-SGF offers a computationally efficient, theoretically grounded alternative to traditional solvers, ensuring that distribution grids can operate safely and optimally in real-time despite the nonconvexity of AC power flow.