Residual Power Flow (RPF) Overview

• Residual Power Flow (RPF) is a physics-based approach that quantifies and corrects mismatches in simplified power flow models using explicit residual mappings.
• It leverages neural and Kirchhoff-based paradigms to enhance AC feasibility, speed up simulations, and improve operational decision-making in modern power systems.
• RPF offers significant error reductions and runtime improvements over traditional methods, making it vital for real-time optimization and probabilistic assessments.

Residual Power Flow (RPF) quantifies and corrects the mismatch between simplified power system approximations and the true nonlinear AC power flow solution by constructing and learning residual mappings. RPF formalizes the infeasibility of operating points via explicit residual functions, enabling differentiable, physics-consistent surrogates suitable for probabilistic simulation, large-scale optimization, and operational decision-making in modern electric grids. RPF approaches facilitate rapid inference, flexible adaptation to evolving operational tasks, and higher accuracy compared to classical neural or linear surrogates.

1. Foundations and Mathematical Formulation

Three primary RPF paradigms have emerged: (1) residual neural learning for direct AC power flow mapping, (2) DC-to-AC optimal power flow correction, and (3) explicit Kirchhoff-based residual minimization.

AC Power Flow Equations:

The nonlinear AC power flow relations between voltage phasors $(V_i, \theta_i)$ at bus $i$ and injections $(P_i, Q_i)$ are: $$P_i = \sum_{j=1}^N V_i V_j [G_{ij} \cos\theta_{ij} + B_{ij} \sin\theta_{ij} ]$$

$$Q_i = \sum_{j=1}^N V_i V_j [G_{ij} \sin\theta_{ij} - B_{ij} \cos\theta_{ij}]$$

with $\theta_{ij} = \theta_i - \theta_j$, $G_{ij}+jB_{ij}$ entries of the admittance matrix $Y$ (Chen et al., 2023).

Kirchhoff-based RPF:

RPF reformulates AC power-flow in terms of explicit residuals:

• Nodal current balance (KCL): $r_{P,i}, r_{Q,i}$
• Cycle angle balance (KVL): $r_\ell = y_\ell \sum_k \delta_k$ over cycles $\ell$ Residual vector: $r(V, \delta; u) \in \mathbb{R}^{2N + (N-1)}$ The RPF solution minimizes

$$J(V,\delta;u) = \frac{1}{2} r(V,\delta;u)^\top W_r r(V,\delta;u)$$

Optionally, a single scalar slack variable $s$ is introduced for exact AC feasibility and uniform bus treatment (Stiasny et al., 14 Jan 2026).

Residual Learning between Approximations:

Given a baseline solution $x_{base}$ (e.g., DC-OPF or linear PF), the AC-feasible solution $x^*_{AC}$ is approximated as: $x_{AC}^* = x_{base} + \Delta x^*$ where $\Delta x^*$ is the learned residual (Za'ter et al., 17 Oct 2025).

2. Model Architectures and Initialization Schemes

Residual MLP Surrogates:

RPF architectures employ a fully-connected linear shortcut between MLP input and output: $$\mathbf{y} = F(\mathbf{x}; \theta) + \mathbf{W}_s \mathbf{x} + \mathbf{b}_s$$ $F$ is a multi-layer nonlinear mapping, $\mathbf{W}_s$ and $\mathbf{b}_s$ form the physics-guided residual shortcut (Chen et al., 2023).

Physics-Guided Initializations:

• Linearized AC-PF (decoupled): $\mathbf{W}_s = \mathbf{F}^\dagger$, $$\mathbf{b}_s = -\mathbf{F}^\dagger \mathbf{E} \mathbf{c}$$
• Jacobian (first-order Taylor): $\mathbf{W}_s = \mathbf{J}^{-1}$, $$\mathbf{b}_s = -\mathbf{J}^{-1}\mathbf{x}_0 + \mathbf{y}_0$$
• Data-driven ridge regression: fitted coefficients from training data (Chen et al., 2023)

Graph Neural Networks for Residual AC-OPF:

RPF correction models utilize topology-aware GNNs with local attention and two-level DC feature integration. Corrections are aggregated at nodes and edges; residual prediction heads generate voltage, angle, power, and flow corrections (Za'ter et al., 17 Oct 2025).

Neural RPF Solvers:

Feedforward neural networks approximate the map $u \mapsto [V^*, \delta^*]$ using either linear or learned features, trained to minimize RPF residuals (Stiasny et al., 14 Jan 2026).

3. Loss Functions, Training Protocols, and Convergence

Objective Functions:

• Mean Squared Error between MLP output and true AC state
• Physics-informed constraint violations (power-flow, box, cost deviation, residual regularization) (Chen et al., 2023, Za'ter et al., 17 Oct 2025)
• RPF residual norm for voltage/angle feasibility (Stiasny et al., 14 Jan 2026)

Training Protocols:

• L-BFGS optimizer for neural RPF surrogates, up to $6000$ epochs
• Mixed feasible/infeasible operating conditions for enhanced generalization (Stiasny et al., 14 Jan 2026)

Convergence Properties:

Physics-guided initializations yield initial MSE two orders of magnitude below random setups and accelerate convergence. RPF surrogates demonstrate rapid training and inference—$\sim$0.01 s for 5k samples, $10$–$30\times$ speedup over Newton-Raphson (Chen et al., 2023, Stiasny et al., 14 Jan 2026).

4. Comparative Performance and Benchmark Results

Empirical evaluations cover both deterministic and probabilistic metrics and include multiple reference systems:

Method/System	Angle ARMSE (IEEE-118)	Volt. ARMSE (IEEE-118)	AWD (IEEE-118)
LPF	289.3	11.93	—
FC MLP	7.36	4.07	$1.96 \times 10^{-4}$
ResNet/random-short	7.14	2.93	$2.22 \times 10^{-4}$
RPF-Data	2.70	0.79	$0.92 \times 10^{-4}$
RPF-LinPF	2.46	1.24	$0.86 \times 10^{-4}$
RPF-Jacobian	2.72	0.85	—

RPF outperforms classical linear PF, vanilla MLP, "guided" TPBNN, KNN, RR, SVR, and ResNet architectures by $2$–$10\times$ in accuracy, matching their speed and greatly surpassing MC/Quasi-MC approaches (speed-up $1000$–$3000\times$) (Chen et al., 2023). DC-to-AC residual learning yields $25$–$50\%$ lower MSE, up to $3\times$ reduction in feasibility error, and $13\times$ runtime improvement over AC-IPOPT—even for networks with $2000$ buses and N–1 topology variants (Za'ter et al., 17 Oct 2025).

5. Applications in Probabilistic, Optimal, and Real-Time Power System Tasks

Probabilistic Power Flow (PPF):

RPF enables rapid surrogate-based quantification of voltage phasor distributions under stochastic injections, reducing simulation time by orders of magnitude (Chen et al., 2023).

AC Optimal Power Flow (OPF):

Residual neural models correct DC-OPF baselines to near-AC-feasible points; GNN-based RPF architectures enforce operational limits while preserving scalability (Za'ter et al., 17 Oct 2025).

Predict-then-Optimise (PO) Framework:

RPF neural solvers embed directly into downstream optimization tasks. For AC-OPF, the cost and operational constraints are minimized together with the RPF residual norm; for quasi-steady state, slack variables (frequency, distributed slack) are optimized to minimize residuals (Stiasny et al., 14 Jan 2026).

Handling Infeasible Operating Conditions:

RPF neural solvers trained on mixed feasible/infeasible data generalize to predict minimal-residual states outside the AC-feasible region. Errors and residuals are more uniformly distributed, while classical bus-type approaches exhibit larger severity in infeasible regimes (Stiasny et al., 14 Jan 2026).

6. Implications, Advantages, and Practical Considerations

Physical Consistency and Symmetry:

Kirchhoff-based residuals eliminate bus-type asymmetries, leveraging universal nodal and cycle balance measures and a single slack variable for global feasibility restoration (Stiasny et al., 14 Jan 2026).

Interpretability and Scalability:

Physics-guided initializations and topology-aware architectures enhance model interpretability and allow efficient scaling to large systems and complex contingencies (Chen et al., 2023, Za'ter et al., 17 Oct 2025).

Speed and Flexibility:

RPF neural solvers and residual correction architectures offer $10$–$30\times$ speed-up over Newton-Raphson and AC-IPOPT methods, with subsecond inference across large grid models (Za'ter et al., 17 Oct 2025, Stiasny et al., 14 Jan 2026).

Data Generation and Security Assessment:

By decoupling the correction from baseline models, RPF enables fast generation of synthetic AC-OPF datasets and rapid N–1 contingency evaluation, supporting real-time grid security applications (Za'ter et al., 17 Oct 2025).

7. Summary and Research Directions

RPF subsumes a class of formulations centered on learning or computing residual corrections to baseline power system outputs, rigorously grounded in network physics and optimization. Its distinct features—explicit infeasibility quantification, unified bus treatment, physics-guided neural initialization, and operational speed—contribute to its superiority over previous linear, regression, and unconstrained ML methods. RPF neural solvers demonstrated accurate replication of AC solutions, flexible embedding in multi-stage control pipelines, and robust behavior under infeasible scenarios. The paradigm's continued development is expected to expand its reach into stochastic operational planning, adaptive grid control, and scalable probabilistic simulation (Chen et al., 2023, Za'ter et al., 17 Oct 2025, Stiasny et al., 14 Jan 2026).