GENCO - A Unified Neural Solver Embedded in a Development Framework for Steady-State Grid Analysis

GENCO (GEometric Neural Corrective Optimizer) is a unified neural solver for steady-state transmission grid analysis that handles power flow (PF), optimal power flow (OPF), and state estimation (SE) within a single architecture, eliminating task-specific pipelines by operating on a shared network representation. To support advances in neural power system solvers, we introduce the open-source GridFM Development Framework (Fig. 1) that standardizes synthetic data generation and training in a low-code environment. We also release large-scale datasets with millions of PF and OPF scenarios across diverse grid topologies to further support reproducible benchmarking.

The paper presents two complementary contributions: GENCO, a unified neural solver architecture as a common backbone that operates on a shared network representation and scales to large grids of up to 10k buses, supporting Power Flow (including off-nominal operating conditions and high-order contingencies), Optimal Power Flow (supporting variable generator costs), and State Estimation (robust to partial observability and inaccurate grid parameters); and the GridFM Development Framework, released under the Apache 2.0 license via the LF Energy OpenGridFM project, which integrates tools for synthetic data generation (gridfm-datakit), model development (gridfm-graphkit), and standardized evaluation, along with open datasets containing 4 million PF/OPF instances across 8 grid topologies hosted on Hugging Face.

Building on this framework, the paper evaluates GENCO and introduces a unified benchmarking and analysis pipeline: comprehensive performance, runtime, and scalability analysis against state-of-the-art neural solvers (PF: GNS, CANOS-PF, PFNET; OPF: HH-MPNN) and classical AC/DC solvers (PowerModels) on PF∆ and OPFData benchmarks, and validation on real-world data using SCADA measurements from the Hydro-Quebec grid, demonstrating that GENCO can leverage synthetic pretraining and limited real-world finetuning to compute complete AC solutions on real-world data.

The architecture comprises a Multi-Layer Perceptron (MLP)-based input projection layer that projects grid parameters and known states into a high-dimensional latent space, followed by N iterative correction steps. At each step, a Heterogeneous Graph Transformer (HGT) layer is followed by a Solution Decoder MLP that decodes the primary variables (P̂gen, V̂, θ̂). The decoded variables can optionally be bounded by a sigmoid transformation to enforce box constraints before passing through a task-specific physics decoder to obtain a complete intermediate solution containing all injections and voltage states. Power Balance Residuals (PBRes) are then computed and fed back through a Residual Encoder MLP, enabling feasibility-aware refinement of the latent representation of the intermediate solution. The model is trained against ground-truth solutions from classical methods by minimizing a loss that combines a supervised distance term to the ground-truth solution with physics-informed residuals accumulated across all correction steps to encourage early feasibility.

For Power Flow, evaluated on the PF∆ benchmark, GENCO achieves substantially lower errors and variance than all baselines across all tasks and topology variants. In the baseline setting, the loss corresponds to only 0.28%, 0.53%, and 0.69% of the mean apparent power under (N), (N-1), and (N-2) conditions, respectively, while the best baseline (CANOS-PF) remains within the single-digit percentage range (4.2%, 6.4%, and 7.3%). GENCO is more data-efficient and more robust to close-to-infeasible cases than competing methods. On unseen grids (Task 3.1), GENCO achieves the lowest mean power-balance loss on both IEEE 57 and GOC 500, although all models suffer from substantial degradation under cross-grid evaluation.

For Optimal Power Flow, evaluated on OPFData, GENCO achieves competitive optimality (worst-case gap ≤ 0.3%) while substantially improving feasibility, with all mean normalized feasibility violations remaining below 0.35% across all grids. GENCO consistently achieves lower reactive power-balance residuals (PBResQ), often by one to two orders of magnitude compared to HH-MPNN, because reactive generation is reconstructed analytically from the reactive power-balance equations using the decoded voltage state rather than predicted directly. On larger systems (500- and 2000-bus systems), HH-MPNN typically achieves lower objective gaps, whereas GENCO attains substantially smaller branch thermal and reactive power-balance violations.

For State Estimation, GENCO reconstructs states more accurately than WLS on the small grids—its median voltage magnitude error is up to an order of magnitude lower on the IEEE 14- and 30-bus cases—with a degradation on the larger grids. As measurements become sparser and more corrupted, GENCO remains robust, increasingly outperforming WLS on small grids and closing the gap on larger grids. GENCO provides a complete estimator: unlike WLS, it returns an estimate even when the measurement configuration is insufficient for convergence, degrading gracefully rather than failing abruptly. GENCO also remains robust to inaccurate network parameters, treating the grid model as a soft prior rather than a hard constraint, and can jointly recover missing and noisy states as well as grid parameters, achieving a 30–50% reduction in noise magnitude across all tested IEEE grids.

For runtime and performance scaling, GENCO Tiny is approximately 28–29× faster than AC-PF on GOC 2,000 and GOC 10,000, making the learned approach increasingly attractive at scale. On these grids, GENCO provides a good alternative to DC-PF: unlike DC-PF, it predicts a complete AC state, including voltage magnitudes and reactive generation, achieving similar residuals while being approximately 2× slower than DC-PF. For OPF, GENCO Small is 16–85× faster than AC-OPF and 4–6× faster than DC-OPF, while reducing DC-OPF optimality gaps by 1.9–55× and feasibility violations by 2.4–86×.

For robustness and generalization, GENCO generalizes to contingency levels up to N-20 despite training only on N-2, achieving 2.85× lower median residuals than DC-PF for N-1 contingencies and 1.5× lower median residuals for N-20. GENCO remains robust under out-of-operating-limit scenarios despite training on fewer than 0.01% such out-of-limit elements, outperforming DC-PF for line-loading prediction and detecting voltage violations beyond the capabilities of DC-PF. A pretrained GENCO transfers effectively to new grids, with only 1,000 grid-specific samples sufficient to outperform DC-PF on IEEE 118, although zero-shot generalization remains an open challenge.

For validation on real data from the Hydro-Quebec grid, GENCO Tiny reduces active power-balance residuals by approximately 2.5× relative to DC-PF and substantially improves reactive power-balance residuals compared with GRIT-HQ on the synthetic HQ1200 dataset. After fine-tuning on 15,000 Hydro-Quebec SCADA samples, GENCO reaches a mean active power-balance residual of 3.36 ± 0.05 MW, approaching the 2.90 MW achieved by DC-PF on the same evaluation set. Pretraining on synthetic HQ1200 scenarios generated with gridfm-datakit enables substantially more data-efficient adaptation to real SCADA measurements, reducing the residual by an order of magnitude relative to training from scratch with the same 15,000 SCADA samples.

The paper concludes that GENCO offers a middle ground between classical AC and DC solvers: it sustains high throughput while recovering the full set of grid variables, including voltage magnitudes and reactive power. For PF on grids with ≥ 2000 bus nodes, runtime improvements of 30× w.r.t. AC-PF at residuals comparable to DC solvers were achieved. For OPF, the speedups are larger, reaching 85× for case 2000 w.r.t. AC-OPF at optimality gaps ≤ 0.3%, and feasibility within 0.35% of the mean bounds. For SE, GENCO can even outperform the weighted least squares (WLS) method, and always returns a high-quality estimate even when WLS fails to converge. Compared to task-specific machine learning models, GENCO reached state-of-the-art performance across all three steady-state grid tasks, at a fraction of development time and costs, due to the unified architecture acting as a common backbone, sharing model and learning parameters, so hyperparameter optimization is required only once, rather than repeated for every task.

Improvements to AI Systems:

1. Unified Multi-Task Neural Solver Architecture: Implement a single graph-neural-network backbone (like GENCO) that handles power flow, optimal power flow, and state estimation simultaneously, eliminating task-specific pipelines. This reduces development cost and hyperparameter tuning by 3×, as one model serves all tasks.

2. Physics-Informed Iterative Correction with Residual Feedback: Integrate a correction loop where the model decodes intermediate solutions, computes physics-based residuals (e.g., power balance equations), and feeds them back into the latent representation. This improves feasibility by 10–100× compared to direct prediction, enabling the AI to self-correct toward physically valid states.

3. Analytical Reconstruction of Secondary Variables: Instead of predicting all variables, decode only primary variables (generation, voltage magnitude, phase angle) and reconstruct secondary variables (reactive power) analytically from physical equations. This reduces reactive power balance violations by 1–2 orders of magnitude.

4. Synthetic Pretraining with Real-World Fine-Tuning: Use a standardized synthetic data generation framework (like GridFM) to pretrain on millions of diverse grid topologies, then fine-tune on limited real-world SCADA data. This achieves an order-of-magnitude improvement in data efficiency—only 1,000 real samples are needed to outperform classical DC solvers on new grids.

5. Graceful Degradation Under Incomplete/Corrupted Inputs: Design the AI to treat the grid model as a soft prior rather than a hard constraint, allowing it to return estimates even when measurements are sparse, corrupted, or parameters are inaccurate. This prevents abrupt failures and enables joint recovery of missing states and parameters (30–50% noise reduction).

6. Scalable Runtime Optimization for Large Systems: Implement a lightweight model variant (e.g., GENCO Tiny) that achieves 28–85× speedup over classical AC solvers on grids with 2,000–10,000 buses, while predicting complete AC states (including voltage magnitudes) at residuals comparable to DC solvers. This makes real-time grid analysis feasible at scale.

7. Cross-Topology Generalization with Minimal Fine-Tuning: Build the model to transfer to unseen grids with only 1,000 grid-specific samples, outperforming DC-PF on new topologies. This enables rapid deployment across different power networks without retraining from scratch.

8. Contingency and Out-of-Limit Robustness: Train on limited contingency levels (e.g., N-2) but generalize to extreme cases (N-20), achieving 1.5–2.85× lower residuals than DC-PF. The AI can also detect voltage violations and line loadings beyond DC-PF’s capabilities, even when trained on <0.01% out-of-limit scenarios.

What the Improved AI System Can Do:

• Real-Time Grid Operation: Solve power flow, optimal power flow, and state estimation in milliseconds for grids up to 10,000 buses, enabling dynamic dispatch and contingency analysis faster than classical solvers.

• Resilient Monitoring: Continuously estimate grid states from noisy, incomplete SCADA data, and even recover inaccurate network parameters—critical for aging infrastructure or cyber-physical attacks.

• Zero-Shot Adaptation: Deploy to new grids with minimal data (1,000 samples), reducing onboarding time from weeks to hours.

• High-Fidelity Decision Support: Provide complete AC solutions (voltage magnitudes, reactive power) with feasibility violations <0.35%, making it safe for operational decisions where DC approximations fail.

• Unified Development: Use one AI framework for all steady-state analysis tasks, cutting development costs by 3× and enabling rapid iteration across research and industry applications.

Foundation models are transforming business workflows and boosting productivity, yet they remain largely absent from engineering domains such as power system analysis, where strict physical consistency must be enforced. We present GENCO (GEometric Neural Corrective Optimizer), a unified neural solver for steady-state transmission grid analysis that handles power flow (PF), optimal power flow (OPF), and state estimation (SE) within a single architecture and shared network representation. To support advances in neural power system solvers, we introduce the open-source GridFM Development Framework, which standardizes synthetic data generation and training in a low-code environment. We also release large-scale datasets with millions of PF and OPF scenarios across diverse grid topologies to support reproducible benchmarking. We evaluate GENCO on the PFDelta and OPFData benchmarks against state-of-the-art neural solvers and classical solvers, including Newton-Raphson and IPOPT, as well as on real-world Hydro-Qu'ebec SCADA data. For large-scale PF, GENCO recovers the full AC operating state, including voltage magnitudes and reactive power that DC-PF cannot provide, while matching DC-PF-level active power-balance residuals. It achieves up to 30x speedups over Newton-Raphson at only 2x the runtime of DC-PF. For OPF, it achieves up to 85x speedups over IPOPT while improving feasibility, optimality, and runtime over DC-OPF. For SE, GENCO is more robust than classical weighted least squares to noisy measurements and network parameter errors, and always returns a high-quality estimate even when weighted least squares fails to converge. Together, the unified architecture and development framework provide a new approach to large-scale steady-state grid analysis, lowering the barrier to entry for power system engineers and marking a step toward Grid Foundation Models.