Bayesian Distribution System State Estimation with Flow Matching

Introduction

Distribution system state estimation (DSSE) is notoriously difficult on medium‑voltage feeders because real‑time sensor coverage is sparse. Traditional approaches rely on pseudo‑measurements at secondary substations, effectively treating the unknown injections as a Bayesian prior. Weighted‑least‑squares (WLS) methods preserve the exact power‑flow constraints but collapse the prior to a simple mean‑variance description, while supervised deep‑learning estimators learn both the prior and the power‑flow mapping, tying them to a specific measurement set and often providing unreliable uncertainty estimates.

The recent PoSE (Posterior Sampling for State Estimation) framework tackles these shortcomings by separating what the operator knows (network topology, power‑flow equations, and measurements) from what must be learned (the statistical distribution of substation injections). PoSE trains a neural network on historical data using flow matching, a technique that learns the injection distribution without ever approximating the power‑flow equations. At inference time, the network supplies a prior, and a Kalman‑type correction incorporates the current noisy measurements, yielding posterior samples that are guaranteed to satisfy the physical constraints of the grid.

How PoSE Works

1. Learning the Injection Prior – Historical substation injection data are fed to a neural network that is trained to match the flow of a diffusion process toward the target distribution. This flow‑matching objective sidesteps the need for explicit likelihoods and produces a flexible, high‑dimensional prior that captures correlations across buses.
2. Sampling with Physical Constraints – To generate a posterior sample, PoSE draws a latent vector from the learned prior, pushes it through the network to obtain a candidate injection vector, and then runs a single iteration of a noise‑aware Kalman update. The update respects the full AC power‑flow equations, ensuring that each sample is a physically admissible operating point.
3. Adaptability – Because the neural network only models the injection distribution, the same trained prior can be reused for any measurement configuration, noise level, or even a reconfigured network topology. No retraining is required when the feeder layout changes.

Performance Highlights

• IEEE 33‑bus feeder – PoSE reduces the Continuous Ranked Probability Score (CRPS) for injection estimates by roughly 7 % compared with a full‑covariance Gaussian prior across twelve test scenarios. Its 90 % credible intervals capture the true injections at the nominal rate, whereas supervised deep‑learning estimators fall short, covering the truth only 37‑78 % of the time.
• 240‑bus unbalanced Iowa feeder – PoSE achieves the lowest errors among all tested methods for injections, voltages, and branch currents. A fast variant processes each snapshot in 0.2–0.6 seconds, matching the accuracy of the full model while remaining suitable for real‑time operation.
• Robustness to topology changes – When the feeder is reconfigured into seven different topologies, PoSE (without any additional training) outperforms both Gaussian‑prior baselines and supervised models that have been retrained for each new layout.

Practical Takeaways

• Physical consistency: Every posterior sample respects the AC power‑flow equations, eliminating the risk of infeasible estimates that can plague purely data‑driven models.
• Uncertainty quantification: By producing a full posterior distribution rather than a point estimate, PoSE enables operators to assess risk and make informed decisions under uncertainty.
• Scalability: The flow‑matching prior is learned once and reused, making the approach computationally attractive for large distribution networks and for scenarios where measurement infrastructure evolves over time.

Conclusion

PoSE demonstrates that a clean separation between learned statistical priors and deterministic physical models can dramatically improve DSSE. By leveraging flow‑matching neural networks to capture the complex, correlated behavior of substation injections, and by embedding these priors within a Kalman‑style correction that enforces power‑flow constraints, PoSE delivers accurate, uncertainty‑aware state estimates that remain robust to changes in measurement sets, noise levels, and network topology. For utilities seeking to modernize their distribution monitoring without costly sensor roll‑outs, PoSE offers a compelling blend of physics‑based rigor and machine‑learning flexibility.