SenseFlow: A Physics-Informed and Self-Ensembling Iterative Framework for Power Flow Estimation
Zhen Zhao, Wenqi Huang, Zicheng Wang, Jiaxuan Hou, Peng Li, Lei Bai
TL;DR
SenseFlow is a novel framework that combines physics-informed neural networks and self-ensembling techniques to improve the accuracy of power flow estimation in complex electrical grids, addressing challenges like network sparsity and the Slack node.
Contribution
The paper introduces SenseFlow, integrating a physics-informed neural network with a self-ensembling iterative process to enhance power flow estimation accuracy in power systems.
Findings
Outperforms existing power flow estimation methods.
Effectively handles network sparsity and Slack node influence.
Provides high-precision voltage and angle predictions.
Abstract
Power flow estimation plays a vital role in ensuring the stability and reliability of electrical power systems, particularly in the context of growing network complexities and renewable energy integration. However, existing studies often fail to adequately address the unique characteristics of power systems, such as the sparsity of network connections and the critical importance of the unique Slack node, which poses significant challenges in achieving high-accuracy estimations. In this paper, we present SenseFlow, a novel physics-informed and self-ensembling iterative framework that integrates two main designs, the Physics-Informed Power Flow Network (FlowNet) and Self-Ensembling Iterative Estimation (SeIter), to carefully address the unique properties of the power system and thereby enhance the power flow estimation. Specifically, SenseFlow enforces the FlowNet to gradually predict…
Peer Reviews
Decision·ICLR 2026 Conference Withdrawn Submission
The paper makes an interesting attempt to tailor graph attention mechanisms specifically for power flow by emphasizing the role of the slack bus and physical dependencies. Also, the iterative, self-ensembling refinement idea shows awareness of the classical Gauss–Seidel/Newton–Raphson structure, which is conceptually appealing.
The authors seem to have missed the fundamental point of the power flow formulation — bus types (PQ, PV, slack) are not intrinsic physical categories but numerical conveniences to make the equations solvable. Also the slack bus itself merely absorbs system mismatch, not a source of “global influence” that must be learned. In reality, system works in distributed slack fashion (Dhople et. al. ) where all generators shares a mismatch of power, compared to schedule dispatch. Also, existing GNNs alre
1. The problem addressed is timely and relevant to the increasing complexity of modern power systems. 2. The presentation is clear, and the technical pipeline is well-structured. 3. The proposed architecture and iterative refinement are technically valid and leverage recent advances in graph neural networks.
1. The paper’s core contribution is not clearly distinguished from existing machine learning approaches to power flow problems, particularly end-to-end methods that enforce physical correctness. 2. There is no theoretical analysis or guarantee provided for the prediction quality, convergence speed, or robustness of the proposed method. 3. The approach does not explicitly address the issue that power flow equations can have multiple feasible solutions, raising concerns about the uniqueness and co
## Strengths - This paper embeds the neural network into the iterative process of the traditional PFA method, which differs from most existing end-to-end learning methods and makes the model’s training more conservative. - This paper distinguishes between different types of buses in PFA in order to learn representations that are more effective for PFA.
## Weakness - The term “power flow estimation” can be easily misunderstood, and from the perspective of power systems, the motivation of this paper is unclear. PFA has already been extensively studied, and even in large-scale systems with tens of thousands of buses, fast batch PFA can be achieved through high-performance computing techniques. So, why is a data-driven PFA necessary? - The description of the correlation between the two components in SenseFlow is not very clear; they appear to be
* The paper introduces an attention-based mechanism to propagate information efficiently across the power network graph, which can potentially mitigate the fact that power network graphs have high diameter. * The "self-ensembling" mechanism adds an iterative correction scheme to mitigate potential prediction errors.
* The paper does not state the mathematical problem it is trying to solve. Namely, Section 3.1 only includes an English description, it would be beneficial for non-power systems experts to include a complete mathematical statement (i.e., definitions + math equations) of the problem to be solved. * The following works are relevant for the literature review: * [Neural networks for power flow: Graph neural solver](https://doi.org/10.1016/j.epsr.2020.106547): one of the first works on graph neural
1. This work targets power flow estimation, achieving significantly higher estimation quality compared to neual-network-based baselines, and lower estimation time compared to traditional numerical solver (PyPower). 2. This work covers the scenario of missing data in a power system, which is difficult to handle in numerical solvers. 3. The domain specific technologies proposed in this work are reasonable and effective as experimental results demonstrate.
1. The motivation may be more clearly stated. 2. The impact of this work may be generalized to other domains. 3. Missing robustness discussion.
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsOptimal Power Flow Distribution · Power System Optimization and Stability · Thermal Analysis in Power Transmission
MethodsSoftmax · Attention Is All You Need