Sensor-Aware Joint Data Fusion and State Estimation under Sparse Multi-rate Distribution Grid Sensing

arXiv:2610.15156 · cs.LG · Submitted 2026-10-10 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "Sensor-Aware Joint Data Fusion and State Estimation under Sparse Multi-rate Distribution Grid Sensing".

Jane: The gist This work proposes a heterogeneous latent graph ODE and a graphtransformer state estimator to solve distribution system state estimation under sparse multi-rate sensing conditions,

Tom: First, who's behind it and why it matters.

Paper summary: Tom: So we're looking at this paper, "Sensor-Aware Joint Data Fusion and State Estimation under Sparse Multi-rate Distribution Grid Sensing". It tackles the problem of figuring out the state of a power grid when you have lots of new things like local storage and generation that operators can't even monitor directly.

Jane: Exactly. Classical methods rely on everything being perfectly timed, but in reality, we get these sparse, heterogeneous sensors—like PMUs and DERs—that sample at different rates. This paper is proposing a way to handle that complexity by using a two-stage pipeline: first fusing the data from all these different sensors into something usable.

Lu: The core idea they're pushing here is a heterogeneous latent graph ODE, where the measurement functions and how the data is sampled are typed and localized right on specific buses in the grid. They pair this with a GraphTransformer state estimator that specifically solves distribution system state estimation for these multi-rate, heterogeneous sensor placements.

Meng: From an engineering standpoint, I'm interested in how they handle that heterogeneity because in production, we never get uniform data streams. If you can model different sensor types separately before fusing them, that sounds like it could make the whole pipeline more robust when things break down at scale.

Lalam: This approach extends latent graph ODEs by adding type conditioned encoders and attention mechanisms to handle the observation structure of a real feeder, which is what makes this paper different from previous work <ref:2610.15156#pg2>.

Tom: Right, so we're talking about how they build this model. The training involves two distinct objectives: first, a fusion stage that minimizes the negative ELBO of the generative process to learn the underlying data distribution, and then an estimation stage that optimizes both parameters under a masked squared error on the final state.

Jane: And they test this pipeline across three IEEE standard topologies: thirteen-bus, thirty-four-bus, and one hundred twenty-three-bus systems. The results show performance metrics for the per-unit MSE and also how it handles packet loss probability across those systems.

Lu: Looking at the numbers on page three they show that for the IEEE one hundred twenty-three-bus feeder, using DF plus SE yields a per-unit MSE of zero point zero five four three, which is compared to just the estimation alone being worse than a naive estimate <ref:2610.15156#pg3>.

Paper summary: Meng: Wait, there's more detail on page three <ref:2610.15156#pg3>. When they look at data fusion model performance across all selected state estimators on the IEEE one hundred twenty-three-bus system, they see that the GraphTransformer estimator gets an MSE of zero point one two zero one, and GCN gets a score of zero point one three seven three, which is much better than others like LG-ODE at zero point five one two three on that system <ref:2610.15156#pg3>.

Lalam: And when they look at the pipeline test MSE under per-observation packet loss probability, for the IEEE one hundred twenty-three-bus system, their pipeline error is zero point one seven five at a ninety percent loss probability <ref:2610.15156#pg3>.

Tom: That's significant because they also show that an estimator operating directly on sparse masked snapshots starts performing worse than a naive estimator returning the per-bus mean near sixty percent loss <ref:2610.15156#pg3>.

Jane: So, what does this mean for us who aren't running these massive models? It suggests that the full pipeline, fusion and estimation together, remains roughly two times better than a naive estimate or just a state estimator on its own even when you lose ninety percent of the packets.

Lu: The discussion on heterogeneity is interesting because Het-ODE separates physically distinct regimes while maintaining cross-sensor alignment for observations associated with similar grid states <ref:2610.15156#pg2>. The type-conditioned maps absorb modality specific differences before and during fusion, allowing the shared latent coordinates to represent the underlying grid state more consistently.

Meng: That sounds like a lot of complexity in the training phase, though. How much does that localized typing add to the inference cost when we are trying to deploy this on actual hardware?

Lalam: The authors place devices according to their functional role and the electrical characteristics of the distribution feeder using a rule-based assignment algorithm <ref:2610.15156#pg3>. This produces a layered measurement configuration where SCADA monitors utility infrastructure, AMI represents the customer edge, PMUs provide sparse high-resolution feeder observability, and DERs occupy physically feasible and electrically relevant connection points.

Tom: So the practical reading is that this pipeline can tolerate roughly half of all packets going missing before accuracy meaningfully suffers <ref:2610.15156#pg3>. That's a strong statement about its resilience.

Paper summary: Jane: It really shows how composition helps when things get messy in the real world, especially when observability degrades over time as you move through different grid states or load changes. This paper, "Sensor-Aware Joint Data Fusion and State Estimation under Sparse Multi-rate Distribution Grid Sensing", is showing us a path forward for estimation in these modern distribution systems.

Lu: The authors characterization of the conditions under which a composed data fusion and state estimation model fails across three grid topologies subject to real load changes and different sensing regimes is important because it defines the boundaries of this solution <ref:2610.15156#pg3>.

Meng: I'm just curious about the limitations they mentioned. What stops them from saying this works everywhere, or what is their biggest sticking point right now?

Lalam: They note that models are trained per topology on a set of fixed sensor placements, and a topology-independent estimator is future work, as is the co-design of sensor placement with the estimator <ref:2610.15156#pg3>. Also, an estimator trained on masked raw snapshots and then fed a dense fused reconstruction is evaluated off its training distribution, and composing one degrades the pipeline below fusion alone <ref:2610.15156#pg3>.

Tom: So the limitation is that it's currently topology-dependent. That means if you deploy this in a completely new grid structure, you might need to retrain or reconfigure it.

Jane: And another limitation is that they acknowledge the pipeline can fall below a trivial baseline error when observability declines and can only be rescued by composing it with a data fusion model <ref:2610.15156#pg3>. That points to where the system needs more robust design choices for truly long-term operation.

Lu: Overall, this paper is pushing the Latent Graph ODE framework forward by extending it with type conditioned encoders and attention mechanisms to properly represent the heterogeneous observation structure of a distribution feeder <ref:2610.15156#pg2>.

Meng: It's clear that for real deployment, you'd need to figure out how to make those sensor types and their roles consistent across different hardware setups, because right now it seems tied to fixed placements.

Lalam: This work provides a joint approach for distribution system state estimation under sparse multi-rate sensing conditions <ref:2610.15156#pg1>. That's the main contribution of this paper.

Conclusion: Tom: So we’re wrapping up this deep dive on "Sensor-Aware Joint Data Fusion and State Estimation under Sparse Multi-rate Distribution Grid Sensing."

Jane: Basically, they take all these messy, different sensors from a power grid—like PMUs and local storage—and figure out how to fuse that data while simultaneously estimating the actual state of the whole system.

Tom: That’s right. The title is pretty descriptive of what they did. They are using this heterogeneous latent graph ODE and this graph transformer state estimator to handle those sparse, multi-rate measurements.

Jane: The big takeaway for me is how they tackle that data mismatch between sensors without needing perfect synchronization or full observability everywhere in the grid.

Tom: Exactly. It shows a pipeline approach where you fuse the data first, and then feed that fused information into the state estimator separately trained for this specific multi-rate setup.

Jane: And their results across those IEEE test feeders show it keeps working reasonably well even when the system gets trickier, like when observability starts to degrade over time.

Tom: It’s interesting because they prove that composing a data fusion model with a state estimation model can actually rescue the estimation error, even under ninety percent packet loss scenarios.

Jane: So what does this mean for someone running an operation? It suggests that relying on just one part of the system—either just the fusion or just the estimator—isn't enough when you have these modern, complex grids.

Tom: It means you need that combined approach to stay accurate when things get sparse and noisy, which is a real-world condition for distribution systems today.

Jane: Next up, we’re going to look at the specific technical details of how this latent graph ODE actually works in their training setup.

Pooja Algikar, *, Stefano Fenu, *, Kumarsinh Jhala, *, Siby Plathottam

Argonne National Laboratory

cs.LG

Submitted: 2026-10-10

Updated: 2026-10-10

License: http://creativecommons.org/licenses/by/4.0/

The gist: The gist This work proposes a heterogeneous latent graph ODE and a graphtransformer state estimator to solve distribution system state estimation under sparse multi-rate sensing conditions,

Key concepts

Heterogeneous Latent Graph ODE (Het-ODE): This model represents the system state evolution as a continuous process on a graph. It uses type-conditioned encoders to handle different sensor types and sampling rates. This allows the model to learn a shared latent space where measurements from diverse sensors, despite their differences, can interact according to unified physical laws.
Graph Transformer State Estimator: This is the second stage of the pipeline that estimates the system state based on fused measurements. It uses transformer architecture applied to graph data, allowing it to effectively process and integrate information from heterogeneous sensor inputs into a coherent estimate of bus voltages and other system variables.
Type-Conditioned Adapters: These components map raw, heterogeneous measurement data into a shared latent space. They are trained to absorb modality-specific differences in sensor readings before fusion. This ensures that features from different sensors are aligned and represented consistently within the overall state estimation framework.

Terminology

Data Fusion — 1× in this paper · explained in 3 episodes

Summary

The gist

This work proposes a heterogeneous latent graph ODE and a graphtransformer state estimator to solve distribution system state estimation under sparse multi-rate sensing conditions, demonstrating effectiveness across multiple IEEE standard topologies as observability degrades.

Problem Context

Distribution system state estimation is necessary because the addition of local storage and generation adds complexity that an operator or aggregator is not able to directly monitor Classical methods work well when all system elements are time-synchronized and at least partly observable, but multi-rate monitoring of sparse heterogeneous sensors typically relies on machine learning approaches Modern developments have made distribution systems significantly more complex, with distributed grid resources (DGRs) and behind-the-meter storage violating both assumptions Operational tasks that hinge on system visibility, such as fault localization, congestion management, hosting capacity assessment, volt/VAR optimization, DGR setpoint dispatch therefore require state estimation in order to infer bus voltage phasors across the feeder from whatever measurements the deployed sensor mix provides

Proposed Methodology

The authors propose a heterogeneous latent graph ODE in which measurement functions and sampling processes are typed and localized on specific buses, paired with a graphtransformer state estimator solving DSSE specifically for multi-rate heterogeneous sensor placements The feeder is represented as a graph G = (V, E) with buses V = 1 to N, and the state xi(t) obeys the AC power-flow equations si(t) = Pj∈N(i) vi(t)(Yijvj (t))∗ Het-ODE builds on LG-ODE [4], which infers a per-node initial latent with a graph recognition model, evolves it with a neural ODE, and decodes the trajectory Each observation is masked to its class catalog and mapped into the shared space by a type-specific adapter, then concatenated with a learned type embedding h t,0 i = [h t,0 i eτ(i)], (1) which aligns heterogeneous feature sets before fusion The latent state evolves as z T i = z 0 i + R T 0 g i(Z t; τ)dt with shared relation and object functions and a type-specific readout z˙t i = Wτ(i) o (·), so all modalities obey one interaction law while the latent derivative at a PMU bus may be expressed differently from that at a DER bus

Training Objectives

The training proceeds in two stages with different objectives, and they are stated explicitly because they differ The fusion stage minimizes the negative ELBO of the generative process: Lrec = − Ez 0∼qϕ hP i,t log p τ(i) θ xi(t) z t i i + βk KL qϕ(z 0 O) N (0, I) The estimation stage then optimizes the fusion and estimation parameters jointly under a masked squared error on the final estimate: Lest = P i,t mi(t) xˆi(t) − xi(t) 2 P i,t mi(t)

Performance Results

The model was evaluated on three IEEE test feeders: 13-bus (13 buses), 34-bus (36), and 123-bus (130) Per-unit MSE results show that for the IEEE 123-bus feeder, the pipeline error is 0.0543 (.00331) when using DF + SE The pipeline test MSE under per-observation packet loss shows that for the 123-bus system, the error is 0.175 at a 90% loss probability The authors observe that an estimator operating directly on sparse masked snapshots begins to perform worse than a naive estimator returning the per-bus mean near 60% loss (0.180 against 0.153) The full pipeline remains roughly 2× better than a naive estimate or a state estimator alone even at 90% loss

Discussion on Heterogeneity

Het-ODE more clearly separates physically distinct regimes while maintaining cross-sensor alignment for observations associated with similar grid states This indicates that the typeconditioned maps absorb modality-specific differences before and during fusion, allowing the shared latent coordinates to represent the underlying grid state more consistently The proposed rule-based assignment algorithm places devices according to their functional role and the electrical characteristics of the distribution feeder This produces a layered measurement configuration in which SCADA monitors utility infrastructure, AMI represents the customer edge, PMUs provide sparse high-resolution feeder observability, and DERs occupy physically feasible and electrically relevant connection points The practical reading is that this pipeline tolerates roughly half of all packets going missing before accuracy meaningfully suffers

Limitations

Models are trained per topology on a set of fixed sensor placements, and a topologyindependent estimator is future work, as is the co-design of sensor placement with the estimator The authors note that an estimator trained on masked raw snapshots and then fed a dense fused reconstruction is evaluated off its training distribution, and composing one degrades the pipeline below fusion alone

Conclusion

We present a joint approach to DSSE for heterogeneous multi-rate sensing conditions that scales robustly to real feeder system sizes Our fusion model, Het-ODE, extends latent graph ODEs with type conditioned encoders and attention We characterize the conditions under which a composed data fusion and state estimation model fails across 3 grid topologies subject to real load changes under different sensing regimes and observability levels As observability declines, state estimation pipelines develop error that can fall below a trivial baseline and can be rescued by composition with a data fusion model

References

[4] Zijie Huang, Yizhou Sun, and Wei Wang. Learning continuous system dynamics from irregularlysampled partial observations. Advances in Neural Information Processing Systems, 33:16177–16187, 2020

[5] Ognjen Kundacina, Mirsad Cosovic, and Dejan Vukobratovic. State estimation in electric power systems leveraging graph neural networks. arXiv preprint arXiv:2201.04056, 2022

[6] Shiva Moshtagh, Behrouz Azimian, Mohammad Golgol, and Anamitra Pal. Topology-aware graph neural network-based state estimation for pmu-unobservable power systems. IEEE Transactions on Power Systems, 2025

[7] Arbel Yaniv, Kilian Golinski, and Christoph Goebel. Robust and interpretable graph neural networks for power systems state estimation. Electric Power Systems Research, 262:113595, 2027

--- Page 3 ---

Figure 2: Comparison of the latent representations learned by vanilla LG-ODE and Het-ODE. Small dots denote sampled latent states; open and filled circles mark the start z 0 and end z T of each mean trajectory, respectively; stars indicate sensor-type centroids; colored curves trace sensor-averaged temporal trajectories; and the black curve highlights bus 611 Het-ODE more clearly separates physically distinct regimes while maintaining cross-sensor alignment for observations associated with similar grid states

--- Page 4 ---

Table 2: IEEE 34 feeder results by SE model and modality

SE backbone se only composed GCN 0.0804 0.0002 GIN 0.1019 0.

Improvements for AI systems

Improved AI system can perform state estimation in distribution systems characterized by multi-rate heterogeneous sensor placements by utilizing a heterogeneous latent graph ODE framework that extends the latent graph ODE framework with typed encoders, dynamics, and decoders so that the observation structure of a real feeder is represented natively rather than approximated away.
Improved AI system can effectively fuse disparate sensor data by employing a mechanism where observations are mapped into a shared space via a type-specific adapter and concatenated with a learned type embedding, ensuring that all modalities obey one interaction law while the latent derivative at a PMU bus may be expressed differently from that at a DER bus.
Improved AI system can achieve superior state estimation accuracy by incorporating a graph-transformer estimator that emits a per-bus residual, defined as xˆi(t) = ˆx DF i(t) + ∆i(t) with ∆i(t) = Eϑ([ˆx DF(t)m(t)]; G), which refines the initial reconstruction.
Improved AI system can be made robust to observation loss by demonstrating that the full pipeline remains roughly 2× better than a naive estimate or a state estimator alone even at 90% loss, indicating that the composition of data fusion and estimation rescues performance when observability declines.
Improved AI system can handle unobservable systems more coherently by using Het-ODE which more clearly separates physically distinct regimes while maintaining cross-sensor alignment for observations associated with similar grid states, leading to latent representations with clearer organization by physical regime and more coherent continuous-time trajectories.

Abstract

Distribution-system state estimation is a growing necessity for the modern power grid as the addition of local storage and generation add system complexity that an operator or aggregator is not able to directly monitor. Classical methods work well when all system elements are time-synchronized and at least partly observable, but multi-rate monitoring of sparse heterogeneous sensors typically relies on machine learning approaches. It is a common practice in the literature to structure this task as a two-stage pipeline, first fusing data from disparate sensors and then feeding the fused pseudo-measurements to a separately trained state estimator. We propose a heterogeneous latent graph ODE in which measurement functions and sampling processes are typed and localized on specific buses, paired with a graph- transformer state estimator solving DSSE specifically for multi-rate heterogeneous sensor placements, and demonstrate its effectiveness across multiple IEEE standard topologies as observability degrades.

Sources

State Estimation in Electric Power Systems Leveraging Graph Neural Networks