Augmented State Estimation of Network Parameters

• The topic introduces a method that augments the state vector with network parameters to enable simultaneous inference of dynamic states and structural quantities.
• It leverages techniques like Extended Kalman Filters, dual optimization, and compressed sensing to overcome observability and computational challenges.
• Key applications include grid monitoring, brain network reconstruction, and cyber-physical security, offering practical insights for complex networked systems.

Augmented state estimation of network parameters refers to a family of methodologies for the simultaneous inference of both network states (e.g., nodal variables, dynamic states) and structural or parametric quantities (e.g., admittances, line constants, coupling strengths) from observed time series or measurement snapshots. This dual-estimation paradigm is essential in complex networked systems, such as power grids and distributed dynamical networks, where both the hidden system trajectory and the underlying system parameters are incompletely known or subject to uncertainty. Augmented state estimation is typically realized by extending the state vector to include unknown parameters, enabling their joint inference, or by formulating bi-level estimation schemes that alternate or couple state updates with parameter learning. Key applications include grid monitoring, system identification, and the reconstruction of functional connectivity in biological and engineered networks.

1. Formulation of Augmented State-Parameter Estimation

The canonical approach to augmented state estimation augments the ordinary state vector $x \in \mathbb{R}^n$ with a parameter vector $\theta \in \mathbb{R}^p$ to form the extended state:

$$x_a(k) := \begin{bmatrix} x(k) \ \theta(k) \end{bmatrix} \in \mathbb{R}^{n+p}$$

with $\theta(k+1) = \theta(k)$. The discrete-time nonlinear state-space model is then: $$\begin{align*} x(k+1) &= F(x(k), u(k), \theta), \ y(k) &= H(x(k), \theta). \end{align*}$$ This augmented model admits estimation by algorithms such as the Extended Kalman Filter (EKF), Unscented Kalman Filter (UKF), or moving-horizon estimation (MHE), operating on the composite state vector $x_a(k)$ (Liu et al., 2020).

In network settings (e.g., power systems), the problem often extends to the inference of network connectivity matrices or line parameters, which enter as $\theta$ and are typically high-dimensional, sparse, or structured objects (Basiri et al., 2017, Pagnier et al., 2021).

2. Observability and Sensitivity Analysis in Augmented Systems

Simultaneous identification of $x$ and $\theta$ is generically possible only if the augmented system is locally observable. In linearized or finite-horizon settings, observability is assessed by constructing an empirical observability matrix $S_a$, stacking the sensitivities of outputs $\theta \in \mathbb{R}^p$0 to initial conditions $\theta \in \mathbb{R}^p$1. For a window $\theta \in \mathbb{R}^p$2,

$\theta \in \mathbb{R}^p$3

with each block given recursively by the system and measurement Jacobians (Liu et al., 2020).

In cases where $\theta \in \mathbb{R}^p$4 is rank-deficient, estimation is fundamentally ambiguous for portions of $\theta \in \mathbb{R}^p$5. Sensitivity-based variable selection, via column-norm ranking and orthogonalization (e.g., a Gram–Schmidt approach), admits extraction of the maximally estimable subset of states and parameters, constraining unobservable variables to follow open-loop dynamics within the MHE framework.

3. Algorithmic Architectures and Computational Scaling

Several principal architectures exist for augmented estimation:

• Joint Kalman filtering (EKF/UKF on augmented state): The state vector is enlarged to include parameters; update and covariance propagation must then be performed on $\theta \in \mathbb{R}^p$6-dimensional structures. With $\theta \in \mathbb{R}^p$7 as in dense networks, computational cost scales as $\theta \in \mathbb{R}^p$8 per time step (Singh et al., 2021).
• Dual or bi-level optimization: State estimation (e.g., via Kalman filter) is embedded as an inner loop, treating parameters as fixed, while an outer numerical optimization minimizes a prediction-error functional over $\theta \in \mathbb{R}^p$9. Automatic differentiation or explicit backpropagation through observer steps enables efficient gradient computation; complexity is typically $$x_a(k) := \begin{bmatrix} x(k) \ \theta(k) \end{bmatrix} \in \mathbb{R}^{n+p}$$0 per time step plus $$x_a(k) := \begin{bmatrix} x(k) \ \theta(k) \end{bmatrix} \in \mathbb{R}^{n+p}$$1 per update for gradients, which is significantly more tractable for large $$x_a(k) := \begin{bmatrix} x(k) \ \theta(k) \end{bmatrix} \in \mathbb{R}^{n+p}$$2 (Singh et al., 2021, Pagnier et al., 2021).
• Compressed sensing recovery: In sparse-network settings (typical power grids), compressed sensing via $$x_a(k) := \begin{bmatrix} x(k) \ \theta(k) \end{bmatrix} \in \mathbb{R}^{n+p}$$3 minimization is applied to underdetermined measurement equations (e.g., DC power-flow), leveraging sparsity of parameter vectors (e.g., rows of the susceptance matrix $$x_a(k) := \begin{bmatrix} x(k) \ \theta(k) \end{bmatrix} \in \mathbb{R}^{n+p}$$4). Iterative and one-shot convex programs provably recover line parameters from minimal measurements, provided restricted-isometry conditions are met (Basiri et al., 2017).

Approach	Model size (state + param)	Complexity
Joint KF (Augmented)	$$x_a(k) := \begin{bmatrix} x(k) \ \theta(k) \end{bmatrix} \in \mathbb{R}^{n+p}$$5	$$x_a(k) := \begin{bmatrix} x(k) \ \theta(k) \end{bmatrix} \in \mathbb{R}^{n+p}$$6
Dual Optimization EKF	$$x_a(k) := \begin{bmatrix} x(k) \ \theta(k) \end{bmatrix} \in \mathbb{R}^{n+p}$$7 (state-step)	$$x_a(k) := \begin{bmatrix} x(k) \ \theta(k) \end{bmatrix} \in \mathbb{R}^{n+p}$$8
Compressed Sensing	Sparse parameters, per row	$$x_a(k) := \begin{bmatrix} x(k) \ \theta(k) \end{bmatrix} \in \mathbb{R}^{n+p}$$9 per row (dense case)

4. Applications in Networked and Power Systems

In power system applications, augmented estimation addresses the scenario where both the instantaneous grid state ($\theta(k+1) = \theta(k)$0: voltages, angles) and uncertain network parameters ($\theta(k+1) = \theta(k)$1: admittances, taps, shunts) must be inferred from incomplete and heterogeneous measurement data (PMUs at generators, SCADA at loads) (Pagnier et al., 2021). The network’s power-flow constraints yield nonlinear algebraic equations, which are embedded within a differentiable optimization loop using deep-learning frameworks (e.g., PyTorch), enabling first-order or advanced gradient methods to learn both state and parameters (remarkably, even with only a few unrolled Newton-Raphson steps) (Pagnier et al., 2021).

In neuroscience, the inversion of large-scale brain network models from MEG sensor data requires joint estimation of latent neural dynamics and region-to-region effective connectivity. Dual-optimization EKF approaches with embedded backpropagation achieve efficient runtime and competitive accuracy compared to joint-augmented Kalman filters, with scalability demonstrated across 100-region brain models (Singh et al., 2021).

Compressed-sensing augmented estimation reconstructs entire grid topologies (sparse $\theta(k+1) = \theta(k)$2 matrices) from sparse temporal or spatial measurements. Two classes of algorithms—one-shot, row-wise and symmetry-exploiting iterative strategies—offer practical polynomial-time reconstruction and error guarantees, even when parts of the topology or measurement streams are missing (Basiri et al., 2017).

5. Observability Limitations and Variable Selection

When the system’s outputs are insufficient to render all states and parameters simultaneously observable (as detected by the rank of $\theta(k+1) = \theta(k)$3), estimation suffers from structural ambiguities. In such cases, restricting the estimator to a variable subset—selected by orthogonalized sensitivity analysis to maximize output excitation—enables robust identification of the “most estimable” variables, freezing the rest. This approach produces near-optimal estimation error profiles in benchmark nonlinear systems, reducing root-mean-square errors by up to 50% compared to full estimation when the full set is not observable (Liu et al., 2020).

Tuning the selection threshold (cutoff $\theta(k+1) = \theta(k)$4 relative to measurement and process noise) provides robustness and prevents overfitting, with best performance empirically observed at intermediate variable set sizes, matching the empirical rank of the observability matrix.

6. Implementation and Case Study Performance

Implementation varies with application and system scale:

• In high-dimensional networks, dual-optimization schemes enable rapid convergence (e.g., hours for 100-node MEG inversions), with error and reliability scaling favorably versus traditional joint KFs. Empirical studies report parameter estimation RMSEs remaining competitive or outperforming joint approaches, especially as network size increases (Singh et al., 2021).
• Compressed-sensing–based topology recovery on realistic power systems (e.g., IEEE 118-bus, up to 300 buses) achieves perfect reconstruction of sparse admittance matrices with $\theta(k+1) = \theta(k)$5 measurement samples, where $\theta(k+1) = \theta(k)$6 is nodal degree, confirming theoretical guarantees (Basiri et al., 2017).
• In moving-horizon estimators with variable selection, estimation accuracy is maximized when the cardinality of the selected variable set matches the effective observability of the system. Freezing low-sensitivity variables mitigates ill-conditioning and noise-induced drift in the estimates (Liu et al., 2020).

7. Extensions, Limitations, and Open Problems

Augmented state estimation is fundamentally limited by system observability, sensitivity profile, and identifiability under the available measurement regime. Assumptions of known noise structure ($\theta(k+1) = \theta(k)$7 in filtering methods) are commonly made, yet online estimation of noise parameters remains challenging and is an active area for extension. The use of sparsity-promoting or structured priors (e.g., for $\theta(k+1) = \theta(k)$8) offers improved identifiability in underdetermined settings (Basiri et al., 2017, Singh et al., 2021).

For severely nonlinear or heavy-tailed noise regimes, extensions to particle filtering, unscented or higher-order Kalman filters, and nonlinear optimization techniques are plausible. Embedding physically accurate process models as differentiable layers within machine learning frameworks provides a route to seamless integration of physics-based constraints with data-driven parameter inference (Pagnier et al., 2021).

The applicability of these frameworks extends to cyber-physical security monitoring, attack detection (in dynamic grid topologies), and the real-time identification of abrupt changes in network structure or dynamical regime.

---

References

• (Basiri et al., 2017) "From State Estimation to Network Reconstruction"
• (Liu et al., 2020) "Simultaneous state and parameter estimation: the role of sensitivity analysis"
• (Pagnier et al., 2021) "Embedding Power Flow into Machine Learning for Parameter and State Estimation"
• (Singh et al., 2021) "Efficient state and parameter estimation for high-dimensional nonlinear system identification with application to MEG brain network modeling"