Identification of AC Networks via Online Learning

Emanuele Fabbiani, Pulkit Nahata, Giuseppe De Nicolao, Giancarlo Ferrari-TrecatearXiv 2020 · OpenReview

TL;DR

This paper introduces an online learning algorithm to estimate the admittance matrix of AC power distribution networks, crucial for managing the increasing complexity from distributed energy resources. By combining a recursive identification method with an optimal design-of-experiment strategy, the approach efficiently learns network topology and line parameters, even in dynamically changing grids.

What Problem Does This Paper Address?

Modern power grids are becoming increasingly complex due to the integration of distributed energy resources (DERs) like solar and wind power. This shift introduces challenges such as intermittency and reverse power flows, necessitating a detailed understanding of the grid's topology and electrical parameters for reliable operation and control. However, this information is often missing, outdated, or difficult to obtain, especially in distribution networks that can experience frequent changes for maintenance or fault recovery. Traditional offline methods struggle with these dynamic environments, requiring re-estimation from scratch after any change. This paper aims to bridge this gap by providing an online learning method that can continuously update its knowledge of the grid in real-time.

Methods

The core of the proposed solution is a recursive identification algorithm that estimates the network's admittance matrix (Y) using phasor measurements of voltages and currents at each bus. This algorithm is based on the Recursive Least Squares (RLS) method, adapted for the specific structure of power network equations.

To enhance the learning process, the authors introduce a design-of-experiment (DoE) procedure. This DoE component actively influences the system by computing optimal voltage excitations for controllable DERs. The goal is to maximize the information content of the measurements collected at each step, thereby accelerating the convergence of the identification algorithm and improving its accuracy. This proactive approach contrasts with passive methods that rely solely on whatever data is available.

The paper also leverages structural properties of the admittance matrix, specifically symmetry and the Laplacian property (where applicable), to reduce the number of parameters to be estimated. They provide mathematical tools (transformation matrices D and T) to handle these structural constraints efficiently.

Datasets

The paper evaluates its proposed methods using two standard power system testbeds:

  • Grid T: A 6-bus transmission network from Wood and Wollenberg.
  • Grid D: A modified IEEE 13-bus radial feeder, augmented with controllable generators to simulate distributed generation.

These grids are used to generate synthetic voltage and current phasor measurements under various scenarios, including static networks and networks experiencing faults (line tripping).

Key Insights and Ideas

  • Online Learning for Dynamic Grids: The most significant insight is the demonstration that online learning is not just feasible but highly effective for identifying power networks that change over time. This is crucial for modern grids with DERs and reconfigurable topologies.
  • Proactive Data Collection (DoE): Actively designing the input (voltages) to maximize information gain is a powerful strategy. It significantly speeds up convergence and improves accuracy compared to purely passive recursive methods.
  • Leveraging Network Structure: Exploiting the inherent symmetry and Laplacian properties of admittance matrices dramatically reduces the estimation problem's complexity and improves the robustness of the algorithms.
  • Adaptability to Topology Changes: The online nature of the algorithm allows it to adapt quickly to sudden changes, such as line faults, by updating its estimates within a few iterations, a capability batch methods lack.

Key Results

  • The proposed DoE-enhanced recursive identification method consistently outperforms baseline recursive methods (RLS1, RLS2) and batch methods (OLS, Adaptive Lasso) in terms of accuracy (measured by Frobenius norm, max norm, and relative error) across various scenarios.
  • In scenario 1 (static network), DoE achieved error metrics comparable to or better than batch methods, while being an online algorithm.
  • In scenario 2 (network with a fault), the online methods, especially DoE, demonstrated rapid adaptation, with estimated line admittances dropping to near zero within a few iterations after a line was tripped. For example, after a fault, the estimated admittance for a tripped line was significantly lower with DoE (0.48) compared to RLS1 (2.55) after 7 iterations.
  • The method shows robustness to noise, performing acceptably even when voltage measurement noise is of the same order as current noise.
  • The DoE formulation, even when using an estimated admittance matrix from the previous step, shows only a marginal performance degradation compared to using the actual admittance matrix.

Limitations and Future Work

  • Observability Assumption: The current method assumes full observability, meaning voltage and current measurements are available at every node. Future work could explore scenarios with partial observability.
  • Error-in-Variables Models: The current approach primarily models measurement noise on the output (currents). Future research aims to extend the framework to handle error-in-variables models, accounting for noise in both voltage and current measurements.
  • Scalability: While the method is theoretically applicable to larger networks, practical implementation might face numerical stability issues due to collinearity, which DoE helps mitigate but doesn't entirely eliminate.
  • Supervisory Control: The authors suggest exploring the utility of grid identification schemes in the context of supervisory control of microgrids.

Summary updated

Abstract

The increasing penetration of intermittent distributed energy resources in power networks calls for novel planning and control methodologies which hinge on detailed knowledge of the grid. However, reliable information concerning the system topology and parameters may be missing or outdated for temporally varying electric distribution networks. This paper proposes an online learning procedure to estimate the network admittance matrix capturing topological information and line parameters. We start off by providing a recursive identification algorithm exploiting phasor measurements of voltages and currents. With the goal of accelerating convergence, we subsequently complement our base algorithm with a design-of-experiment procedure which maximizes the information content of data at each step by computing optimal voltage excitations. Our approach improves on existing techniques, and its effectiveness is substantiated by numerical studies on realistic testbeds.
BibTeX
@misc{fabbiani2020identification,
  author = {Emanuele Fabbiani and Pulkit Nahata and Giuseppe De Nicolao and Giancarlo Ferrari-Trecate},
  title = {{Identification of AC Networks via Online Learning}},
  year = {2020},
  eprint = {2003.06210},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2003.06210},
}
Figure