Data-Aware Retrodiction for Asynchronous Harmonic Measurement in a Cyber-Physical Energy System

Youda Liu, Xue Wang, Yanchi Liu, Sujin Cui

TL;DR

This paper tackles the "out-of-sequence measurement" (OOSM) problem in cyber-physical energy systems (CPES) where data from sensors arrives at the central processing unit with varying delays. They propose a novel "data-aware retrodiction" method using a Nonlinear Autoregressive Exogenous (NARX) neural network to accurately reconstruct the timing of these delayed measurements, significantly improving harmonic analysis and identification in smart grids.

What Problem Does This Paper Address?

Modern smart grids, or Cyber-Physical Energy Systems (CPES), rely on a vast network of sensors to monitor electrical parameters for safety, reliability, and efficiency. However, in large-scale, distributed sensing networks, data from different sensors can arrive at the central data management system with varying latencies. This "out-of-sequence measurement" (OOSM) problem disrupts the temporal integrity of the data, leading to inaccurate harmonic analysis, which is crucial for power quality assessment, fault localization, and smart grid management. Traditional methods often struggle with the complex, non-linear, and time-varying nature of harmonic data, and many rely on computationally intensive models that are difficult to implement in real-time.

Methods

The core of this paper is the proposal of a data-aware retrodiction method using a Nonlinear Autoregressive Exogenous (NARX) neural network. Unlike traditional methods that rely on predefined kinematic models of harmonic behavior, the NARX network learns the complex dynamics directly from the practical electrical data.

Here's a breakdown of their approach:

  • NARX Network: This type of recurrent neural network is well-suited for time-series prediction and filtering. It takes past inputs (exogenous) and past outputs (autoregressive) to predict future values. In this context, it learns the patterns of harmonic data, even when some data points are missing due to OOSM.
  • Data-Aware Retrodiction: The NARX network is trained to "retrodict" or reconstruct the missing time-ordered data. When an out-of-sequence measurement arrives, the network uses its learned model to estimate what the measurement should have been at its correct time. This is done by propagating the error backward through the network and updating its weights.
  • Multi-Lag OOSM: The paper specifically addresses multi-lag OOSM, where data can be delayed by multiple time steps. They present algorithms for both single-lag and multi-lag scenarios, with the multi-lag solution recursively applying the single-lag correction.
  • Comparison with Existing Methods: The proposed NARX-based method is compared against established OOSM algorithms, including Kalman filter-based methods (like Bl, ALG-I) and particle filter-based methods (like SEPF).

Datasets

The paper uses experimental data collected from a two-way distributed metering framework built to simulate a cyber-physical energy system. This testbed includes:

  • 100 sensing nodes (84 end devices, 13 routers, 3 coordinators).
  • Linear and nonlinear loads (lamps, rotor rigs, microwaves, air-conditioners) to generate harmonic disturbances.
  • Electrical parameters (voltages, currents, frequency, harmonic amplitudes) measured at a sampling frequency of up to 14 kHz.
  • Data transmission via IEEE 802.15.4 protocol.

The experiments involve:

  • Non-stationary harmonic injection: Simulating harmonics from inverters in speed-varying rotor machines.
  • Transient harmonic injection: Simulating harmonics from switching large nonlinear loads (e.g., a light wave oven).
  • Frequency variance: Simulating power supply switches between public power and an uninterruptible power supply (UPS).

The data was collected from 100 channels and analyzed using 3 typical channels for visualization. The performance was compared against measurements from a Fluke 435 instrument.

Key Insights and Ideas

  • NARX Network's Adaptability: The NARX network's ability to learn complex, non-linear dynamics directly from data, without relying on predefined physical models, is a significant advantage for harmonic analysis, which is inherently dynamic and irregular.
  • Data-Awareness is Key: The "data-aware" aspect of their retrodiction method means it adapts to the actual observed harmonic behavior, making it more robust than model-based approaches that might struggle with unexpected system changes.
  • OOSM is a Real Problem: The experiments clearly demonstrate that OOSM significantly impacts harmonic measurement accuracy and identification, especially for higher-order harmonics and transient events.
  • Retrodiction Improves Accuracy: The proposed NARX-based retrodiction method consistently reduces measurement errors and improves harmonic identification accuracy across various scenarios.
  • Computational Trade-offs: While NARX-based methods can be computationally intensive due to neural network training, their performance is comparable to or better than particle filter methods, and they offer a stable real-time performance once trained.

Key Results

  • Reduced MSE: The NARX-based OOSM method significantly reduced the Mean Square Error (MSE) of harmonic measurements. For non-stationary harmonics, MSE was reduced by more than 33%. For transient harmonic events, MSE was reduced by more than 24%.
  • Improved Harmonic Identification: The method achieved better harmonic identification accuracy, particularly for the 3rd-9th odd harmonics, outperforming other OOSM algorithms.
  • Meets Standards: The NARX-based OOSM method, along with SEPF, met the harmonic measurement error limits specified by the IEC 61000-4-7 standard, whereas other methods did not.
  • Computational Complexity: The computational complexity of the NARX-based method was found to be stable with respect to delay but increased with input data size. It generally performed better or comparably to particle filter methods and was more stable than some Kalman filter-based approaches.
  • Memory Consumption: The NARX method had a higher memory consumption compared to Kalman filter-based methods but was considered reasonable given its ability to model complex dynamics.

Limitations and Future Work

  • Generalizability of Training Data: The paper notes that the NARX network's performance can be sensitive to the training dataset. While it performs well when trained on data similar to the test conditions, its flexibility across vastly different electrical environments needs further examination.
  • Dynamic System Instability: In real-world power networks, system topology and load types change dynamically, meaning the underlying harmonic dynamics might not be stable. While the NARX model can update itself, the convergence time to a new model might be too long for very rapid changes.
  • Transfer Learning: The authors suggest transfer learning as a future direction. Pre-training the NARX network on various load types and topologies and then deploying these pre-trained weights when the grid conditions change could significantly reduce adaptation time.
  • Larger Network Scales: For extremely large networks or highly complex power quality analysis tasks, the computational and memory costs of NARX might become a concern, requiring further investigation.
  • Hybrid and Unsychronized Networks: Future work will also involve examining the data-aware OOSM method in hybrid and unsynchronized network environments.

Summary updated

BibTeX
@inproceedings{liu2016data,
  author = {Youda Liu and Xue Wang and Yanchi Liu and Sujin Cui},
  title = {{Data-Aware Retrodiction for Asynchronous Harmonic Measurement in a Cyber-Physical Energy System}},
  booktitle = {Sensors},
  year = {2016},
  url = {https://openreview.net/forum?id=kWmLQtI1d0t},
}
Figure