Instability Prediction in Power Systems using Recurrent Neural Networks

Ankita Gupta, Gurunath Gurrala, Pidaparthy S. SastryIJCAI 2017 · DOI

TL;DR

This paper introduces an Online Monitoring System (OMS) using a stacked Gated Recurrent Unit (GRU) based Recurrent Neural Network (RNN) for early prediction of transient instability in power systems. Unlike previous methods that require precise fault information, this GRU-based RNN continuously monitors system variables and predicts instability without needing fault clearance times, demonstrating robustness to noise and topology changes.

What Problem Does This Paper Address?

Power systems are susceptible to disturbances like faults, which can lead to transient instability and potentially blackouts. Early detection of this instability is crucial for operators to take timely preventive actions. Traditional methods for assessing transient stability often rely on detailed simulations or require accurate, real-time information about fault occurrences and clearance times. This fault information is difficult to obtain in practice due to limitations in instrumentation and potential errors in relay tripping times. Existing machine learning approaches, such as Support Vector Machines (SVMs), also struggle when faced with even slight inaccuracies in fault clearance time or measurement noise, making them brittle for real-world applications. The paper addresses this gap by proposing a system that can predict instability without needing precise fault information, making it more practical and robust.

Methods

The core of the proposed solution is a stacked Gated Recurrent Unit (GRU) based Recurrent Neural Network (RNN). GRUs are a type of RNN designed to better capture long-range temporal dependencies in data, which is essential for analyzing time-series data from power systems. The system operates as an Online Monitoring System (OMS), continuously analyzing a sliding window of past measurements.

Instead of a simple binary classification (stable/unstable), the RNN is trained to classify the system state into five categories: No Disturbance (Class 0), Fault Occurrence (Class 1), Fault Duration (Class 2), Fault Clearance (Class 3), and Alert (Class 4 - indicating impending instability). This multi-class approach helps the RNN learn the nuances of system behavior during disturbances.

For identifying the critical generator that might lead to instability, the authors employ Principal Component Analysis (PCA). They project the data (specifically, generator voltage measurements during the predicted instability window) onto the first two principal components. A robust Gaussian density is then fitted to these projected data points, and generators with the minimum likelihood under this density are identified as outliers and thus critical.

Datasets

The authors utilize a simulation of the IEEE 118-Bus system for generating training and testing data. This system is a standard benchmark for power system stability research, comprising 118 buses, 19 generating units, 91 loads, and 177 transmission lines.

Data is generated through offline dynamic simulations using MATLAB. The simulations involve imposing three-phase-to-ground faults on each bus and transmission line at various locations and with randomly selected clearing times (4-8 cycles). These scenarios are repeated under different loading levels. The simulator records variations in generator voltages (magnitude, phase angle, and derivative of phase angle) over a 1000-cycle time window.

Key Insights and Ideas

  • Fault-Information-Free Instability Prediction: The most significant insight is the development of a system that can predict instability without requiring precise fault clearance times. This overcomes a major practical limitation of existing methods.
  • GRU's Efficacy for Temporal Dependencies: The paper highlights the effectiveness of GRUs in capturing the complex temporal dynamics of power system variables, which is crucial for early instability detection.
  • Multi-Class Classification for Robustness: Training the RNN for a 5-class classification problem (instead of binary) allows it to better distinguish between normal variations and those indicative of impending instability.
  • PCA for Critical Generator Identification: The use of PCA to reduce the dimensionality of generator data and identify outliers provides a novel way to pinpoint critical generators, aiding operators in targeted control actions.
  • Robustness to Real-World Imperfections: The proposed OMS demonstrates remarkable resilience to measurement noise and changes in network topology, making it suitable for practical deployment.

Key Results

  • The proposed GRU-based RNN OMS achieves high accuracy in predicting transient instability.
  • The system demonstrates zero missed detection rates even under significant measurement noise (up to 3% uniform noise).
  • The false alarm rate increases slightly with noise but remains manageable.
  • The OMS is shown to be robust to network topology changes, performing well even when transmission lines are disconnected.
  • The critical generator identification method using PCA is effective in identifying outliers in approximately 85% of unstable cases, even with measurement noise.
  • The OMS predicts instability typically within 4-7 cycles from the fault clearance time (tc), providing valuable lead time for control actions.

Limitations and Future Work

  • Centralized Measurement Assumption: The current OMS assumes that measurements from all generators are available at a central location synchronously.
  • Future Work: The authors suggest extending the current system into a distributed system as an interesting problem for future research. This would involve handling decentralized measurements and communication delays.

Summary updated

BibTeX
@inproceedings{gupta2017instability,
  author = {Ankita Gupta and Gurunath Gurrala and Pidaparthy S. Sastry},
  title = {{Instability Prediction in Power Systems using Recurrent Neural Networks}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2017},
  doi = {10.24963/ijcai.2017/249},
}
Figure