Fast and Reliable N - k Contingency Screening with Input-Convex Neural Networks

Nicolas Christianson, Wenqi Cui, Steven H. Low, Weiwei Yang, Baosen Zhang

TL;DR

This paper introduces a novel machine learning approach using Input-Convex Neural Networks (ICNNs) to reliably screen power system contingencies, ensuring zero false negatives (i.e., never misclassifying an unsafe scenario as safe). By integrating a differentiable convex optimization layer into the training process, they achieve significant speedups (10-20x) over traditional methods while maintaining high accuracy, making power grid operations safer and more efficient.

What Problem Does This Paper Address?

Power systems are increasingly vulnerable to cascading failures caused by component outages (contingencies). Ensuring grid stability requires operators to identify and plan for these contingencies. However, checking all possible combinations of component failures, especially for higher-order "N-k" contingencies (where k > 1), becomes computationally intractable due to the exponential growth in possibilities. Existing methods either rely on heuristics that lack reliability guarantees or use exact methods that are too slow. This paper addresses the critical need for a method that is both fast and provably reliable for screening N-k contingencies.

Methods

The core innovation lies in using Input-Convex Neural Networks (ICNNs) for contingency screening. ICNNs are a class of neural networks that are guaranteed to be convex, which is crucial for this application. The authors propose a method to train ICNNs such that their predicted feasible region is a provably reliable inner approximation of the true feasible region of safe operating points.

Key methodological contributions include:

  • Certifying Reliability: A method to verify if an ICNN is reliable (zero false negatives) by solving a collection of convex optimization problems.
  • Enforcing Reliability: A technique to transform an unreliable ICNN into a reliable one by scaling and translating its parameters, achieved by solving optimization problems.
  • Differentiable Convex Optimization Layer: A novel training methodology that incorporates the reliability-enforcing scaling procedure directly into the ICNN training via a differentiable convex optimization layer. This allows the model to learn to be reliable from the ground up, avoiding the conservativeness of post-hoc scaling.
  • Fast Scaling Procedure: A simplified scaling procedure (Theorem 2) that requires only a single differentiable optimization step, improving computational efficiency during training.

Datasets

The paper uses the IEEE 39-bus test network for their case study. They generated 14,000 random demand vectors from a multivariate normal distribution to create a dataset for training, validation, and testing. The true feasible region was constructed by considering N-2 contingencies, excluding islanding and highly infrequent ones.

Key Insights and Ideas

  • ICNNs as Reliable Classifiers: The convexity property of ICNNs is fundamental to their ability to provide reliability guarantees. Their 0-sublevel set naturally forms a convex region, which can be rigorously checked for containment within the true feasible region.
  • Differentiable Training for Reliability: Integrating the reliability-enforcing scaling directly into the training loop via a differentiable layer is key. This allows the model to learn parameters that inherently satisfy the reliability constraint, rather than applying a fix afterward, which can lead to overly conservative solutions.
  • Trading Off Online vs. Offline Computation: The proposed method shifts the computational burden from real-time contingency screening to an offline training phase. This is a critical trade-off for ensuring safety in time-sensitive power system operations.
  • Inner Approximation for Efficiency: The ICNN learns an inner approximation of the feasible region. This is sufficient for guaranteeing reliability (no false negatives) and allows for faster screening because the ICNN only needs to check its own predicted region, which is computationally cheaper than checking all contingencies.

Key Results

  • Speedup: The ICNN-based approach achieves a significant speedup of 10-20x for contingency screening compared to the exhaustive method.
  • Zero False Negatives: The method guarantees a zero false negative rate, a critical requirement for reliable power system operation.
  • Low False Positive Rate: The ICNNs achieve a false positive rate between 2% and 5%.
  • Faster Preventive Dispatch: When applied to security-constrained optimal power flow (SC-OPF), the ICNN-based approach can halve the runtime (for deeper models) or speed it up by nearly 10x (for shallower models) with minimal increase in dispatch cost (average < 0.1%) and infeasibility (~1%).
  • Model Performance: The performance of the ICNNs is sensitive to hidden depth and positive class weights, with shallower models and lower positive weights generally showing better results for SC-OPF.

Limitations and Future Work

  • Scalability: The authors acknowledge that scaling the approach to larger power systems and higher-order contingencies remains an open challenge.
  • Combined Methods: Exploring combinations with other techniques, such as group testing, could potentially accelerate the full contingency analysis problem.
  • Broader Applications: The methodology of learning data-driven inner approximations to complex feasible sets using ICNNs could be applied to other energy systems and control problems, such as AC-OPF or electric vehicle flexibility.
  • End-to-End Training for SC-OPF: The authors suggest modifying the training methodology to directly incorporate the SC-OPF problem as a differentiable convex optimization layer for potentially improved performance.

Summary updated

Abstract

Power system operators must ensure that dispatch decisions remain feasible in case of grid outages or contingencies to prevent cascading failures and ensure reliable operation. However, checking the feasibility of all $N - k$ contingencies -- every possible simultaneous failure of $k$ grid components -- is computationally intractable for even small $k$, requiring system operators to resort to heuristic screening methods. Because of the increase in uncertainty and changes in system behaviors, heuristic lists might not include all relevant contingencies, generating false negatives in which unsafe scenarios are misclassified as safe. In this work, we propose to use input-convex neural networks (ICNNs) for contingency screening. We show that ICNN reliability can be determined by solving a convex optimization problem, and by scaling model weights using this problem as a differentiable optimization layer during training, we can learn an ICNN classifier that is both data-driven and has provably guaranteed reliability. Namely, our method can ensure a zero false negative rate. We empirically validate this methodology in a case study on the IEEE 39-bus test network, observing that it yields substantial (10-20x) speedups while having excellent classification accuracy.
BibTeX
@inproceedings{christianson2024fast,
  author = {Nicolas Christianson and Wenqi Cui and Steven H. Low and Weiwei Yang and Baosen Zhang},
  title = {{Fast and Reliable N - k Contingency Screening with Input-Convex Neural Networks}},
  booktitle = {L4DC},
  year = {2024},
  doi = {10.48550/arxiv.2410.00796},
}
Figure