Trustworthy Machine Learning for Power System Applications

Rahul Nellikkath*

*Corresponding author for this work

Research output: Book/ReportPh.D. thesis

21 Downloads (Pure)

Abstract

The rapidly increasing shares of renewables in our electricity system is triggering an unprecedented change in the power grid static and dynamic behavior. Unlike the traditional power systems of the past decades, where disturbances propagated comparatively slower due to the inertia from the conventional generators, converter-based sources such as wind turbines, solar PVs, and electric vehicles introduce much faster dynamics and complex control architectures. Existing simulation tools struggle to keep up with the modeling and compuation needs that these changes bring about, highlighting the urgent need ot develop more efficient methods to screen thousands or millions of scenarios rapidly. Besides power system dynamic studies, the current transition to higher shares of renewables and converter-interfaced resources makes power system optimization, as well, orders of magnitude more complex.
Machine learning methods have become increasingly popular for accurately estimating the stability of power systems and approximating their optimal operation, offering the ability to evaluate up to 1000 scenarios in the time it takes to assess a single scenario using traditional methods. However, when it comes to power system applications, machine learning algorithms face significant challenges: first, they are dependent on the availability of high-quality datasets, which are often difficult to obtain. Second, they are considered a black box, and their (lack of) trustworthiness poses a major barrier for their adoption in safety-critical systems such as power systems. Considering that data-driven approaches require equal share of normal and abnormal operating points to accurately estimate the security of power systems, historical data – where abnormal situations appear rarely in power systems – are not enough. On the other hand, generating suitable datasets through simulations, especially for large power systems, is often impractical, even with many of the recent techniques proposed in the literature. In contrast, Physics-informed neural networks (PINNs) are able to address this challenge by incorporating the underlying physical models inside the training, thus reducing the need for external datasets and accelerating the overall training process. Current PINN models for power systems are limited to very few applications, mainly focusing on traditional synchronous generators. This thesis significantly advances the state of the art for PINNs in power systems, by introducing first-of-their-kind methods both for the static and the dynamic operation of power systems.
This thesis is the first that proposes Physics-Informed Neural Networks for both DC-OPF and AC-OPF problems, integrating the Karush-Kuhn-Tucker conditions in the NN training to ensure that the model generates solutions adhering to both the optimality conditions and the physical constraints of the power system. Equally importantly, this thesis explores the application of PINNs in power system time-domain simulations, generating first-of-a-kind results in approximating the non-linear dynamics of a renewable resource equipped with an inverter controller. We demonstrate that our PINN was 100 times faster than conventional simulation approaches; it is worth highlighting that in this time comparison we also include the computation time for the dataset generation and the PINN training; and, still, PINNs are 100 times faster than conventional methods. Our results showcase the significant potential PINNs have for very fast approximations of the power system dynamic behavior and their ability to perform rapid screening of thousands of critical disturbances.
The second challenge of machine learning methods applied in power systems involves the black-box nature of machine learning algorithms, which makes it difficult to trust in safety-
critical applications, such as power systems. Removing the barriers for a widespread acceptance of Neural Network (NN) algorithms in power systems demands a highly accurate approximation of the nonlinear processes and guarantees that NNs will not compromise system constraints.
Recently, a method has been introduced to extract the worst-case performance guarantees for linear DC-OPF problems via Mixed Integer Linear Programming. Although this is a significant step towards ensuring that the NN output does not violate any system-critical constraints, there is still significant work to be done. This thesis builds upon and significantly expands this work in several directions. First, it introduces algorithms that determine the worst-case violations not only for DC-OPF problems, which are linear, but also for the significantly more challenging (and more relevant for the power system industry) AC-OPF problems, which are nonlinear and non-convex. Second and most importantly, this thesis introduces methods that not only determine but also reduce and potentially eliminate these violations; this is a significant new challenge. This thesis takes upon this task by (i) designing a new neural network training procedure that reduces worst-case violations during the neural network training, (ii) efficient resampling around poorly performing data regions across the entire input domain, and (iii) by scalable algorithms employing advanced bound tightening and gradient-based methods.
Original languageEnglish
Place of PublicationKgs. Lyngby, Denmark
PublisherDTU Wind and Energy Systems
Number of pages102
Publication statusPublished - 2024

Fingerprint

Dive into the research topics of 'Trustworthy Machine Learning for Power System Applications'. Together they form a unique fingerprint.

Cite this