Explainable AI for Electric Power Systems - Paid Master's Thesis

At Eneryield we help energy companies predict power outages with AI systems that identify developing faults. Many faults in electric power systems don't appear out of nowhere. They develop over time and often leave early signs in the electrical measurements. The goal is to allow for earlier intervention, help improve reliability and support the continued integration of renewable energy. We work at the intersection of machine learning, software engineering and electric power systems, using data collected from real industrial sites worldwide. Our systems work with both waveform data and derived scalar features, and because our applications are used in practice, we place a strong emphasis on AI that is explainable and understandable. We are now looking for Master’s students to work on several related research projects. The data and research problem The projects use measurements from operational electric power systems, collected over several years and from multiple locations. The data is recorded irregularly and often arrives in bursts, rather than as a regularly sampled time series. The central research question is how pre-fault behavior shows up in these measurements. This includes distinguishing normal operation from fault patterns and understanding how this behavior evolves over time. Because our systems are used in an industrial setting, explainability and reliability are essential. Predictive performance is important, but it is not the only measure of success. Projects may also evaluate robustness, stability, interpretability, generalization and the quality of knowledge discovered from the data. Three main project topics

1. Uncertainty Quantification and Selective Prediction

A fault warning is only useful if operators know when to trust it. This project explores how to quantify the uncertainty behind each prediction, so the system can tell when it is on solid ground and when it should qualify or hold back a warning. You may work with methods such as calibration, ensemble uncertainty, conformal prediction and detection of unfamiliar inputs, or explore explainable models with uncertainty built into the prediction itself.

1. Robust Temporal Modeling under Irregular and Variable Sampling

How can a model learn from how system behaviour evolves over time, without simply learning each customer’s recording habits? This project explores temporal point processes and related continuous-time models for irregularly sampled data, either to learn representations of recent system behaviour or to produce a pre-fault risk signal directly. By simulating different sampling regimes, you will test whether models stay robust to changes in the recording process while still responding to real changes in the system.

1. Learning Human-Readable Rules for Developing Faults

Can machine learning automatically discover a compact set of human-readable rules for identifying developing faults? This project explores fuzzy, neuro-fuzzy and neuro-symbolic approaches for learning such rules directly from power-system data. You will investigate how measurements and temporal trends should be expressed as concepts in the rules, and evaluate the results on predictive performance as well as compactness, stability and coverage. What we offer

• A paid Master’s thesis with competitive compensation
• The chance to apply your academic knowledge to field-recorded industrial data
• Mentorship from Eneryield engineers and researchers
• The opportunity to develop methods that may contribute directly to our future systems

Requirements

• Enrolled in a Master's program in Computer Science, Engineering Physics, Mathematics, Statistics, Electrical Engineering or similar
• A genuine interest in deep analysis and problem solving with machine learning and AI; interest in energy and electric power systems is a merit
• Strong Python skills, comfort with the scientific stack (NumPy, pandas, scikit-learn), PyTorch and/or Tensorflow.
• Experience working with Git and GitHub
• Good written and spoken English

How to apply Please send your CV, transcripts and a brief cover letter about your background and interests to isabel@eneryield.com before November 15th. You are welcome to indicate which theme interests you the most, although a specific project does not have to be selected in your application. Each theme also has smaller extensions and focus areas we can tailor to your interests, and we are open to well-motivated ideas of your own. You are welcome to apply either individually or in pairs. If you are applying together, please submit one joint application. We review applications continuously, so we encourage you to apply as soon as possible. We look forward to hearing from you!