Given the presence of rapidly evolving factors such as increased renewable energy integration and large, dynamic consumers like data centers, solving the optimal power flow problem on modern power grids is becoming increasingly more challenging and dynamic. Classical solvers for optimal power flow are reliable but computationally expensive, while machine learning-based approaches are fast but often fail to guarantee feasibility or generalize to new conditions. This project will focus on building upon ‘hybrid’ methods which combine these two approaches by investigating how existing classical solvers can be relaxed to become differentiable and improving neural network models for predicting warm starts.

I’m participating in SuperUROP because I’m interested in learning how I can combine my longstanding interest in applied machine learning with new skills I’ve gained through coursework in mathematics and optimization. I also hope to gain a better understanding of technical challenges related to working with power systems, a domain with high potential for societal impact.