Machine Learning Group

Probabilistic Impact Scenarios for Extreme Weather Event Resilience

Power system operators struggle to plan for events like extreme weather, despite forecasts. Uncertainties in timing, location, and grid vulnerabilities complicate planning. Stochastic optimization algorithms enhance system redispatch in the face of unknown threats. Yet, improved methods are required for realistic probabilistic scenario creation, essential for these algorithms. Leveraging modern statistical, artificial intelligence (AI), and machine learning (ML) techniques can aid in scenario generation, but uncertain behavior remains a challenge.

This multi-lab collaboration project tackles the challenge of proactive grid management in the face of unknown threats. Goals include developing statistical sampling methods for extreme temperature and wind scenarios, using AI/ML for forecast error generation, and creating a visualization for model comparison and explanation. CSI is leading ongoing research using unsupervised architectures, such as generative adversarial networks (GANs), to generate power scenarios without manual labeling. Central to this effort is the expansion of SeqGAN, a recent GANs model, tailored for power generation. By seamlessly merging GANs with reinforcement learning, this model can grasp temporal correlation.

The ultimate goal for this effort is to craft seamless power scenarios, intricately woven without dependence on prior knowledge or intricate fine-tuning.