GE Vernova researchers outline AI tools to speed interconnection studies and market clearing

Yang Guo, a senior engineer for electric power at GE Vernova Advanced Research, and a colleague, Ritam (presenter, GE Vernova Advanced Research), outlined a three-part research program to shorten distribution interconnection timelines, reduce day-ahead market inefficiencies driven by uncertainty, and accelerate market-clearing computations.

Guo said long, semi-manual interconnection studies can take “one to two months and can even go beyond,” and that the team’s goal is to reduce that turnaround to days through an automated, physics-aware workflow. Ritam described a four-stage pipeline that first uses machine learning to shortlist high-risk days, then validates shortlisted scenarios with a GridOS power-flow solver, applies staged mitigation, and produces interactive summaries for planners.

The presentation emphasized three mitigation options when an interconnection would otherwise be denied: first, redispatching existing distributed energy resources; second, deploying temporary relocatable energy storage at the interconnection site; and third, controlled load curtailment that allows a conditional connection with operational constraints. As Ritam put it, one bridging option is “relocatable energy storage at the interconnection site.”

On the screening step, Ritam described training supervised deep neural networks (including graph neural networks) on a year of P/Q forecasts labeled by AC power-flow runs, then clustering predicted violation hours to produce representative days for physics validation. He reported the model had a very high detection rate in the test dataset: “the true positive rate (TPR) … is 100%,” while noting the model produced about a 2–2.5% false-positive rate on that sample; the workflow keeps physics “in the loop” so those false positives are removed in later validation.

Addressing market-side uncertainty, Ritam said procuring extra reserves in the day-ahead market can reduce utilization and raise costs, and proposed a scenario-based, two-stage clearing approach in which slower generators are committed in stage 1 and faster units respond in stage 2. In a small 24-bus test case, he said stochastic unit commitment cut reserve costs in several scenarios by roughly 12–20% compared with a deterministic approach, while deterministic reserve procurement sometimes produced infeasible or inefficient market outcomes.

Yang Guo discussed methods to accelerate the computationally intensive unit-commitment problem: identifying and dropping nonbinding dynamic constraints, warm-starting solvers from similar historical cases, and using graph/convolutional neural networks to predict binary commitment statuses. He reported proof-of-concept results showing about a 50% solution-time reduction after removing certain constraints and overall 2–3× faster solves with warm starts. Guo also described an ongoing U.S. Department of Energy Grid Deployment Office-funded project to integrate distributed energy resource (DER) aggregators into wholesale markets; that work is in customer pilot testing and Guo said the team expects 2–5× speedups and the ability to handle significantly more market participants in the tested configurations.

The researchers framed the approach as a pipeline from research to product: identify industry pain points, build POCs, demo with real-world customer data, and then deploy features into commercial products. The presentation closed with an invitation for questions; none were recorded in the transcript.

Next steps the presenters identified include continued customer pilots and DOE-funded integration work; the team emphasized that ML screening is intended to prioritize cases for physics-based validation, not replace it.