- Problem
- Data center electricity demand in the United States nearly doubled between 2023 and 2026, and grid capacity has become the binding constraint on where new compute can be sited. Ranking which zones absorb that load safely means holding several moving quantities together at once: price volatility, congestion stress, existing data center concentration, and the queued generation that might relieve any of it.
- Approach
- Roughly a gigabyte of raw hourly market, load, infrastructure, interconnection-queue, industrial-demand, and geospatial data feeds a pipeline supporting several complementary methods. Panel fixed-effects regression isolates within-zone variation, XGBoost predicts locational marginal price volatility as a function of load, and event-study difference-in-differences tests whether observed shifts follow data center interconnections rather than merely coinciding with them. A composite score combines volatility, spike frequency, and data center exposure, less a discounted relief term for queued capacity.
- Result
- All 21 PJM zones ranked and re-ranked in real time as load growth and queued solar assumptions move, presented on an interactive map for readers who will not open a notebook. Because normalization anchors are fixed to the baseline period, a zone's score means the same thing across scenarios, which is what makes the comparison usable rather than merely responsive.
Why several methods rather than one
Each method answers a different objection. XGBoost captures nonlinear structure in the load-to-volatility relationship but says nothing about causality. Panel fixed effects removes stable zone characteristics, so the remaining variation is within-zone over time. The event study asks whether volatility shifts actually follow interconnection events rather than merely correlating with them.
Agreement across three methods with different assumptions is a stronger claim than a single well-fit model. Where they disagree, that disagreement is the finding.
Why volatility rather than price
Price level reflects fuel costs and demand in ways largely exogenous to siting. Volatility captures something closer to the question: how hard the zone is working to balance, and how much headroom remains before congestion shows up in prices. A zone can be expensive and stable, or cheap and fragile. Only the second is a siting risk.
The relief term
Queued renewable capacity is counted at a discount rather than face value. Interconnection queues clear slowly and incompletely, so treating a queued megawatt as equivalent to one in service would overstate relief. The relief term carries one sixth the weight of the vulnerability components, which keeps queued capacity meaningful without letting unbuilt plant dominate the ranking.
Fixed normalization anchors
Scoring anchors are computed once from the baseline period and held constant. Recomputing them per scenario would rescale every zone whenever one zone changed, making scenarios incomparable. Fixing them costs dynamic range at extreme inputs and buys the ability to say one scenario is worse than another, which is the entire purpose of the tool.
Limits
The model predicts volatility from historical relationships in PJM data, so it extrapolates rather than simulates. It does not model transmission topology, unit commitment, or the capacity market directly. It is a screening tool for narrowing attention across zones, not a substitute for a production cost model.
Analyses on this site are for general information only. They are not investment, financial, legal, or tax advice, and not an offer or solicitation to buy or sell any instrument. Past results do not indicate future outcomes. Views are the author's own.