NoVA Data Center Water Forecasting
A companion piece to my paper on the Digital Gateway data center fight: instead of arguing that water use is a governance blind spot, this forecasts it. Annual water demand and cost for Loudoun and Prince William counties, 2000-2030, from three models validated by walk-forward backtesting. The point forecast clears its accuracy bar. The 90% prediction interval doesn't, and every number in the output says so.
The target metric was wrong before any model ran
The obvious data source, USGS public-supply withdrawal, gave only 4 real years per county and, for Prince William, folded in Fairfax Water's regional Occoquan withdrawal at roughly 2.5x actual county deliveries. I replaced it with water actually delivered, hand-transcribed from Loudoun Water and Prince William Water's Annual Comprehensive Financial Reports, which took real annual coverage from 4 of 25 rows to 34 of 50. fetch_data.py --verify-acfr downloads each source PDF and asserts every transcribed figure appears in it verbatim, so a typo in the manifest fails loudly instead of silently.
Switching from quantile regression to SARIMAX made interval coverage worse, not better.14-20 real annual points per county isn't enough to calibrate a 90% interval, and no model choice fixes a sample-size problem.
Converged doesn't mean identified
SARIMAX reported clean convergence on every county, no warning, no error. The exogenous coefficients, capacity, temperature, rate, told a different story: standard errors at or near zero and nan, meaning the likelihood had no real curvature in that block. The same small-sample signature shows up in the quantile models, where the coefficient on capacity flips sign between the 5th and 95th percentile fits, and the interval itself crosses its own bounds (Q05 > Q95) on 11 of 22 folds. Because the exogenous terms don't identify, the 2024-2030 forecast runs as an exog-free ARIMA with no drift. It comes out near-flat, likely an underestimate against 2-4%/yr historical growth, but a flat number I can stand behind beats a confident one I can't.
A risk flag that only half-works
A county-year is flagged high-risk at or above the 75th percentile of that county's real history. Across 22 backtest folds that scores 0.78 precision and 0.78 recall, which reads fine until you split it: Loudoun trips the flag in every scenario and every forecast year, so it never actually discriminates there. The flag only does useful work for Prince William's efficiency scenario. The report says that directly instead of letting an aggregate 0.78 stand in for a metric that's silently useless for half its counties.