Time-aware evaluation
The last 48 hours are held out; the model is fit only on preceding observations instead of using a random split that leaks future information.
02 / MACHINE LEARNING / TIME SERIES
GridCast.
A time-series forecasting experiment using historical U.S. electricity demand, with a comparison against a simple baseline.
PROJECTElectricity demand forecastingINCLUDEDWeb app, source code and tests
APPLICATION CAPTURE / LOCAL TESTINGREAL OBSERVATIONS336 hours
HOLDOUT48 hours
BEST TEST MAE667 MW*
01 / THE PROBLEM
The problem
A forecast chart can look convincing even when the model does not beat a simple alternative. This project tests performance on observations held out from training.
What I built
GridCast uses hourly demand and lag features, keeps the final 48 hours out of training, and compares ridge regression with a forecast based on demand at the same hour the previous day.
02 / FEATURES
What the app can do
01
Explore 336 consecutive historical ERCOT hourly demand observations
02
Fit browser-side ridge regression using lagged demand and calendar features
03
Compare MAE against a 24-hour seasonal-naive baseline on held-out observations
04
Adjust forecast scenarios and inspect future 24-hour estimates
05
Import another valid hourly CSV or load supported public historical archive data
03 / IMPLEMENTATION
Implementation choices
Baselines before claims
The naive forecast is evaluated alongside ridge regression. In the included reference experiment, the simple baseline performs better.
Explicit uncertainty
The displayed uncertainty envelope is a heuristic derived from validation error, not a calibrated probabilistic interval.
04 / TESTING & LIMITATIONS
Testing and limitations
Test evidence
336 consecutive ERCOT demand observations from July 2–15, 2015. Independent Python backtest: ridge MAE 823.37 MW; 24-hour seasonal baseline MAE 667.42 MW on the 48-hour holdout. *667 MW is the baseline, not the trained model.
Limitations
Fourteen days of 2015 data are insufficient to establish robust modern grid forecasting. The reference report uses observed lags for one-step-ahead evaluation; the forward 24-hour projection is recursive. No real-time operational claims.
Technical stack
Pythonscikit-learnRidge regressionJavaScriptEIA Form-930
05 / DATA SOURCES