GridPulse reduced 2025 holdout MAE from 3,302.6 MW for EIA to 1,842.7 MW on the same 8,690 valid hours.
PROJECT 01 / Energy Analytics · Forecasting · Live Operations
GridPulse Energy Grid Analytics
An operational energy analytics system built on EIA-930 PJM hourly grid data, combining forecast benchmarking, source-data QA, stress screening, error intelligence, and a live/replay Streamlit control room.
DATASETEIA-930 PJM hourly grid operations
SIZE35,064 hourly rows · 2022–2025
SOURCEU.S. Energy Information Administration · Form EIA-930
MEASURED EVIDENCE
What can be verified.
Metrics and outputs drawn from the project artifacts—not estimates added for presentation.
Peak MAE fell from 4,330.5 MW for EIA to 2,770.4 MW using one shared top-decile demand threshold.
The GridPulse candidate beat EIA on both overall and peak-demand MAE in every valid 30-day future fold.
All headline 2025 forecasts were scored on identical valid rows rather than different convenience subsets.
VALIDATION
The headline comparison is an out-of-time 2025 holdout using identical scoring rows and one shared peak threshold. Expanding-window 30-day rolling-origin validation produced 13 valid future folds, with the candidate passing the same EIA promotion gate in every fold.
PLAIN-LANGUAGE INTERPRETATION
GridPulse is a residual correction to EIA's reported day-ahead forecast, not a replacement weather-and-load forecasting stack. Its live control room uses an hourly operational feed and reports freshness explicitly rather than presenting EIA-930 as sub-second telemetry.
Why this project matters
Electricity demand can become difficult to forecast and serve when high load, rapid ramps, forecast misses, generation conditions, and interchange pressure coincide. Grid operators and analysts need a way to separate routine variation from periods that deserve closer operational attention without overstating what the data can prove.
What was developed
GridPulse ingests a frozen 2022–2025 PJM snapshot from eight official EIA-930 balance files, verifies source hashes, normalizes the hourly operating data, preserves QA anomalies, engineers lagged forecasting features, benchmarks EIA against a weekly baseline and a gradient-boosted residual correction model, runs rolling-origin validation, decomposes errors by operating slice, and exposes the results through a multipage Streamlit dashboard with frozen replay and live EIA API modes.
What it means
On 8,690 common 2025 holdout hours, the GridPulse ML-corrected EIA forecast reduced overall MAE by 44.2% and peak-demand MAE by 36.0% versus EIA's reported day-ahead forecast. The candidate also beat EIA on both overall and peak MAE in all 13 valid rolling-origin folds, while the project still documents where the model regresses or remains biased.
04 / WHERE THIS WORK APPLIES
From project to practical use.
The workflow can support load-forecast review, utility planning, balancing-authority analytics, peak-demand preparation, forecast-error diagnostics, source-data QA, grid operations research, and control-room decision support. The same pattern can be extended to additional balancing authorities once timing, data quality, and operating-context differences are validated.
05 / LIMITATION
What the project does not solve yet.
The verified result is specific to PJM and the frozen 2022–2025 EIA-930 snapshot. The model is not better in every local-hour slice: 08:00 PJM local shows a 0.8% MAE regression. At the highest-demand decile, average peak misses are smaller but the remaining misses are more frequently underforecasts. The stress score is an analytical screen, not an official reliability rating or blackout probability.
06 / NEXT ITERATION
Future advancement
Replicate the benchmark across additional balancing authorities, tighten feature-availability checks against an explicit forecast-issuance clock, investigate the 08:00 local regression, and test peak-focused calibration or asymmetric-loss approaches for the highest-demand hours.
TECHNICAL TOOLKIT
PythonStreamlitPandasScikit-learnEIA-930Gradient BoostingTime-series Validation