Arimancy/pjm-demand-weather

PJM Hourly Electricity Demand + Population-Weighted Weather (2015 to 2026)

A model ready hourly time series for forecasting load on the largest US grid: PJM RTO demand (MWh) pre joined to population weighted weather, with calendar, holiday, degree hour, irradiance, cloud cover, apparent temperature, and snowfall features. 96,865 hours from 2015-07-01 to the most recent build, no missing hours on the UTC spine, every imputed, interpolated, or preliminary value explicitly flagged, in Parquet and CSV with an easy chronological split. The weather is genuine ERA5 (the model is pinned, not the archive default), and every build ships a JSON manifest recording the exact upstream sources.

The monthly chart above is the dataset's signature. PJM is the most dual peaked grid in this series: January and July both rise well above the spring and fall shoulders in every complete year (January runs 1.17 to 1.33 times the shoulder months, July 1.24 to 1.47), because both electric heating and air conditioning bind across a 13 state footprint. Winter is when the system is stressed hardest: Winter Storm Elliott (December 2022) and the January 2025 cold snap are both inside the data range, and both stand visibly above their adjacent weeks while the weighted apparent temperature collapses to -22.5 and -20.0 C. The year gradient also shows the newest story: data center growth is pushing demand toward all time records, with mean July demand up about 10 percent from 2024 to 2026. Weather explains the swings; that trend is load growth, and the dataset says so plainly.

Why this dataset

The hard part of load forecasting is not finding demand data, it is the join. This dataset does the error prone work correctly and documents it:

• UTC canonical join. Demand and weather are joined strictly on UTC, which has no DST pathologies. Local time (America/New_York, Eastern Prevailing Time per PJM market convention) ships as a derived column; spring forward days have 23 local hours and fall back days 25, flagged via is_dst_transition_day rather than "fixed."
• Population weighted weather, ten points. ERA5 series at Chicago, Newark, Washington DC, Philadelphia, Baltimore, Pittsburgh, Cincinnati, Cleveland, Columbus, and Norfolk, weighted by Census Vintage 2025 populations (0.2135, 0.1845, 0.1464, 0.1433, 0.0647, 0.0548, 0.0524, 0.0490, 0.0507, 0.0407). Ten points because PJM spans continental Midwest, Appalachian, Mid Atlantic coastal, and Southeast coastal climates; the points cover roughly two thirds of the served population.
• Winter columns. Population weighted apparent_temp_pop_wt_c (deg C) and snowfall_pop_wt_cm (cm per hour), weighted with the identical code as temperature. In the baseline notebook's controlled ablation, apparent temperature is a small but robust incremental gain beyond core weather (about 0.06 MAPE points, consistent across seeds): wind chill genuinely drives heating load beyond what dry bulb temperature shows. Snowfall does not measurably improve mean accuracy and the dataset makes no accuracy claim for it; it is provided for storm window and peak risk marking, a feature for downstream use.
• Solar columns. Population weighted shortwave_radiation_pop_wt_wm2 and cloud_cover_pop_wt_pct, the series standard introduced on CAISO, carrying the summer side.
• Measured value of the weather join. In the baseline notebook's ablation on the identical evaluation frame, removing all weather raises LightGBM MAPE by about 0.8 percentage points, from about 3.4 percent to 4.23 percent (calendar plus demand lags only), roughly a fifth of the error. That is the value of actual observed ERA5 weather; a live deployment using weather forecasts would see less. The full model runs about 38.7 percent of the seasonal naive error (8.79 percent), below the project's 40 to 65 leakage tripwire band; the notebook documents why that is weather value and not leakage (the no weather control sits in band at 48.1 percent, every demand feature is at least 24 hours old, and weather is exogenous time t actuals).
• Provenance you can filter on. The long backbone is Catalyst Cooperative's PUDL cleaned EIA-930 (anomalies imputed, with reason codes); the freshest roughly 2 months come straight from the EIA API and are flagged demand_source="eia_api" (preliminary, unimputed). Where the two overlap, PUDL's cleaned data wins, with one exception stated plainly: a real API report beats a PUDL placeholder. Hours the BA had not yet reported carry PUDL's imputation code missing_value with no reported value; where the API has the actual report for such an hour, the report ships, flagged eia_api. Weather gap handling only ever interpolates interior gaps of 3 hours or less, flagged. The weather side carries the same honesty: weather_model and weather_source record the pinned ERA5 model and the Open-Meteo archive, and weather_is_preliminary with weather_final_through mark the ERA5 reanalysis-lag tail (the last week or so, which ERA5 has not settled yet) so you can drop or keep it.

Footprint

demand_mwh is RTO level total demand: the whole PJM footprint of 13 states plus DC, about 65 million people, the largest electricity market in the United States. It is not a zonal series; the zonal breakdown (AEP, ComEd, Dominion, PSEG and the rest) is a candidate v2 layer. The ten weather points cover roughly two thirds of the served population; West Virginia, Kentucky, Delaware outside the Philadelphia metro, and rural load are not separately represented. Regional weather, not exact service territory population, is the demand driver being modeled. Two deliberate weighting choices are documented in the pipeline config: the Newark weight uses New Jersey's population minus its four Philadelphia CBSA counties (the full New York CBSA would cross into NYISO), and the Chicago CBSA keeps its MISO served northwest Indiana slice per series convention.

Sources

Data	Source	License
Demand backbone	EIA-930 via Catalyst Cooperative PUDL (out_eia930__hourly_operations)	US public domain; PUDL processing credited below
Demand fresh tail	EIA Open Data API v2, electricity/rto/region-data, respondent PJM	US public domain
Weather	Open-Meteo Historical Weather API (ERA5 reanalysis)	CC BY 4.0
Population weights	US Census Bureau CBSA estimates, Vintage 2025	US public domain

Methodology summary

Hourly PJM demand (hour ending UTC) from PUDL's cleaned parquet, topped up past PUDL's lag with the EIA API, deduplicated preferring PUDL wherever PUDL carries cleaned real data (hours PJM had not yet reported, which PUDL extrapolates under imputation code missing_value, ship the real API report instead). ERA5 2 m temperature, shortwave radiation, cloud cover, apparent temperature, and snowfall pulled at each of the ten grid points in GMT with the Open-Meteo model pinned to era5 (never the Best Match default), each variable combined as X(t) equals the weight sum of the per point series with fixed Vintage 2025 population weights (a weighted value is NaN unless all ten points report). Left join on the UTC hour, reindexed to the complete hourly range, gaps flagged, short interior gaps (3 hours or less) linearly interpolated and flagged. Calendar features computed from local wall clock time; US federal holidays ship as a statutory and observed pair (is_holiday_statutory, is_holiday_observed, the latter also marking the weekend shifted observed weekdays; no single state set fits a 13 state footprint); degree features are hourly degree hours and daily degree days, base 65 F, from the weighted temperature, with degree_day_is_partial flagging the dataset's two partial boundary days. Full column by column reference: see the data dictionary.

Files

• pjm_demand_weather_hourly_v1.parquet: recommended; typed, compressed, timezones survive round trip.
• pjm_demand_weather_hourly_v1.csv: universal access; timestamps are ISO 8601 with explicit offset.

Known limitations

• The most recent stretch of demand is preliminary EIA data (demand_source="eia_api", unimputed) until PUDL coverage overtakes it, about 2 to 3 weeks against recent nightly PUDL objects. The weather model is pinned to ERA5, which trails real time by about a week, so the trailing ERA5-lag window has no weather: those hours ship with null weather and are flagged weather_is_preliminary and temp_is_missing while their demand values are kept, never trimmed. weather_final_through marks the last settled ERA5 hour.
• The demand backbone reads PUDL's experimental nightly build, which is subject to upstream revision between builds. Each build's JSON manifest records the exact PUDL object fingerprint, so the backbone behind any release is reconstructable.
• ERA5 is a reanalysis (model) estimate, not station observations; apparent temperature is itself a derived quantity, one more step removed.
• RTO total demand, not zonal (see Footprint above).
• Population weights are static (Vintage 2025).
• Data center growth is a structural trend the weather features do not explain; models leaning on recent demand lags absorb it.
• A backtest on this data uses actual weather; do not compare its error head to head with PJM's live day ahead forecast error, which uses live weather forecasts and analyst correction.

License and attribution

This dataset is published by Arimancy LLC under CC BY 4.0. The full license text and the attribution target also ship in the package as LICENSE.txt.

• EIA-930 demand data: U.S. Energy Information Administration, Form EIA-930 / Hourly Electric Grid Monitor (US public domain). Source: U.S. Energy Information Administration (Jul 2026).
• PUDL: cleaned and imputed EIA-930 demand via Catalyst Cooperative's Public Utility Data Liberation (PUDL) project. PUDL's data and documentation are published under CC-BY-4.0.
• Weather data by Open-Meteo.com (CC BY 4.0). ERA5: Hersbach et al. (2023), ECMWF, doi:10.24381/cds.adbb2d47.
• Population estimates: U.S. Census Bureau (US public domain).

Built and maintained by Arimancy (Arimancy LLC). Dataset page: arimancy.com/datasets/pjm-hourly-demand-weather. Also available on Kaggle and Hugging Face. Column-by-column reference: the included data_dictionary.md.