xpertsystems/enr003-sample

ENR003 — Synthetic Electricity Demand & Load Forecasting Dataset (Sample Preview)

XpertSystems.ai | Synthetic Data Factory | Energy & Climate Vertical

A single-table, load-research-calibrated electricity demand dataset spanning 8 utility demand zones across diverse IECC climate zones (Hot-Humid to Very Cold), with 15-minute interval resolution. Each row joins zone-level load composition, weather, day-ahead / hour-ahead / week-ahead probabilistic forecasts, peak event flags, ERCOT 4CP detection, demand response activations, TOU pricing tiers, LMPs, and behind-the-meter DER (rooftop solar, BESS, EVs, V2G). Calibrated benchmark-first against EPRI PRISM load research, PJM Load Forecast Archive, EIA-861, FERC Electric Power Reports, and DOE EV Charging Study 2023.

This is the sample preview — 8 zones × 1 week × 15-min cadence (5,376 rows × 99 columns). The full product covers 50 zones × full annual cycle (1.75M rows) with multi-season seasonal load factors, 1-in-50-year peak exceedance modeling, and N-1 grid stress scenarios.

---

Dataset summary

Property	Value
Rows	5,376
Columns	99
Cadence	15-minute
Time span	1 week (2022-01-01 → 2022-01-08)
Zones	8 utility demand zones
Climate zones	1A / 3B / 4A / 5A (IECC)
Load categories	Residential / Commercial / Industrial / Agricultural / EV
Forecast horizons	1h ahead, 24h (day-ahead), 168h (week-ahead)
File formats	Parquet (preferred) + CSV

The 99 columns are grouped into eight blocks that join on zone_id × timestamp_utc: identifiers, load profile, seasonal/calendar, load curve, peak event, forecast (multi-horizon), weather, market, and EV/DER.

---

Calibration sources

All ten validation metrics target named industry sources, not generator self-metrics:

• PJM Load Forecast Archive — published MAPE benchmarks at 1h / 24h / 168h forecast horizons (1.8% / 3.2% / 5.1%)
• EPRI PRISM Load Research — daily peak-to-valley ratios and load factors for residential / commercial / industrial / EV / agricultural load shapes
• EIA-861 — utility-level customer counts and seasonal load factor amplification
• FERC Electric Power Reports — temperature sensitivity (MW per °C per 1000 customers) by load category
• DOE EV Charging Study 2023 — L2 / DCFC charging session distributions, V2G adoption rates
• NREL TR-65-72701 / IEA Wind Task 36 — probabilistic forecast interval coverage benchmarks
• ASHRAE 55-2020 — comfort temperature thresholds and humidity ranges
• IECC Climate Zones — geographic temperature and degree-day modeling

---

Validation scorecard (seed = 42)

10/10 PASS · Grade A+ (100%) across all six canonical seeds (42, 7, 123, 2024, 99, 1).

#	Metric	Observed	Target	Tol	Type	Source
1	day_ahead_forecast_mape_pct	3.24	3.2	±1.5	two-sided	PJM 24h ahead
2	hour_ahead_forecast_mape_pct	1.81	1.8	±1.0	two-sided	PJM 1h ahead
3	week_ahead_forecast_mape_pct	5.08	5.1	±2.5	two-sided	PJM 168h ahead
4	firm_plus_interruptible_equals_peak_rate	1.000	0.995	±0.005	FLOOR	Structural
5	weather_demand_abs_correlation_strongest_zone	0.746	0.25	±0.15	FLOOR	FERC sensitivities
6	peak_to_valley_ratio_end_of_day_mean	2.17	2.0	±1.0	two-sided	EPRI PRISM
7	load_factor_end_of_day_mean	0.684	0.70	±0.15	two-sided	EPRI PRISM
8	naive_persistence_model_share	0.0487	0.05	±0.03	two-sided	Generator model mix
9	humidity_pct_mean	50.14	50.0	±5.0	two-sided	ASHRAE 55 + Beta(3,3)
10	p10_p90_interval_coverage_pct	100.0	90.0	±10.0	FLOOR	NREL / IEA T36

---

Schema highlights (99 columns)

Identifiers (3): zone_id, timestamp_utc, climate_zone (IECC 1A / 2A / 3A / 3B / 3C / 4A / 5A / 6A).

Load profile (10): total_demand_MW, residential_demand_MW, commercial_demand_MW, industrial_demand_MW, agricultural_demand_MW, ev_charging_demand_MW, net_load_MW, load_density_MW_per_km2, customer_count, avg_consumption_kWh_per_customer.

Seasonal & calendar (12): season, month, day_of_week, hour_of_day, is_holiday, heating_degree_days, cooling_degree_days, seasonal_load_factor, summer_peak_flag, winter_peak_flag, shoulder_period_flag, load_shape_type (Residential_Weekday / Residential_Weekend / Commercial_Weekday / Commercial_Weekend / Industrial_Flat / EV_TOU).

Load curve (12): daily_peak_MW, daily_valley_MW, peak_to_valley_ratio, morning_ramp_MW_per_hour, evening_ramp_MW_per_hour, load_factor, coincident_peak_flag, non_coincident_peak_MW, base_load_MW, flexible_load_MW, ramp_event_flag, duck_curve_depth_MW.

Peak event (14): peak_demand_event_id, peak_event_type (None / Summer_Peak / Winter_Peak / Shoulder_Spike / Weather_Extreme), peak_magnitude_MW, peak_duration_minutes, peak_probability_exceedance, firm_peak_MW, interruptible_peak_MW, demand_response_activation_flag, demand_response_MW_curtailed, peak_temp_C, peak_humidity_pct, heat_index_C, wind_chill_C, ercot_4cp_flag.

Forecast — multi-horizon (16): forecast_horizon_hours, forecast_target_timestamp_utc, forecast_issued_at_utc, forecast_demand_MW, forecast_p10_MW, forecast_p50_MW, forecast_p90_MW, forecast_error_MW, forecast_mape_pct, weather_forecast_temperature_C, weather_forecast_error_C, model_type (LSTM / XGBoost / ARIMA / Prophet / Ensemble / Naive_Persistence), feature_set_version, fc1h_demand_MW, fc1h_mape_pct, fc168h_demand_MW, fc168h_mape_pct.

Weather (11): temperature_C, temperature_normal_C, temperature_anomaly_C, humidity_pct, dew_point_C, solar_irradiance_W_per_m2, wind_speed_m_per_s, cloud_cover_pct, precipitation_mm, extreme_weather_flag, urban_heat_island_C.

Market (10): real_time_lmp_per_MWh, day_ahead_price_per_MWh, tou_rate_tier (Super-Off-Peak / Off-Peak / Mid-Peak / On-Peak), tou_rate_per_kWh, demand_charge_per_kW, price_elasticity_demand, demand_response_incentive_per_kWh, energy_cost_forecast_per_MWh, carbon_price_per_tonne, renewable_energy_credit_per_MWh.

EV & DER (10): ev_penetration_pct, ev_charging_sessions_count, ev_l2_charging_MW, ev_dcfc_charging_MW, v2g_discharge_MW, smart_charging_active_flag, rooftop_solar_generation_MW, battery_storage_dispatch_MW, virtual_power_plant_flag, der_capacity_MW.

---

Suggested use cases

• Day-ahead load forecasting models — train LSTM/XGBoost regressors for total_demand_MW 24h ahead using weather forecast + calendar + historical load features. Benchmark against the included PJM-calibrated forecast columns
• Probabilistic forecasting — evaluate P10/P50/P90 interval quality on multi-horizon forecasts (1h / 24h / 168h) with the included forecast_p*_MW columns
• Peak demand prediction — classifier for summer_peak_flag, winter_peak_flag, coincident_peak_flag, ercot_4cp_flag from weather and load shape features
• Demand response targeting — predict demand_response_activation_flag given temperature, humidity, heat index, and price signals
• EV charging load disaggregation — decompose total_demand_MW into EV-driven components using ev_l2_charging_MW, ev_dcfc_charging_MW, and TOU rate tier features
• V2G dispatch optimization — model v2g_discharge_MW as a function of evening peak, LMP, and SoC proxies
• Behind-the-meter DER aggregation — combine rooftop_solar_generation_MW, battery_storage_dispatch_MW, and virtual_power_plant_flag for net-load forecasting
• Climate zone transfer learning — train per-climate-zone load models and test cross-zone generalization
• Load duration curve construction — sort total_demand_MW descending per zone for resource adequacy / capacity planning analyses
• Price elasticity estimation — use the included price_elasticity_demand values and TOU tier features as targets / instruments
• LMP forecasting & energy trading — train short-term price models conditioned on demand, weather, and TOU signals

---

Loading examples

from datasets import load_dataset

ds = load_dataset("xpertsystems/enr003-sample", split="train")
print(ds.shape)

import pandas as pd
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="xpertsystems/enr003-sample",
    filename="enr003_demand.parquet",
    repo_type="dataset",
)
df = pd.read_parquet(path)

# Multi-horizon forecast MAPE evaluation
for col, h in [("fc1h_demand_MW", "1h"), ("forecast_demand_MW", "24h"), ("fc168h_demand_MW", "168h")]:
    mape = ((df[col] - df["total_demand_MW"]).abs() / df["total_demand_MW"]).mean() * 100
    print(f"{h:>5} ahead MAPE: {mape:.2f}%")

# Per-climate-zone temperature–demand sensitivity
for cz, sub in df.groupby("climate_zone"):
    corr = sub["temperature_C"].corr(sub["total_demand_MW"])
    print(f"Zone {cz}: temp-demand correlation = {corr:+.3f}")

# Build a 15-min load duration curve for a single zone
zone_a = df[df["zone_id"] == df["zone_id"].iloc[0]]
ldc = zone_a["total_demand_MW"].sort_values(ascending=False).reset_index(drop=True)
print(ldc.describe())

---

Limitations and honest disclosures

This sample is calibrated for structural fidelity, not bit-exact reproduction of any specific utility's load archive. Specifically:

• HDD / CDD columns use a non-standard temperature base. The generator computes hdd = max(0, 65*5/9 - temp_C) ≈ max(0, 36.11 - temp_C), which uses an unconverted 65×(5/9) ≈ 36.11°C base instead of the standard ASHRAE 65°F → 18.33°C base. This inflates HDD (~30 for typical winter temps vs ~5–15 expected) and floors CDD at zero except in extreme heat. Use temperature_C directly for degree-day analyses; do not consume heating_degree_days / cooling_degree_days columns as-published. The full product release ships an HDD/CDD recompute patch.
• peak_to_valley_ratio and load_factor are CUMULATIVE within each day. The generator's daily_peak_tracker builds up as the day progresses, so these columns at hour 0 reflect only midnight data, not the full day. By hour 23 they reflect the entire day. Use end-of-day (hour 23) readings for full-day load curve metrics, or recompute from total_demand_MW grouped by date.
• forecast_mape_pct is a random draw, not a computed MAPE. The resulting forecast_demand_MW produces an actual MAPE close to (but not exactly equal to) the published forecast_mape_pct column. Compute MAPE directly: (forecast_demand_MW - total_demand_MW).abs() / total_demand_MW.
• P10/P90 forecast intervals over-cover by design. The generator sets sigma = abs(error) * 1.2 per-row, which builds the interval AROUND the realized error rather than reflecting uncalibrated forecast uncertainty. Empirical coverage is ~100%. For probabilistic forecast research, treat these as upper-bound conservative intervals.
• **Component demands (residential + commercial + industrial + agricultural
  • EV) sum to ~65% of total_demand_MW.** Weather adjustment (weather_adj, computed from temperature sensitivities) is added separately to total, so components don't sum to total. Use components for share analysis, not reconciliation.
• Single-season sample (Winter only) at 1-week scale. Multi-season load factor amplification (Summer +28%, Spring -18%, etc.) cannot be validated in this sample; the full product covers the full annual cycle.
• Solar irradiance uses simplified mid-latitude sunrise/sunset model without longitude awareness — all zones share "solar noon ≈ 12:00 UTC." Fine for fleet-aggregate ML; don't expect timestamp ↔ local-clock alignment for any specific geography.
• On-Peak TOU tier fires only in summer (line 446 of generator). In winter weeks like this sample, tou_rate_tier takes values Super-Off-Peak / Off-Peak / Mid-Peak only.
• Climate zone draw varies per seed — at n=8 zones from a pool of 8 IECC zones, typically 4–6 zones appear per sample.
• load_density_MW_per_km2 uses per-row random divisor uniform(50, 500), not a fixed per-zone area. Use for fleet aggregates, not zone-level density studies.

The full ENR003 product addresses these by ASHRAE-correct HDD/CDD, post-loop load curve recomputation, calibrated probabilistic forecasts, multi-season annual cycle, and longitude-aware solar — contact us for the licensed commercial release.

---

Companion datasets in the Energy & Climate vertical

• ENR-001 — Synthetic Power Grid Operations Dataset (bus telemetry, line flows, generation dispatch, frequency, contingency)
• ENR-002 — Synthetic Renewable Energy Generation Dataset (solar/wind/ hybrid SCADA, weather, forecast, PCC, BESS)
• ENR-003 — Synthetic Electricity Demand & Load Forecasting Dataset (you are here)

Use ENR-001 + ENR-002 + ENR-003 together for full grid + renewables + demand ML workflows: dispatch decisions (ENR-001) conditioned on plant-level renewable telemetry (ENR-002) and zone-level demand forecasts (ENR-003).

For the broader catalog, see:

• Oil & Gas — OIL-001 through OIL-004
• Materials & Energy — MAT-001
• Insurance & Risk — 10 SKUs
• Cybersecurity — 11 SKUs

---

Citation

@dataset{xpertsystems_enr003_sample_2026,
  author       = {XpertSystems.ai},
  title        = {ENR003 Synthetic Electricity Demand and Load Forecasting Dataset (Sample Preview)},
  year         = 2026,
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/datasets/xpertsystems/enr003-sample}
}

---

Contact

• Web: https://xpertsystems.ai
• Email: pradeep@xpertsystems.ai
• Full product catalog: Cybersecurity, Insurance & Risk, Materials & Energy, Oil & Gas, Energy & Climate, and more

Sample License: CC-BY-NC-4.0 (Creative Commons Attribution-NonCommercial 4.0) Full product License: Commercial — please contact for pricing.