Energy Dashboard

Clearbox · Energy Telemetry Italian Residential Grid Intelligence

Zone

Loading data...

Households by Province (Top 20)

House Type Distribution

Heating System

Energy Performance Label

Contract Power

Technology Adoption (%)

Cooking Type

PV Adoption by House Type (%)

Methodology: Real ENTSO-E zone load data decomposed into sectors using Terna annual shares, IMCEI/IMSER monthly indices, and day-type profiles. The residential share (green) is compared against our bottom-up simulation predictions.

Daily Zonal Load -Sectoral Decomposition (GWh)

Sectoral Share by Day (%)

Residential: Bottom-Up Simulation vs Top-Down Decomposition (GWh/day)

Solar PV: Simulation vs ENTSO-E — Residential (GWh/day)

Zone Solar PV Generation (MWh)

Backtesting

Methodology: Scatter plot comparing daily residential consumption predictions (bottom-up simulation) against ENTSO-E top-down decomposition estimates across all available historical weeks. Points on the diagonal indicate perfect agreement.

Predicted vs Actual Residential Consumption (GWh/day)

Accuracy by Forecast Lead (daily, %)

P10–P90 Band Calibration by Hour (% of hours)

Error by Hour of Day (%, Sim − Actual)

Hourly Bias by Week (%, Sim − Actual)

Reading the hourly panels: there is no measured hourly residential load, so the hourly reference combines the two calibration references: the day's residential volume from the sectoral decomposition, spread across the hours by ARERA's metered withdrawal profile for the zone, month and day type. The simulation is compared as withdrawal (consumption minus self-consumed PV), which is what ARERA measures. ARERA is a typical shape for the month, not that day's, so these panels diagnose shape rather than verify single hours. Bands: sub-hourly P10/P90 are summed to the hour and hours to the day, i.e. errors are treated as fully correlated (the widest consistent band). A calibrated band holds the actual ~80% of the time, with ~10% falling on each side. The hourly panels ignore the Hours filter; the lead panel ignores the Lead filter.

Residential Load, Grid Draw & PV

Desk Indicators — Peak, Ramp, Threshold

Power in MW (energy per step ÷ step length). Peak and trough at native resolution; ramps on hourly means (MW/h, within the day). The band is P10–P90 at the peak step. P(peak > threshold) assumes normal errors fully correlated across the day (the same convention as the backtest bands), under which it equals the highest single-step exceedance probability. Check the backtest band calibration before relying on it.

PV Generation

Self-Consumption & Battery

Province Load Heatmap (MWh)

Temperature (°C)

Consumption by Component — Average Household

Consumption by Province (MWh, Top 20)

How these stories are generated. For every household sampled by the forecast pipeline we ask Gemini (gemini-2.5-flash-lite) to write a 7-day activity schedule conditioned on (a) the household's sociodemographic profile — occupants, ages, occupations, presence patterns — (b) the home's physical attributes — surface, energy label, heating type, PV/EV/storage — and (c) the week's weather (per-day temperature, cloud, precipitation). The model returns XML covering six behavioural dimensions: occupancy (24×N hour-by-person grid), meals, showers, appliances (washing machine, dishwasher, dryer, oven, vacuum, iron), EV charging, and trips. Identical archetypes (family size + house type + region) re-use a small pool of donor narratives — what you see below are the unique stories actually generated by the model. Use the filters to find weekend, rainy, or prosumer-specific stories; click any card for the full week.
All Has PV Has EV Has Battery Heat Pump Weekend included Rainy day Hot week (>28°) Cold week (<5°) Single occupant Family 3+
ARERA domestic withdrawal statistics (2024). Mean hourly grid withdrawal per domestic customer, aggregated by ARERA from validated meter readings for every domestic withdrawal point via the Sistema Informativo Integrato — a census, not a survey. Values are gross withdrawal at the meter (prosumers included, not netted against injections). The simulation's calibration uses this dataset as its intra-day shape reference only. The published level (watts) is an unweighted mean over the five contract-power classes and is biased by 12–17%; it is shown here as a nominal reference and is never used for calibration — zonal volumes come from the sectoral decomposition instead (see CALIBRATION_NOTES.md). Source: arera.it.

Hourly Withdrawal Shape (Index, Daily Mean = 1)

Nominal Level by Month (W) — Unweighted Class Mean, Not Used for Calibration

Zone Comparison — Weekday Shape (Index)

Weekday Shape by Month × Hour (Index)