iaac-maai/day3-example-daily-load-profiles

Daily household load profiles (Sceaux, 2006–2010), with typical-day clusters

1,411 days of electricity use of one household in Sceaux (near Paris, France), between December 2006 and November 2010. Each row is one day: the average power in each of the 24 hours, the daily total, calendar information, and the typical-day profile the day belongs to ("workday, home in the morning", "at home all day", "workday, empty house", "nobody home"), found by clustering.

It turns a well-known minute-by-minute dataset into a ready-to-use daily table for studying load profiles, and it is the worked example dataset for the Day 3 assignment of AI Methods I: Machine Learning for the Built Environment (IAAC MaAI 2026/27).

• GitHub repository (how it was made): https://github.com/iaac-maai/ai-methods-1-2026 (notebook Day_3/docs/application_of_unsupervised_learning.ipynb, Application B and section 8)
• Model fitted on it: https://huggingface.co/iaac-maai/day3-example-load-profiles
• Authors: AI Methods I course team, IAAC

Source and attribution

Original dataset	Individual Household Electric Power Consumption by Georges Hebrail and Alice Berard, UCI Machine Learning Repository, https://doi.org/10.24432/C58K54
Original licence	CC BY 4.0
Original content	About 2 million one-minute measurements (December 2006 – November 2010) of the household's active and reactive power, voltage, current and three sub-meters

This dataset is a derived work, released under the same licence (CC BY 4.0) with attribution to the original authors.

What we changed

• Selection: only Global_active_power (the total active power of the household, in kW) is used.
• Aggregation: one-minute readings → hourly averages. An hour with fewer than 45 valid minutes (out of 60) is treated as missing.
• Cleaning: missing values (? in the original file) are ignored in the averages; only days with all 24 hours valid are kept: 1,411 days.
• Feature engineering: the daily energy use (daily_kwh) and calendar features (weekday, weekend, month, season).
• Clustering: each day's 24-hour curve is divided by its own total (to compare shapes, not volumes) and grouped with K-Means into 4 typical-day profiles (cluster, cluster_name). The day's position on three 2D maps of the curves is included (PCA, t-SNE and UMAP coordinates), so the clusters can be plotted without recomputing anything. Details, choice of k and stability checks: see the model card.

What the clusters look like

The average day of each profile, in kW (computed from the h00 … h23 columns):

• workday, home in the morning: a sharp breakfast peak at 7:00 (about 2.0 kW), moderate use through the morning, a second peak around 20:00.
• workday, empty house: the same breakfast peak, then the lowest daytime use of all profiles (about 0.55 kW at 15:00, only the base load) and the highest peak of all at 21:00 (about 2.6 kW).
• at home all day: use rises steadily through the day and peaks in the evening (about 2.4 kW at 19:00): someone is at home cooking, heating and using appliances.
• nobody home: a flat line between 0.6 and 1.1 kW, with no morning or evening peak: only the fridge and devices on standby.

Left: how the clustering behaves. Each dot is one day, placed on a 2D map (umap_x, umap_y) so that days with similar curves are neighbours. The four profiles occupy neighbouring regions of one continuous cloud, with no empty space between them: the two workday profiles mix along their border, and only the "nobody home" days form a tight group of their own (top left). The clusters are a useful way of cutting a continuum, not separate islands; this is why the silhouette score is low (about 0.1) and why another algorithm would draw the borders differently.

Right: the profiles through the year (month and cluster_name columns). "Nobody home" makes up 55% of August and 28% of July (summer holidays); "at home all day" is most common from November to January (34–41% of the days); the two workday profiles dominate the rest of the year.

Three maps of the same days

The pca_*, tsne_* and umap_* columns place every day on three 2D maps of its 24-hour shape. Top row: coloured by cluster_name. Bottom row: coloured by weekend, a column that was not used for clustering.

• PCA is a linear projection, so its axes have a meaning, read from the component loadings. PC1 (21% of the variance) grows with a breakfast peak at 7:00, an evening peak at 19:00–22:00 and an empty middle of the day. PC2 (15%) separates morning-heavy days (top) from evening-heavy days (bottom). The four profiles are four slices of one round cloud: "workday, empty house" on the right, "workday, home in the morning" at the top, "at home all day" at the bottom and "nobody home" (flat days) on the left. There are no gaps between them: K-Means cut a continuous cloud. Two components keep only 36% of the variance, so this map is a rough view.
• t-SNE and UMAP keep neighbourhoods instead. The "nobody home" days form a tight knot (the holidays): the only group that is clearly set apart. The other borders stay fuzzy. On these maps the axes, the distances between groups and the group sizes mean nothing.
• Bottom row: on all three maps the weekends gather where the "at home all day" days are, although weekend was never given to the algorithm. The clusters capture real behaviour. They are not the same thing, though: 39% of the "at home all day" days are weekdays (winter days at home, days off).

To redraw any of these maps, see How to load it.

Columns

Column	Type	Unit	Description
date	date	–	The day (YYYY-MM-DD)
h00 … h23	float	kW	Average active power of the whole household during hour 0 (00:00–00:59) … hour 23
daily_kwh	float	kWh	Total electricity used that day (sum of the 24 hourly averages)
weekday	text	–	Day of the week (Monday … Sunday)
weekend	bool	–	True on Saturday and Sunday
month	integer	–	Month (1–12)
season	text	–	Meteorological season: winter (Dec–Feb), spring, summer, autumn
cluster	integer	–	Typical-day profile assigned by the model (0–3)
cluster_name	text	–	Name of the profile (our interpretation): workday, home in the morning (33.5% of days), at home all day (26.3%), workday, empty house (25.6%), nobody home (14.7%)
pca_x, pca_y	float	–	Position of the day on the first two principal components of the 24-hour shapes (a linear projection: distances are roughly meaningful)
tsne_x, tsne_y	float	–	Position of the day on a 2D t-SNE map of the shapes (perplexity=30). Only for plotting: distances between groups and group sizes are not meaningful
umap_x, umap_y	float	–	Position of the day on a 2D UMAP map of the shapes (n_neighbors=15, min_dist=0.1). Only for plotting, with the same caveat as t-SNE

Limitations and biases

• One household. These are the habits of one family in one house near Paris; they do not represent other households, building types or climates.
• Gaps. Days with missing hours were removed, so some periods (including a few longer outages in the original data) are absent.
• The clusters are one reading of the data. Daily habits form a continuum (silhouette ≈ 0.1); another algorithm or another number of clusters would draw the borders differently. The names are our interpretation of the curves and the calendar, not information about the occupants.
• The t-SNE and UMAP coordinates are not reproducible exactly on another computer or library version.

How to load it

import pandas as pd

df = pd.read_csv(
    "hf://datasets/iaac-maai/day3-example-daily-load-profiles/daily_load_profiles_clustered.csv",
    parse_dates=["date"],
)
df.groupby("cluster_name")["daily_kwh"].median()

# Draw the clusters on any of the three maps: pca_x/pca_y, tsne_x/tsne_y or umap_x/umap_y
import matplotlib.pyplot as plt

for name, days in df.groupby("cluster_name"):
    plt.scatter(days["umap_x"], days["umap_y"], s=8, label=name)
plt.legend()
plt.show()

Reading hf:// paths requires the huggingface_hub package.

Authors and licence

AI Methods I course team. Worked example for AI Methods I: Machine Learning for the Built Environment, IAAC MaAI 2026/27.

Released under CC BY 4.0, based on Individual Household Electric Power Consumption by G. Hebrail and A. Berard (UCI Machine Learning Repository, https://doi.org/10.24432/C58K54, CC BY 4.0).