iaac-maai/day3-example-load-profiles

Household load profiles: grouping days by the shape of their electricity use

This model assigns one day of household electricity use (24 hourly values in kW) to one of four typical daily profiles: two kinds of workday, an "at home all day" profile and a "nobody home" profile. It is a scikit-learn pipeline (Normalizer + KMeans) fitted on 1,411 days measured in one house near Paris between 2006 and 2010.

It is the worked example for the Day 3 assignment of AI Methods I: Machine Learning for the Built Environment (IAAC MaAI 2026/27): it shows students what a published clustering model and its model card should look like.

• GitHub repository (full analysis): https://github.com/iaac-maai/ai-methods-1-2026 (notebook Day_3/docs/application_of_unsupervised_learning.ipynb, Application B and section 8)
• Dataset with the cluster of every day: https://huggingface.co/datasets/iaac-maai/day3-example-daily-load-profiles
• Authors: AI Methods I course team, IAAC

Intended use

• Intended users and use: students, building managers and energy designers who want to label days by their pattern of use, for example to count how many "empty house" workdays a dwelling has (candidates for lowering the heating setpoint during the day) or to spot holiday periods in meter data. It is a teaching model.
• Out-of-scope uses:
  • Other building types (offices, schools, shops) or whole districts: their days look nothing like a household's, but the model will still force them into one of the four profiles.
  • Households with very different equipment (electric car charging at night, heat pumps, rooftop PV): their curves were not in the data.
  • Predicting how much electricity a day will use: the model only looks at the shape of the curve, never the total.
  • Any decision about a specific household or its occupants.

Model details

Task	Clustering (unsupervised)
Algorithm	KMeans
Key hyperparameters	n_clusters=4, n_init=10, random_state=42
Preprocessing	Normalizer(norm="l1"): each day is divided by its own total, so the model groups the shape of the curve (the share of the day's electricity used in each hour), not the volume. No StandardScaler: all 24 columns are in the same unit
File	model.joblib: the whole fitted scikit-learn Pipeline (8 KB)
Library versions	scikit-learn 1.9.1, pandas 3.0.6, numpy 2.5.3, Python 3.12.3

Input features

The model expects a pandas.DataFrame with exactly these 24 columns, one row per day:

Column	Type	Unit	Description	Preprocessing
h00 … h23	numeric	kW	Average active power of the whole household during hour 0 (00:00–00:59) … hour 23	Each row divided by its sum (Normalizer(norm="l1"))

All values must be present (no missing hours) and non-negative.

Output and clusters

model.predict(X) returns the profile number of each day:

Cluster	Name	Share of the days	Shape of the day	Calendar (never used by the model)
0	workday, home in the morning	33.5%	Sharp peak at 7:00, a second peak around 20:00, moderate use in between	90% weekdays; ~26 kWh/day
1	at home all day	26.3%	Low at night, rising through the day, highest in the evening (peak 19:00)	61% weekends, most common in winter (Nov–Jan); ~32 kWh/day
2	workday, empty house	25.6%	Morning peak, almost nothing during the day, strong evening peak at 21:00	91% weekdays; ~24 kWh/day
3	nobody home	14.7%	Flat: no peaks, only the base load (fridge, standby)	55% of all August days, 28% of July; ~17 kWh/day

The names come from our interpretation: we compared each cluster with information the model never saw (day of the week, month, daily total).

Red: the centroid (average day) of each profile. Grey: 40 real days from it. Values are shares of the day's electricity.

Training data

Dataset	Individual Household Electric Power Consumption (Hebrail & Berard), processed into daily profiles
Source	UCI Machine Learning Repository (also on Kaggle); our processed version: iaac-maai/day3-example-daily-load-profiles
Licence	CC BY 4.0
One row is	One day of one household (24 hourly averages)
Size used	1,411 complete days
Coverage	One house in Sceaux (near Paris, France), 17 December 2006 – 25 November 2010

Processing: about 2 million one-minute readings of Global_active_power → hourly averages (an hour counts as missing if fewer than 45 of its 60 minutes are valid) → one row per day → only days with all 24 hours valid are kept (1,411 days).

Method and evaluation

• Choice of k: there is no natural number of clusters: the silhouette score is low for every k from 2 to 8 (0.13 at k = 2, 0.10 at k = 4) and the elbow curve has no clear bend. Daily habits form a continuum. We chose k = 4 for interpretability: four profiles that can be named and acted upon.
• Stability (adjusted Rand index, ARI, against the final clusters; 1 = identical grouping):

Check	Mean ARI	Min ARI
10 different random seeds	0.98	0.96
10 random subsamples of 80% of the days	0.85	0.45
Hierarchical clustering (ward), k = 4	0.34	–

• Reading: the profiles are reproducible for K-Means and fairly robust to losing data, but a different algorithm cuts the continuum differently. Treat them as a useful summary of this household's days, not as natural categories.
• Why this model: K-Means can assign new days (.predict()), its centroids are readable "average days", and it is fast.

Limitations and biases

• One household. The profiles describe the habits of one family in one house, in one climate, between 2006 and 2010. Other households have other profiles.
• Always one of four. K-Means has no "I don't know": a day unlike any of the four (an office, a party, a broken fridge) is still assigned to the nearest profile.
• Weak structure. With a silhouette of about 0.1, many days sit between two profiles; a small change in the curve can move them to the neighbouring profile.
• Shape only. The volume of the day is deliberately ignored, so a cold winter day and a mild day with the same pattern get the same profile.
• Names are interpretations. "Nobody home" is our reading of a flat curve that coincides with the summer holidays; the data says nothing about who was actually at home.

How to use

import joblib
import numpy as np
import pandas as pd
from huggingface_hub import hf_hub_download

path = hf_hub_download(repo_id="iaac-maai/day3-example-load-profiles", filename="model.joblib")
model = joblib.load(path)

CLUSTER_NAMES = {
    0: "workday, home in the morning",
    1: "at home all day",
    2: "workday, empty house",
    3: "nobody home",
}

# One day: average kW for each hour 0..23
day = np.full(24, 0.3)                                   # base load
day[[6, 7]] = [1.2, 1.5]                                 # breakfast
day[[18, 19, 20, 21, 22]] = [1.5, 2.5, 3.0, 2.8, 1.5]    # evening
example = pd.DataFrame([day], columns=[f"h{h:02d}" for h in range(24)])

cluster = model.predict(example)[0]
print(cluster, CLUSTER_NAMES[cluster])                   # 2 workday, empty house

⚠️ .joblib files are Python pickles and can execute code when loaded: only load models from sources you trust. Use the library versions listed above; the model may not load with other versions.

Reproducibility

The full analysis (download, processing, choice of k, interpretation, stability, export) is in the GitHub repository: https://github.com/iaac-maai/ai-methods-1-2026. Run Day_3/docs/application_of_unsupervised_learning.ipynb with the packages in the repository's requirements.txt.

Authors and licence

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

Model released under the MIT License. It was fitted on data derived from Individual Household Electric Power Consumption by G. Hebrail and A. Berard, UCI Machine Learning Repository, https://doi.org/10.24432/C58K54, licensed under CC BY 4.0.