IDEAL Household Energy Dataset

• IDEAL Household Energy Dataset is a rich, longitudinal collection of high-frequency sensor data and household metadata from 255 UK residences.
• It features detailed streams for electricity, gas, and environmental conditions, enabling precise aggregation and engineered covariates.
• The dataset supports causal ML applications using tools like PyWhy™ and CausalML™, validated through rigorous refutation and sensitivity tests.

The IDEAL Household Energy Dataset constitutes a richly instrumented, longitudinal resource originating from the University of Edinburgh and is specifically designed to enable finely resolved analysis of domestic energy consumption, associated covariates, and potential causal relationships. Encompassing high-frequency sensor data and detailed household metadata, the dataset enables rigorous ML and causal inference in smart home and IoT-based engineering contexts (Kosioris et al., 17 May 2025).

1. Data Collection and Structure

The IDEAL Dataset (Pullinger et al., 2021) is built from 255 residences in and around Edinburgh, UK, monitored continuously for periods up to 18 months. Analyses typically focused on the 212–208 homes with ≥6 months of non-missing data for key streams. Its granularity and coverage support both aggregate and fine-grained analyses of energy usage.

Measurement streams include:

• Whole-house electricity (kWh) at 1 s intervals.
• Gas pulses at 1 s intervals, converted to volume.
• Room-level temperature (°C), humidity (%), illumination (lux) at 12 s intervals.
• Boiler-tube temperatures for central heating/hot water at 12 s intervals.
• In 39 homes: appliance-level power, sub-circuit real power, radiator/tap-level temperature.

Accompanying metadata and survey features:

• Demographics: income, education, self-reported energy awareness.
• Building characteristics: floor area, number of rooms, presence of energy-intensive appliances.
• External covariates: weather, urbanization indices.

Data organization consists of CSV files per household/room/sensor, each with a unique sensor mapping to physical locations and corresponding metadata linkages.

2. Preprocessing and Feature Engineering

The IDEAL raw sensor data undergoes protocolized transformation for ML and causal analysis:

1. Home selection: Only homes meeting completeness thresholds (≥6 months core data) are retained, resulting in 212 cases for electricity and 208 for gas-focused analyses.
2. Aggregation: Computation of seasonal (cold: Sep–Mar; warm: Apr–Sep) and full-year consumption and temperature metrics.
3. Engineered covariates:
   • Building form: total floor area, room heights, total house volume.
   • Envelope features: numbers of external windows (segregating openable), doors, walls, percentage with TRVs (thermostatic radiator valves).
   • Usage patterns: clothes drying area, tenant count, presence of electric heaters, fans, or AC units.
   • Economic parameters: average regional electricity/gas prices (in £/kWh).
4. Encoding and imputation: One-hot encoding for categorical variables (e.g., urban/rural); missing values imputed to mean within the sample.
5. Label generation: Derivation of binary targets, e.g., high_consumption, high_income, many_electric_heaters, window_rich (num_openable_windows above mean).
6. Dataset partitioning: Generation of separate CSVs for each period (full-year, cold-month, warm-month), each including designated treatment, outcome, and covariate columns.

This schema facilitates sharply defined queries and robust estimation of treatment effects.

3. Causal Framework and Graphical Modeling

Each of the 18 household energy queries is formalized via a compact Directed Acyclic Graph (DAG), encapsulating variable dependencies and domain knowledge. For example, Query 4 (“Do homes with many openable windows consume more electricity?”) uses the following graphical model:

• Nodes:
  • $X$: many_openable_windows (treatment)
  • $Y$: average_electricity_consumption (outcome)
  • $U_1$: occupant_environmental_behavior (unobserved)
  • $U_2$: window_size (unobserved)
• Edges:
  • $U_1 \rightarrow X,\, U_1 \rightarrow Y$
  • $U_2 \rightarrow X,\, U_2 \rightarrow Y$
  • $X \rightarrow Y$

The estimand is expressed in do-calculus as $P(Y\,|\,\mathrm{do}(X=1))$, with ATE given by

$$\mathrm{ATE} = E[Y\,|\,\mathrm{do}(X=1)] - E[Y\,|\,\mathrm{do}(X=0)]$$

Unobserved confounders are explicitly noted and visualized, and domain expertise guides graph construction. Similar representations are used for each causal query.

4. Scope of Causal Queries

The employed causal questions span socioeconomic, structural, and usage interventions, always stated as treatment effect estimands:

#	Treatment ($X$)	Outcome ($Y$0)	Query in do-calculus
1	high_income	electricity_consumption	$Y$1
2	high_education	electricity_consumption	$Y$2
3	many_windows	electricity_consumption	$Y$3
4	many_openable_windows	electricity_consumption	$Y$4
...	...	...	...
17	fans_or_AC	$Y$5 avg_elec_consumption_summer	$Y$6
18	fans_or_AC	$Y$7 avg_indoor_temp_summer	as above

Analyses use both level and difference outcomes, operationalizing climate period and usage pattern interventions.

5. Causal Inference Methodologies

Two Python toolkits were benchmarked using IDEAL data:

5.1 PyWhy™ (DoWhy):

• Models use explicit SCMS (structural causal models) and DAGs.
• Back-door adjustment formula:

$Y$8

• Estimation techniques:
  • Propensity Score Matching
  • Propensity Score Stratification
  • Inverse Propensity Weighting (IPW)
  • Outcome regressions (Linear Regression, MLP)
• Model validation via internal refutation:
  • Random common cause
  • Placebo treatment
  • Data-subset tests

5.2 CausalML™:

• Meta-learner frameworks for HTE/uplift estimation:
  • S-learner (single regressor $Y$9, pseudo-effect $U_1$0)
  • T-learner (distinct $U_1$1; $U_1$2)
  • X-learner (imputation-based enhancements)
  • R-learner for orthogonalized risk minimization:

  $U_1$3

  with $U_1$4, $U_1$5

• Base regressors: Linear Regression, XGBoost

• Validation uses AUUC (Area Under Uplift Curve) and cross-validation

6. Empirical Findings and Validations

Exemplar: Query 4 (openable windows → electricity)

• PyWhy™ ATE estimates (relative % change):

  • Linear regression: +27%
  • Propensity score matching: +24%
  • IPW/stratification: +43%
• CausalML™ XGBoost T-Learner: +23% (other meta-learners: 3%–15%)
• Sensitivity analysis (PyWhy™):
  • Random common cause: ~1% ATE change
  • Simulated unobserved confounder: 11% shift
  • Placebo: 27% reduction toward zero
• Uplift (CausalML™): XGBoost T/X/R highest AUUC (~0.4–0.5), S-learner/linear models <0.2

Aggregate insights:

• XGBoost T-learner estimates broadly concur with PyWhy™ matching/regression.
• Linear regression systematically underperforms due to bias/variance.
• MLP outcomes (PyWhy™) show instability and are omitted from summaries.
• Synthetic 10k-sample benchmarks reveal XGBoost meta-learners minimize MSE; S-learner with Linear Regression exhibits largest KL divergence.

7. Research Implications and Prospective Directions

Several key insights and opportunities emerge from the application of the IDEAL Dataset to causal ML:

• Specification of DAGs is critical for SCM-driven tools (PyWhy™), but data-driven meta-learners (CausalML™) can provide equally robust ATE estimates.
• Internal refutation techniques (e.g., sensitivity to random/unobserved causes, placebos) test model validity within observational constraints but cannot fully confirm ground-truth causal effects absent randomized interventions.
• Uplift metrics (AUUC) provide actionable guidance for selection and deployment of meta-learners.

Limitations and future work:

• 43 homes excluded due to data incompleteness, indicating scope for enhanced robustness via expanded sampling.
• Geographic generalizability requires collection of analogous data across diverse regions and climates.
• Potential for amplifying causal heterogeneity analysis through CATEs over socioeconomic or structural subgroups.
• Toolchain enhancements: graphical user interfaces for SCM model definition (PyWhy™), integration of supplementary causal ML toolkits (CausalLib, Causalinference).
• Real-world validation via architectural/appliance interventions to test estimated effects against actual energy savings.

The IDEAL Household Energy Dataset, as operationalized and evaluated, provides a uniquely detailed and methodologically robust foundation for causal inference in residential energy research (Kosioris et al., 17 May 2025).