DPE Label Prediction
Built an interpretable multiclass model to predict DPE energy labels (A–G) from open ADEME data, with evaluation focused on class imbalance, robustness, and next-step improvements.
The story
If the Ecowatt page was the friendly public face of this project, this is the engine room: the part where I promised to spare you the details and then, unable to help myself, wrote a whole page of them anyway.
The job sounds simple. Look at a Paris flat's basic details and guess its DPE energy grade, A through G. The catch is that French energy records do not hand you clean data. They hand you a couple hundred thousand rows of it, in which the pristine, planet-saving grade A homes are so rare the model quietly suspects they are a myth.
My favourite defeat came not from the statistics but from myself. The finished model flatly refused to reload, and after some dignified panic I found out why: I had defined a little helper function inside my notebook, which is a bit like leaving your house key inside the locked house. The fix was embarrassingly simple, which is the most annoying kind of fix.
It lands at 73% accuracy, which is not a number that wins hackathons but is an honest one. The rest of the page is me showing my work, what it eats, how it thinks, and every place it is still wrong, because pretending otherwise is how you end up trusting a model that does not believe grade A exists.
Data sources and scope
- Geography: Paris
- Time window: 2018–2021
- Primary source: ADEME open DPE datasets (pre-2021 and post-2021 methodology)
- Broader project context: Enedis open data on annual residential consumption by address (Paris-only cleaning)
Target and feature design
The supervised target is the DPE label etiquette dpe, with 7 classes: A, B, C, D, E, F, G. The feature set was intentionally kept lightweight for interpretability.
- Numerical: surface habitable logement, conso 5 usages par m2 ep, emission ges 5 usages
- Categorical: type batiment, periode construction, type energie principale chauffage
- After preprocessing: 24-dimensional vector (scaling + one-hot encoding)
Preprocessing pipeline
The full workflow is implemented as a scikit-learn Pipeline, ensuring identical transformations at training and inference time.
Numerical variables are median-imputed and standardized. Categorical variables are imputed with the most frequent value, coerced to strings, and one-hot encoded with unknown categories ignored to keep inference robust.
Reloading the serialized model (joblib) initially failed because a helper function used in preprocessing was defined inside a notebook. The robust fix is to move helpers into a dedicated Python module and import them consistently during training/inference.
Train/test split and dataset sizes
A stratified 80/20 split was used to preserve the A–G label distribution under class imbalance.
- Dev set: 80,349 rows
- Test set: 150,155 rows
- Note: split is stratified but not grouped (possible leakage if multiple records per building/address)
Model choice and training
We trained a linear multiclass classifier using SGDClassifier with multinomial logistic regression (loss="log_loss"). This choice fits sparse one-hot features and remains interpretable.
Class imbalance is handled via class_weight="balanced". Training raised a non-convergence warning, pointing to tuning opportunities (increase max_iter, adjust tol and/or alpha).
Model form
After preprocessing, the input becomes a feature vector ϕ(x) ∈ ℝ²⁴. For each class k ∈ A–G, the model computes a linear score:
s_k(x) = b_k + Σ_{j=1..24} w_{k,j} ϕ_{j}(x)
Scores are converted to probabilities with softmax, and the prediction is the argmax over class scores.
Evaluation results
Held-out test set (n = 150,155): accuracy = 0.73, balanced accuracy = 0.798, macro F1 = 0.72, weighted F1 = 0.72.
Performance is strongest on high-support labels (notably C and D). Rare labels are harder to evaluate reliably (especially A), and class E is predicted conservatively (high precision, lower recall).
Limitations
- Non-convergence warning suggests further tuning and/or regularization adjustments.
- Potential leakage risk: split is not grouped by building/address identifier.
- Class A scarcity makes its precision/recall estimates noisy.
- DPE methodology shift in mid-2021 complicates cross-period comparability.
Next improvements
- Increase max_iter and tune alpha/tol to improve convergence stability.
- Validate robustness with a grouped split (more conservative generalization).
- Consider merging rare labels or optimizing for broader risk bands.
- Roadmap: train a regression model on measured Enedis consumption + uncertainty intervals.
Report
Download PDFFull model build report (Paris, 2018–2021), including dataset scope, feature design, preprocessing, explicit model form, and evaluation discussion.