Case 04 · Research · KCL MSc

Multimodal residential energy prediction

A building's efficiency is written in three places at once: its measurements, its description, and where it stands. The question was how to read them together.

Role: Research, implementation, write-up
Institution: King's College London, MSc
Built with: PyTorch · BERT · scikit-learn · Streamlit

All work

Remove about 30% of the words from each home's written description and R², the share of the variation the model explains, falls from 0.72⁠–⁠0.73 to 0.41⁠–⁠0.47 for all three methods; noise in the tabular features or the location barely moves it.R² of early, gated and late fusion under eight perturbations of the input, re-plotted from my results. One training run (seed 42), not an average of repeated runs.

01The question

An EPC assessment carries tabular building features, a written description, and a location. Most models take the first and discard the rest. This compares three ways of fusing all three to predict SAP and EI scores.

02What made it hard

The interesting failure was not in the architecture but in the split. Addresses repeat across an EPC dataset, because the same property is reassessed, so a random split leaks a building from training into test and every model looks better than it is. Regrouping the split by property changed the results and meant redrawing the figures.

03Beyond the model

A household energy simulator sits alongside it, so a retrofit scenario can be run on a single property rather than argued about in the abstract.

3: modalities: tabular, text, spatial
3: fusion strategies compared
2: targets: SAP and EI score