The building is now being asked to file its own load forecast
the building is now being asked to file its own load forecast: a foundation model predicts weekly energy use with a 29.2 percent error cut against gradient boosting and stays stable across one to twelve month training windows, a transfer framework cuts errors by 63 percent with sixteen to thirty two source buildings, and a thermal foundation model on citylearn forecasts indoor temperature under half a degree and transfers from two buildings.
the walls now file their own consumption.
Context
The thermal model is described in the arXiv paper Toward a foundational thermal model for residential buildings (May 2026). It is a physics-informed transformer that embeds domain knowledge such as derivative enrichment and Euler-based numerical integration, evaluated on the CityLearn dataset of 247 residential buildings across three climate zones. The excerpt read says models trained on as few as two buildings generalize to unseen buildings and climate zones without fine-tuning, and that models trained on Texas buildings reach RMSE of 0.48 to 0.55 degrees C on California and Vermont buildings.
The transfer framework with 16 to 32 source buildings matches an arXiv paper whose abstract says multi-source transfer learning yields up to 63% lower forecasting errors than single-source transfer learning, and that data from 16 to 32 source buildings over one year is needed for multi-source models to consistently beat time-series foundation models on mean absolute error.
The note says the thermal model forecasts indoor temperature under half a degree. The paper excerpt gives RMSE of 0.48 to 0.55 degrees C on other buildings, which is about half a degree and not strictly under it in every case, and the authors note the data are simulated residential buildings. The note says the 63 percent cut happens with sixteen to thirty two source buildings. The abstract ties up to 63% to multi-source transfer versus single-source transfer, and ties 16 to 32 buildings to when multi-source models beat foundation models, so these are two separate findings joined in the note. The first item, a foundation model cutting weekly error by 29.2 percent against gradient boosting with stable one to twelve month windows, was not located in any source read, so it is unsupported here, not refuted.
Watch next
- The study behind the 29.2 percent weekly forecast figure. Results on measured, not simulated, buildings.
Sources
- Toward a foundational thermal model for residential buildings (arXiv 2605.01364)arxiv.org
- arXiv 2604.16443 (multi-source transfer learning for building energy forecasting; the search result showed no title)arxiv.org
- From RNNs to Foundation Models: An Empirical Study on Commercial Building Energy Consumption (arXiv 2411.14421)ar5iv.labs.arxiv.org
- Can time-series foundation models perform building energy management tasks? (Data-Centric Engineering)cambridge.org
Provenance
The note above is reproduced unedited from the original post, first published on Threads on 4 October 2026 at 12:56 IST. Sources are the papers and datasets the note draws on.
View the original post ↗Embed this note
More notes
The air is now being asked to keep its own ledger
the air is now being asked to keep its own ledger: ecmwf’s aifs compo becomes the first ai model to forecast atmospheric composition globally every three hours, cleanair simulates 365 days of pm2.5 over china in ten seconds, and a unified framework maps six pollutants at one kilometer across the whole country. the air now files its own composition report.
read the note →The current is now being asked to draw its own map
the current is now being asked to draw its own map: china’s langya 2.0 predicts six ocean phenomena including internal waves and mesoscale eddies, a deep net called wenhai resolves eddies globally with air sea flux formulas built in, and scripps infers surface currents from the way temperature patterns deform in satellite images. the ocean now files its own circulation report.
read the note →The soil is now being asked to report its own carbon
the soil is now being asked to report its own carbon: a nix color sensor paired with generative data augmentation predicts soil organic carbon without a lab, random forest drives 74 percent of soil health mapping studies, and sentinel 2 tracks five year carbon change across france and italy from 922 samples. the dirt now files its own carbon account.
read the note →