NeuralBES: A Differentiable, Control-Aware Emulator for Scalable Building Energy Modeling
NeuralBES is a differentiable RC building-energy emulator that stays physics-valid across heterogeneous homes.
NeuralBES parameterizes a resistance-capacitance thermal model with a shared neural encoder that maps building metadata to bounded physical coefficients. A log-space parallel scan solves the linear recurrence, and a predictor-corrector handles thermostat nonlinearity with full-horizon gradients. Trained on ResStock across three climate zones, it is the only data-conditioned model that is both physics-valid and within 4 MAPE points of the strongest raw-error baseline on full-year rollout, using about an order of magnitude fewer parameters than transformer and recurrent baselines.
- Shared encoder maps metadata to bounded RC and equipment coefficients.
- Log-space parallel scan enables scalable differentiable simulation.
- Within 4 MAPE of the best raw-error baseline on annual rollout.
- Physics-valid and far smaller than transformer and recurrent baselines.
- At parameter parity, more than halves grey-box RC MAPE.
From the original article
1 min read at arxiv.orgDemand-side flexibility i. e. forecasting, shifting, and curtailing residential energy loads, depends on thermal models trusted across millions of heterogeneous buildings.
By Ting-Yu Dai · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.