NeuralBES: a differentiable emulator simulates the energy consumption of millions of homes without having to calibrate each building individually

A unique neural encoder converts surface area, year and type of heating into physical thermal parameters. The result: consistent annual trajectories with ten times fewer parameters than competing models.

Managing the flexibility of electricity demand (forecasting, shifting or reducing the consumption of housing) requires reliable thermal models for millions of very different buildings. Physical simulators such as EnergyPlus are accurate but slow and must be calibrated building by building; purely statistical sequence models scale but ignore physics and produce seemingly plausible, physically inconsistent curves.

A network-driven resistance-capacitance model

NeuralBES, presented by Ting-Yu Dai and his co-authors, retains a classical thermal model of the resistance-capacitance type, but entrusts its parameterization to a shared neural encoder. The metadata of a dwelling (surface, period of construction, heating and air conditioning system) is converted into physically bounded capacitances, conductances and coefficients of equipment. The calculation is parallelized and a predictor-corrector loop manages the nonlinear relationship between thermostat and temperature while remaining entirely differentiable.

Trained on the ResStock game in three climate zones, the model covers a wide variety of typologies with a single encoder. Over a full year simulation, it is the only model learned to remain physically valid while approaching the best competitor in raw error by 4 points from MAPE, with about ten times fewer parameters than references based on transformers or recurring networks.