Machine learning for adaptive building façades

Dynamic shading façades must reconcile competing objectives — daylight, glare, thermal comfort and energy — while responding to what the occupant actually wants. Full building-performance simulation is far too slow for real-time operation. This demonstrator shows the pipeline proposed for the research: a deep-learning surrogate trained on simulation data, deployed inside a model-predictive controller that re-optimises the shading configuration every control step.

1 · Generate simulation data

A fast reference simulator encodes the governing physics of an east-facing office room in Wellington: solar position and irradiance, glass and blind transmittance, solar gain, natural cross-ventilation via stack effect, and a capacity-limited HVAC plant with heating and cooling coils. Random sampling across operating conditions and façade configurations produces thousands of labelled scenarios.

inputmeaningrange
t_outoutdoor temperature-838 °C
GHIsolar irradiance01000 W/m²
α_sunsolar elevation080°
n_occoccupancy01
T_setthermal setpoint1826 °C
dblind deployment01
φslat angle090°
vvent opening01

In the full research programme these scenarios would be generated by building performance simulation (thermal + daylight + airflow) and multi-objective optimisation studies. Here the simulator acts as a stand-in, and the surrogate learns its mapping end-to-end.

2 · Train a surrogate

A two-hidden-layer feedforward network (48+48 units, ReLU) maps the 8 operating inputs to the 4 performance targets. Inputs are standardised; the network is trained with Adam on 80% of the data and evaluated on the held-out 20%. The loss curves and parity plots on the dashboard are produced against that held-out set to quantify generalisation rather than memorisation.

E
work-plane illuminance · lux
DGP
daylight glare probability · 0–1
PMV
predicted mean vote · -3–+3
E_hvac
HVAC energy demand · W

Once fitted, the weights are serialised to a portable JSON format and executed as a small hand-rolled inference routine — no deep-learning runtime required at deployment time. Evaluation costs microseconds per façade configuration.

3 · Closed-loop MPC

Every control step the controller evaluates the surrogate across a gridded set of feasible façade configurations and picks the one minimising a weighted objective: daylight deficit, glare exceedance, thermal-comfort violation, HVAC energy, plus a switching penalty that rejects jitter and promotes robust actuation. Real-time sensor readings close the loop: the controller optimises against fresh observations of the operating state, and the simulated response is recorded as telemetry.

cost(x) =

w_d · daylight(x) + w_g · glare(x) + w_t · thermal(x)

+ w_e · energy(x) + w_s · switch(x, x₀)

4 · Occupant agency and recovery

When the occupant expresses a desire for a different environmental state — more light, less glare, warmer conditions — the preference shapes the objective function for the next-best action. Once satisfied, the preference weight decays and the controller guides the façade back toward the energy-efficient operating baseline, exactly the “over-ride and recover” behaviour targeted in the research.

occupant asks “brighter”controller opens shades, slats tiltdaylight target met → returns to baseline

Limitations and next steps

  • The reference simulator is deliberately lightweight; production work would validate against high-fidelity BPS tools (thermal, daylight, CFD airflow).
  • MPC uses a static surrogate; online adaptation and Bayesian calibration to live sensor streams is a natural extension.
  • Switching penalties here are heuristic; reinforcement-learning policies or mixed- integer optimisation could replace the grid search at larger scale.
  • Uncertainty quantification (e.g. ensembles or MC-dropout) is an explicit next step to make the control policy robust to model error.