Machine learning for adaptive building façades

Dynamic shading façades must reconcile competing objectives such as daylight, glare, thermal comfort and energy, while responding to what the occupant actually wants. Full building-performance simulation (BPS) typically runs far slower than real time (often minutes to hours for a single configuration), which makes direct use in an online controller impractical. This demonstrator instead explores the surrogate approach proposed in the research: generate labelled data from a physics-based simulator, train a fast deep-learning surrogate on that data, deploy it inside a model-predictive controller (MPC), and let simulated sensor feedback close the loop, including when the occupant overrides the baseline. It is a working proof of concept, a deliberately simplified single-zone example of the end-to-end idea, not a full implementation of the research programme.

Plain-language summary

A physics simulator works out how each façade setting affects daylight, glare, comfort, and energy. It is accurate but slow, so it is run thousands of times to “train” a small, fast model (a surrogate) that learns to give the same answers almost instantly. From then on the closed loop runs on that fast surrogate: each step the controller scores 60 candidate settings, picks the one that best balances the occupant's preferences, the physics simulator confirms the realised outcome, and the loop repeats. If the occupant overrides, the controller honours it and later recovers to the energy-efficient baseline. All of this runs in your browser on your own machine; nothing is sent to a server.

1 · The problem and the physics

The test case is a single east-facing office module in Wellington, NZ (41.3° S), simulated for a summer day: an 8 m² glazed façade facing a 40 m² floor plate, natural (stack-effect) ventilation, and a capacity-limited HVAC plant (4.5 kW) with separate heating and cooling coils. The shading configuration has three actuators (blind deployment, slat angle, and vent opening) whose non-linear combined effects on daylight, glare, comfort and energy cannot be captured by simple rules.

2 · Generate simulation data

A fast reference simulator encodes the governing physics: solar position and irradiance from Wellington's latitude and the time of day, glass and blind transmittance, solar gain, stack-effect ventilation, and the HVAC plant. Random sampling across the full operating envelope (environmental conditions and façade configurations alike) produces thousands of labelled scenarios, the stand-in for the building-performance simulation and multi-objective optimisation studies described in the research programme.

inputmeaningrange
t_outoutdoor temperature-8 … 38 °C
GHIsolar irradiance0 … 1000 W/m²
α_sunsolar elevation0 … 80°
n_occoccupancy0 … 1
T_setthermal setpoint18 … 26 °C
dblind deployment0 … 1
φslat angle0 … 90°
vvent opening0 … 1

Eight inputs (five that describe the operating state, three that describe the façade configuration) map to four performance targets: illuminance, daylight-glare probability, thermal comfort (PMV), and HVAC energy demand.

3 · Train a surrogate

A two-hidden-layer feedforward network (48 + 48 units, ReLU) learns the simulation mapping end-to-end. Inputs are standardised before training; the network is fitted with Adam on 80% of the scenarios and evaluated on the held-out 20%. The loss curves and parity plot on the dashboard are computed against that held-out set, so the reported performance measures generalisation to unseen operating conditions rather than memorisation of the training data.

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 by a small hand-rolled inference routine, with no deep-learning runtime required at deployment time. A single façade evaluation costs on the order of microseconds, which is what makes real-time re-optimisation feasible.

4 · Closed-loop MPC

The surrogate makes online optimisation tractable. Every control step, the controller scores the feasible grid of configurations (5 blind settings × 4 slat angles × 3 vent openings = 60 candidates) and selects the one minimising a weighted multi-objective cost:

cost(x) =

1.0 · daylight deficit + 1.4 · glare exceedance + 1.2 · thermal violation

+ w_e · energy + 0.14 · switching penalty

w_e scales with the occupant's energy priority (0.7–2.3)

The switching penalty rejects chattering between configurations and promotes robust, reachable actuation. Simulated sensor readings close the loop: each decision is taken against fresh observations of the operating state, the chosen configuration is applied, and the realised response is recorded as telemetry. Because each surrogate call is sub-millisecond, the full re-optimisation (the same principle behind a receding-horizon controller at scale) happens in real time, rather than re-running a full building-performance simulation or optimisation pass at every step.

5 · Occupant agency and recovery

This is the part that most directly mirrors the research focus on responsive control. When the occupant expresses a desire for a different environmental state (more light, less glare, warmer or cooler conditions, or a stronger energy-efficiency priority), the request reshapes the objective function for the next-best action. The controller then picks the feasible configuration that best honours the expressed preference.

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

Once the preference is satisfied, its weight decays and the controller guides the façade back toward the energy-efficient operating baseline (the “override and recover” behaviour targeted in the research, achieved by trusting the surrogate instead of re-running a full optimisation at every step.

6 · Evaluation and robustness

  • The surrogate is judged on held-out data via R² and mean absolute error (the metrics shown on the dashboard), which quantify generalisation to previously unseen operating conditions.
  • The parity plot visually checks spread vs bias across the prediction range; the loss curves expose overfitting if training and validation curves diverge.
  • Computational efficiency is measured directly: the model's evaluation time (µs) is reported on the dashboard as the binding constraint for real-time suitability.
  • Independence from a heavyweight ML runtime (portable JSON weights + hand-rolled inference) supports deployment onto the lightweight edge hardware a physical façade system would use.

Limitations and next steps

  • The reference simulator is deliberately lightweight; production work would validate against high-fidelity BPS tools (thermal, daylight, CFD airflow) and multi-objective optimisation results from the research programme.
  • MPC currently uses a static surrogate; online adaptation and Bayesian calibration to live sensor streams is a natural extension.
  • The closed-loop search is a grid over feasible configurations; reinforcement-learning policies or mixed-integer optimisation could replace it as the action space grows.
  • Uncertainty quantification (ensembles or MC-dropout) is an explicit next step, to make the control policy robust to surrogate error on low-data regions of the operating envelope.