Personalized Building Climate Control with Contextual Preferential Bayesian Optimization

Efficient tuning of building climate controllers to optimize occupant utility is essential for ensuring overall comfort and satisfaction. However, this is a challenging task since the latent utility are difficult to measure directly. Time-varying contextual factors, such as outdoor temperature, further complicate the problem. To address these challenges, we propose a contextual preferential Bayesian optimization algorithm that leverages binary preference feedback together with contextual information to enable efficient real-time controller tuning. We validate the approach by tuning an economic MPC controller on BOPTEST, a high-fidelity building simulation platform. Over a two-month simulation period, our method outperforms the baseline controller and achieves an improvement of up to 23% in utility. Moreover, for different occupant types, we demonstrate that the algorithm automatically adapts to individual preferences, enabling personalized controller tuning.

💡 Research Summary

The paper tackles the problem of adapting building HVAC controllers to individual occupant preferences in the presence of time‑varying external conditions. Traditional rule‑based, PID, or fixed‑parameter Model Predictive Control (MPC) strategies rely on pre‑defined set‑points and cannot accommodate the diverse, dynamic utility functions of occupants, which may weigh energy cost and thermal comfort differently. Direct measurement of such latent utilities is notoriously noisy and biased, whereas humans are more reliable when providing relative judgments (i.e., “I prefer today’s temperature over yesterday’s”).

To exploit this observation, the authors propose a contextual Preferential Bayesian Optimization (contextual PBO) framework that combines binary preference feedback with contextual variables (outdoor temperature, solar irradiance, real‑time electricity price, etc.). The unknown utility function J(θ, z) – where θ denotes the MPC tuning parameters and z the observed context – is modeled as a function belonging to a reproducing kernel Hilbert space (RKHS) with bounded norm. Preference observations are treated as Bernoulli random variables whose success probability follows a logistic sigmoid of the utility difference, yielding a log‑likelihood that can be maximized to obtain a maximum‑likelihood estimate (MLE) of the latent function. A confidence set Bₜ is then defined around the MLE using a hyper‑parameter β that controls the size of the set.

At each daily iteration t, the algorithm selects the next controller parameters θₜ by solving a max‑min problem: it chooses the θ that maximizes the worst‑case improvement over the previous day across all functions in Bₜ, subject to the observed context zₜ. This “optimism under uncertainty” criterion ensures a principled balance between exploration (learning the shape of J) and exploitation (improving utility). The infinite‑dimensional optimization is reduced to a finite‑dimensional problem via the representer theorem, leading to a kernel matrix K that captures similarity between (θ, z) pairs. In low‑dimensional θ spaces a simple grid search suffices; for higher dimensions, multi‑start nonlinear solvers are recommended.

The methodology is evaluated on BOPTEST, a high‑fidelity building simulation platform, using the “singlezone commercial hydronic” test case. An ARX model (order 10, three exogenous inputs) is identified from a week of randomized bang‑bang excitation to capture the building’s thermal dynamics. The economic MPC minimizes a weighted sum of electricity cost and a large penalty on slack variables that enforce soft temperature constraints. Two tunable parameters are introduced: θ₁, a price‑threshold that determines whether the controller applies a low‑temperature set‑point, and θ₂, the daytime low‑temperature bound itself. Context zₜ consists of outdoor temperature, solar irradiance, and a highly dynamic electricity price profile.

Two occupant profiles are simulated: (1) a cost‑focused occupant whose utility is simply the negative of actual energy cost, and (2) a comfort‑focused occupant whose utility penalizes deviations from a preferred temperature range. Each day, the occupant provides a binary preference comparing the current day’s closed‑loop trajectory with the previous day’s. The contextual PBO algorithm updates its confidence set and proposes new θₜ for the next day.

Results show that contextual PBO outperforms both a baseline MPC (fixed parameters) and a static PBO that ignores context. Over a two‑month simulation, cumulative utility (R_sum) and average daily utility (R_avg) improve by up to 23% relative to the baseline. Moreover, the algorithm automatically discovers distinct θ₁–θ₂ combinations for the two occupant types, demonstrating true personalization: the cost‑focused occupant receives aggressive low‑set‑points when electricity is cheap, while the comfort‑focused occupant receives higher set‑points to maintain thermal satisfaction.

The study highlights several contributions: (i) a novel integration of preference learning with contextual Bayesian optimization for real‑time controller tuning; (ii) a practical formulation that works with binary human feedback, reducing measurement noise and bias; (iii) empirical evidence that context‑aware optimization yields significant utility gains and enables occupant‑specific adaptation. Limitations include reliance on daily preference updates (finer‑grained feedback could accelerate learning) and the fact that the simulation does not capture all real‑world occupant actions (e.g., window opening, blind adjustment). Future work is suggested to incorporate multi‑objective preference data (energy, comfort, air quality), online kernel hyper‑parameter adaptation, and field trials in actual buildings to validate scalability and robustness.