Benchmarking Model Predictive Control Algorithms in Building Optimization Testing Framework (BOPTEST)

TL;DR

The paper tackles the high computational cost of nonlinear model predictive control (NMPC) for physics-based building emulators by proposing offline training of differentiable surrogate models that accelerate evaluations and provide gradients for NMPC. It investigates Linear, MLP, and LSTM surrogates to approximate the dynamics $x_{t+1}=f(x_t,u_t,d_t)$ and employs automatic differentiation to solve NMPC with box constraints, comparing gradient-based and SQP solvers across two BOPTEST case studies. The results show that LSTM and MLP surrogates can achieve strong predictive accuracy and enable energy-reducing control under comfort constraints, with trade-offs between model complexity, computation time, and Jacobian conditioning. The framework is modular and adaptable to different models and control formulations, enabling scalable prototyping of predictive controllers for large buildings and fostering robust real-world deployment.

Abstract

We present a data-driven modeling and control framework for physics-based building emulators. Our approach consists of: (a) Offline training of differentiable surrogate models that accelerate model evaluations, provide cost-effective gradients, and maintain good predictive accuracy for the receding horizon in Model Predictive Control (MPC), and (b) Formulating and solving nonlinear building HVAC MPC problems. We extensively evaluate the modeling and control performance using multiple surrogate models and optimization frameworks across various test cases available in the Building Optimization Testing Framework (BOPTEST). Our framework is compatible with other modeling techniques and can be customized with different control formulations, making it adaptable and future-proof for test cases currently under development for BOPTEST. This modularity provides a path towards prototyping predictive controllers in large buildings, ensuring scalability and robustness in real-world applications.

Figures (3)

• Figure 1: Test MSE for different choices of surrogate models in multi-zone test case. LSTM and MLP have comparable performance and outperform the Linear model.
• Figure 2: The set of figures show the results of out-of-training predictive performance for five zone model during three distinct weather events (January, May, and August) for core zone (top). The ambient temperature trajectories is depicted in red (bottom). The orange lines represent the 50-step ahead predictions (12.5 hours) starting from the left most point of the trajectory. The full MSEs are reported in Table \ref{['Table:MLP']}.
• Figure 3: Result comparison for different choices of models and control algorithms. The top figure represents the temperate. The bottom figure is the relevant weather data, and the middle figures are the corresponding control inputs. The results are divided into a cold (Jan) and hot (Aug) weather events. (a) Result for control of core-zone in the multi-zone test case using SLSQP with Linear, MLP, and LSTM models. Using MLP model, the control outperforms LSTM and Linear model-based implementation. (b) MLP-based control results with SLSQP solver slightly outperform the Gradient-based approach.