Monotonic physics-constrained neural networks for model predictive control of building energy systems

Abstract

This paper presents a comprehensive evaluation of monotonic physics constrained neural networks for model predictive control of building energy systems. Incorporating physical domain knowledge in machine learning, so-called physics-guided machine learning, aims to increase the robustness of black box models. Following an extensive literature review on physics-guided machine learning and its applications, the Constrained Monotonic Neural Network (CMNN) architecture is selected and implemented, which enforces physical consistency through monotonicity constraints. The model is evaluated on the standardised Hydronic Heat Pump test case provided by BOPTEST and compared to classical artificial neural networks and a physics-informed linear regression. The results demonstrate that incorporating monotonicity constraints improves training stability and extrapolation capabilities. This prevents catastrophic control failures that can be observed with unconstrained networks. The study demonstrates that model accuracy alone is not always a reliable indicator of control performance, emphasising the importance of application-oriented evaluation. While the standard CMNN performs similarly to a physics-informed linear regression with engineered non-linear features, the architecture’s flexibility offers a distinct advantage. Specifically, the CMNN architecture enables additional convex and concave model constraints, which can significantly outperform other models. Furthermore, the impact of online learning, training data coverageand different monotonicity constraints is investigated. Through this comprehensive evaluation, this study aims to advance the understanding and practical implementation of physics-constrained neural networks. Several further research directions are proposed, including complex use case exploration, extensive model comparison and alternative performance measures.