Buildings alone are responsible for almost 20% of the global greenhouse gas emissions and use almost 40% of the global energy. However, reliable and energy-efficient operation of building energy systems represents an unsolved real-world problem, which requires multidisciplinary expertise and poses a multitude of practical challenges. In this talk, we present a novel physics-constrained deep learning method for modeling and control of building energy systems that overcomes many of the obstacles of classical physics-based and purely data-driven methods. By combining physics and learning, we demonstrate scalability, reliability, and state-of-the-art performance while tackling implementation challenges associated with the high required expertise and computational costs.