Decision-Focused Learning for Neural Network-Constrained HVAC Scheduling

TL;DR

This paper introduces a novel decision-focused learning approach for HVAC scheduling that embeds neural network models of thermal dynamics into an optimization framework, improving cost efficiency and grid services.

Contribution

It develops a stochastic smoothing technique to enable gradient-based training of neural networks within mixed-integer quadratic programs for HVAC control, overcoming previous gradient discontinuity issues.

Findings

Outperforms traditional identify-then-optimize methods in cost savings.

Achieves better grid service performance in simulations.

Demonstrates scalability for real-world building control applications.

Abstract

Heating, Ventilation, and Air Conditioning (HVAC) is a major electricity end-use with a substantial potential for providing grid services, such as demand response. Harnessing this flexibility requires accurate modeling of the thermal dynamics of buildings, a difficult task because nonlinear heat transfer and recurring daily cycles make historical data highly correlated and insufficient to generalize to new weather, occupancy, and control scenarios. This paper presents an HVAC management system formulated as a Mixed Integer Quadratic Program (MIQP), where Neural Network (NN) models of thermal dynamics are embedded as exact mixed-integer linear constraints. Unlike traditional training approaches that minimize prediction errors, we employ Decision-Focused Learning (DFL) to learn the NN parameters with the objective of directly improving the HVAC cost performance. However, the discrete nature of MIQP hinders DFL, as it leads to undefined and discontinuous gradients, thus impeding standard gradient-based training. We leverage Stochastic Smoothing (SS) to enable efficient gradient computation without the need to differentiate the MIQP. Experiments on a realistic five-zone building using a high-fidelity simulator demonstrate that the proposed SS-DFL approach outperforms conventional identify-then-optimize (i.e., the thermal dynamics model is identified on historical data then used in optimization) and relaxed DFL methods in both cost savings and grid service performance, highlighting its potential for scalable, grid-aware building control.