Autonomous Control of Indoor Climate for Commercial Buildings

点击翻译按钮获取中文摘要

Buildings account for 45 percent of the total energy consumption in the United States (U.S.), and maintaining indoor climate, which includes heating, cooling, and ventilation, accounts for approximately half of that energy consumption. A low-cost option for reducing building energy usage is intelligent climate control, moving away from the prevalent "design for steady-state conditions" philosophy into one that exploits the constantly changing conditions a building operates in due to its occupants and the weather. The potential for intelligent climate control has been recognized for many years, especially for commercial buildings that have the requisite sensors and actuators. In particular, control algorithms that make decisions using real-time optimization have been shown to be highly promising. In spite of its promise, such "model-optimization" based control technologies have not been widely adopted by industry. The reason for this lack of translation to practice is the lack of autonomy of existing algorithms. Not only do they require expert human involvement in model creation, which have to be tuned for each building manually, they do not provide guarantees about the quality of real-time decisions. Addressing these weaknesses will lead to the wider adoption of intelligent building climate control technologies, which will contribute to the technological edge U.S. industries enjoy, and reduce the nation's energy usage.This research project seeks to make model+optimization based control of commercial buildings autonomous, thereby aiding wider adoption of such advanced technologies. The approach is to engineer both the modeling and optimization phases specifically for autonomy. The modeling approach is purely data driven so that it can be easily applied to any building. By using recently developed machine learning methods that guarantee certain beneficial model properties (e.g., stability), models can be updated over time purely from data without ever requiring a human expert to check the quality or suitability of the models. Similarly, the optimization problem is made convex by a choice of linear models so that real-time decision making can occur reliably without the optimizer getting stuck in a local minima or failing to converge. The reduction in accuracy due to the restriction to linear models is ameliorated by re-learning models over time, which is made possible by the autonomous data-driven nature of the model fitting algorithms. Finally, special care is taken to ensure that humidity and latent heat considerations, which are critical to hot humid climates, are taken into account both in the modeling and real-time optimization phases.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.