With the rise of cheap sensors and microcomputers, the proliferation of smart power meters, and the reemergence of many machine learning tools, there is an unprecedented opportunity to understand energy-use in buildings. To process all of this data, we are seeking funds to build a low-cost and low-energy computing cluster. With this computing power, we will work on the next generation of control systems for building energy-use management.
Every commercial and residential building exhibits unique energy-use patterns according to its design, climate, mechanical systems, occupancy patterns, etc. As a result, there is no “one-size-fits-all” solution to building energy management. An advanced control system must process sensory data and adapt to the characteristics of a building in real-time. In short, a building control system must be capable of learning. There is a wealth of machine learning algorithms that are well suited to these kinds of tasks, but figuring out when to use which tool or how to tune it according to the data is a computationally demanding task. To address this, we are researching the use of a parallel computing cluster to simulate and solve these building control questions.
Building energy use is an ever-evolving field. The adoption of technologies like electric vehicles, internet-connected devices, and renewable energy introduces interesting questions, such as: Should you use power from your solar panels to charge your vehicle or sell it to the grid? Which option saves money and which reduces environmental impact? Can a thermostat estimate and control the temperature in each room? Can a population of refrigerators be used as grid storage, making it easier to integrate renewable energy? Independently, these questions may have a small impact on the energy use of a single building. However, the additive effect at the grid level has the potential to significantly change the way we use energy and to foster the adoption of renewable technologies.
Our goals are to build a computing cluster designed for building energy use experiments. Simulating our prediction and control algorithms often involves solving the same problem numerous times under slightly varied conditions. This not only helps us to understand the robustness of the algorithm, but to begin answering questions, such as: How important is forecasting accuracy? How far into the future do we need to predict? Can sensor inputs (like motion and temperature) improve our energy use forecasts? A computing cluster will be a valuable resource for processing large data sets and for testing new ideas. We will be able to test multiple forecasting algorithms and compare their accuracy in real time.
It is really exciting and challenging to develop an intelligent control system as we propose to build. At present, the major constraint that we are facing is access to the computing power needed to process years of data and to simulate the behavior of our control algorithms. While cloud services such, as AWS is a possible solution to this hurdle, they require constant funding. Also, these programs are not really compatible with real-time experiments with sensory input. An interesting alternative to this is the advent of low-cost microcomputers that offer the opportunity of creating a high-performance parallel computing cluster. Provided you support us, we can expand our 5node cluster into a 25node computing cluster using BeagleBone Black development boards. Furthermore, we will be able to produce a beginner-friendly, complete guide of how to build this system. This will help other researchers to learn from our experience.