Application of Machine Learning for Energy-Efficient Buildings

With buildings accounting for about 40% of the energy consumed in the United States, the last years have seen an increasing effort in attaining greater energy efficiency in their operations. Several solutions have been proposed to this end, namely, demand response (DR), precooling or preheating, optimal supervisory control of underlying systems such as heating ventilation and air-conditioning (HVAC), and on-site renewables. These solutions demand complex mathematical models in which many factors and their effects are intertwined: set-point temperatures, control systems, building layout, and weather, to name a few. Frequently, modelers describe system dynamics with the use of physics-based models. Such is the case of the widely used EnergyPlus modeling tool. Unfortunately, system dynamics and physics-based models involve the solution of equations that contain parameters that are building-specific, such as building materials and heat transfer constants. This task requires a considerable effort to gather the information needed, both in terms of time and money. In these circumstances, the use of information technology (IT) in buildings nowadays enhances the availability of data and provides an alternative to physics-based models. Consequently, data-driven modeling tools like linear regression, artificial neural networks, and support vector regression are becoming increasingly popular options in this modeling domain. Proper solutions that aim at improving cost efficiency in building operations demand that the models employed are control-oriented. These models must be able to quantify the energy or its cost as a function of a series of control knobs. These can be endogenous (ON/OFF status of devices, set-point temperatures, etc.), or exogenous (building occupancy, weather, etc.). In addition, it is desirable that these models require the lowest amount of data preprocessing, have good predictive ability, and can be updated frequently as more data become available. In this regard, the surge in data availability as well as in computing power has positioned deep learning techniques such as recurrent neural networks (RNNs) as a powerful choice. Albeit RNNs have been applied in this context before, there is much more work to do for fully untapping these tool’s potential for providing accurate control-oriented models. Thus, we aim at demonstrating that these machine learning tools can be included in models that satisfy the aforementioned requirements and enable multiple control use cases. We test our approach with data from a real building and we show that we can outperform other data-driven modeling techniques with errors that are 8.5–52% lower. In addition, we analyze different widths and depths in our RNNs as well as their sensitivity to the outside temperature. A typical prediction-related use case when it comes to showing the effectiveness of a developed controller is the estimation of baseline energy consumption. This scenario is defined as the energy consumption "business as usual" or, in other words, before any implementation that pertains to design, operational, or control improvements. Estimating energy consumption in these circumstances is important, as it helps measure the impact of design retrofits or updates performed to the control systems as a consequence of some sustainability or cost considerations. This estimation, usually referred to Measurement and Verification (M&V), also allows to analyze buildings for participation in grid services in the context of building-to-grid integration. Better understanding of energy consumption in baseline scenarios is beneficial for designing more informed DR programs and to assess their efficacy. In this work, we address the long-term baseline energy prediction problem via a sequential deep neural network (DNN)-based framework. This framework consists of two deep network-based architectures that relate future baseline consumption predictions with past measurements, building zone temperatures, and outside weather conditions. These architectures employ convolution and max pooling layers that allow the extraction of lower-dimensional features. Finally, the last layer consists of tensor train-based gated recurrent unit (GRU) cells that memorize those lower-dimensional features. Also, these cells reduce the total computation time during training, which makes this architecture convenient for future in-field deployment. In addition, a sequential architecture favors the mitigation of prediction error accumulation in the long term. In summary, we evaluate the proposed network on a simulated commercial building dataset. Our approach results in (i) a novel architecture, (ii) more efficient computation time, and (iii) high accuracy in long-term energy prediction scenarios.