Data Predictive Control (DPC)

Model Predictive Control (MPC) is an optimal control algorithm based on a mathematical model of the plant (i.e. the building) and mathematical programming. The model is used to predict the system’s future state depending on the current state, future control inputs and measurable disturbances. According to these predictions, optimal control actions for a limited time horizon are calculated while only the first input is applied to the plant, see Figure below.

M. Zeilinger. Seminar 1: Model Predictive Control. Lecture notes. ETH Zurich. 2018

Applied to a building heating or cooling system the algorithm aims at reducing the energy consumption while preserving or even increasing comfort. Even though MPC has been demonstrated to be a powerful approach to reduce energy consumption of buildings there is no widespread application in the residential building domain so far. An obstacle might be the effort & costs required to identify and maintain the mathematical models of the building used in the MPC framework. As almost each building is unique this process needs to be done for every building individually. However, the number of sensors in buildings increases steadily due to the digital revolution. Therefore, more and more historical data of buildings become available. By that, Machine Learning algorithms capturing the building dynamics become more powerful and expressive. As the name Machine Learning already suggests the model generation is done by a machine (i.e. a computer). Hence, the effort to generate and update thermal dynamics models becomes really small compared to first-principle models. Data Predictive Control combines the advantages provided by Machine Learning and MPC by replacing the first-principle based model in the MPC framework with a Machine Learning model trained on historical building data, see Figure below.

MPC vs DPC

In the NEST unit Urban Mining & Recycling (UMAR) we applied a version of DPC based on Random Forests combined with linear models and compared the performance with a conventional hysteresis controller during cooling season. In these experiments the DPC outperformed the hysteresis controller not only in energy savings but also in terms of comfort. The room controlled by DPC consumed about 25% less cooling energy while the comfort was increased by up to 72%

© Zooey Braun, Stuttgart

For more detailed information you can access our article published in the journal Energy & Buildings.