Optimal control methods are a promising technology that may help to reduce the energy demand of domestic buildings. Paradoxically, the technical simplicity of some of the system components results in difficult optimal control problems. Heat generators and heat pumps, for example, often only accept on-off or discrete input values, which result in mixed integer programs. We tackle the issue by use of deep learning methods. Specifically, we train a neural network on optimal control policies acquired by hybrid model predictive control offline, in order to develop an implementable controller for lean embedded hardware. A case study reveals that the neural net is able to mimic the control performance of the predictive controller. The trained neural net was implemented on lean embedded hardware, where the evaluation could be performed in less than 7 ms.
Paper
Full text
Mimicking Predictive Control with Neural Networks in Domestic Heating Systems
Semantic Scholar · Engineering · 2019
Abstract
Optimal control methods are a promising technology that may help to reduce the energy demand of domestic buildings. Paradoxically, the technical simplicity of some of the system components results in difficult optimal control problems. Heat generators and heat pumps, for example, often only accept on-off or discrete input values, which result in mixed integer programs. We tackle the issue by use of deep learning methods. Specifically, we train a neural network on optimal control policies acquired by hybrid model predictive control offline, in order to develop an implementable controller for lean embedded hardware. A case study reveals that the neural net is able to mimic the control performance of the predictive controller. The trained neural net was implemented on lean embedded hardware, where the evaluation could be performed in less than 7 ms.