HVAC control in buildings using neural network

The energy consumption in buildings become the largest part of energy consumption worldwide, accounting for 40% of total global energy consumption and one third of the green house emission (Ahmed et al., Dec 2021). The optimization of building HVAC system require in many cases a thermodynamic model and mathematical solver which consumes a lot of hardware and time. However the data driven methods that presents 48% are often focusing on one type of HVAC systems (Grassi et al., May 2022). The real engineering problem is that the data driven methods are limited to specific systems and investigated for specific configurations. To overcome this problem, many machine learning solutions for optimized control data for HVAC are proposed. Those solutions are tested for real buildings in Germany. This new controller approach considers the effect of sequential inputs such as weather conditions, internal energy and time, and it consider a historical HVAC optimized data output. The investigation is done for CNN, LSTM, RNN-GRU, RNN-attention, and for each method we try to idealize the hyper-parameters and configurations. The goal of the study is to investigate the accuracy of machine learning routine on building optimization controlling and the ability to learn controlled and optimized data for different systems, to compare difference between generation of single output and multiple outputs, to decide the best inputs prepossessing, and to evaluate the extreme weather effects. This investigation proves that data control solution have limitation especially with extreme weather conditions but it can be improved by working on the pre-processing and different configurations.