How Many Sensors are Necessary?—Sensor Reduction in HVAC Control Modeling Using Feature Attribution and Rule Extraction

Abstract

Reinforcement learning (RL) has become a popular approach for process control design, especially for systems that are complex and not fully understood. The opaquenature of neural networks makes it tempting to use a variety of sensor data as input variables. In case of doubt, additional sensors are often introduced under the assumption that more sensors will also increase the accuracy of the model and, therefore, the control performance. However, this is not necessarily true, and the lack of interpretabilitycomplicates the identification of which sensors influence decision-making and are thus essential for the control function, and which are less important. Gaining insight into the significance of particular sensors allows for dimensionality reduction by fully removing less relevant sensors without compromising the system's performance. Investigating the example of heating, ventilation, and air conditioning (HVAC) systems, we suggest two approaches: rule extraction and feature attribution to identify key sensor inputs that are most relevant for optimizing control performance. In addition, when using rule extraction, we translate RL models into rule-based systems compatible with existing HVAC setups. This approach adds explainability and, in the given example, was able to reduce the required number of sensors by more than 70%.