Deciphering optimal mixed-mode ventilation in the tropics using reinforcement learning with explainable artificial intelligence


The application of mixed-mode ventilation (MMV) in the tropics is challenging, given its hot and humid climate. Consequently, there are limited periods when operating in natural ventilation (NV) is desirable. Furthermore, the potential to use NV diminishes at locations characterized by generally light winds. Given the complex interactions in MMV, this study aims to use reinforcement learning (RL) to identify relationships that can further reduce cooling energy use while maintaining occupant thermal comfort. The control variable variables considered include the percentage of window opening area and dynamic cooling setpoint temperature. Results show that RL can achieve a 52% reduction in cooling energy while maintaining good thermal comfort and indoor air quality compared to the existing baselines that typically involve switching between NV and air-conditioning based on outdoor conditions. We then developed an Explainable AI (XAI) framework to prompt building scientists toward new insights into MMV control in the tropics, which consists of shapley additive explanations (SHAP) and decision tree. The SHAP is able to improve the transparency of RL strategy by revealing the impact of each input on the final decision and the decision tree can extract the key control rules from RL. Using the XAI framework, this study also identified that the RL algorithm was taking advantage of the internal thermal mass to increase cooling efficiency. The extracted rules are further applied to the actual testbed, showing the feasibility of the rules extracted by XAI approach.

Energy and Buildings
Xilei Dai
Postdoctoral Fellow

My research interests include IoT for Building sensing and prediction.