Comparative analysis on predictability of natural ventilation rate based on machine learning algorithms

Abstract

The demand for efficient natural ventilation (NV) systems has increased for the development of sustainable buildings. However, the uncertainty of NV remains a challenging issue for appropriate utilization strategies of NV. For the successful implementation of NV systems in buildings, it is essential to clarify when and how to use NV systems in advance. In order to achieve the objectives, this study investigated the predictive models of NV rate (NVR) through eight machine learning (ML) algorithms, which are suitable for the interpretation of non-linear relationships between the measured indoor and outdoor environmental variables. Among all of the algorithms, deep neural network (DNN) ensured the best prediction performance for the NVR and it was shown that 40%, 46%, and 38% better predictive performance in terms of mean absolute error (MAE), root mean squared error (RMSE), and mean absolute percentage error (MAPE) than multivariate linear regression (MLR), which had the highest error rate, respectively. Based on the Shapley additive explanation (SHAP), the most influential features that affected to results of predictive models were examined and most of the ML approaches, except for MLR, had similar features (the pressure difference, outdoor temperature, wind speed, indoor relative humidity, solar radiation, the difference of indoor/outdoor relative humidity, and wind direction). The results of this study can improve the prediction performance for NVR, and this would contribute to the development of an intelligent NV system. Future work needs to develop the optimal operating strategies for hybrid ventilation systems integrating NV and mechanical systems.