Comparison of different machine learning algorithms for predicting air-conditioning operating behavior in open-plan offices

Abstract

Air conditioning (AC) operating behavior has a significant influence on the energy consumption of buildings. Machine learning (ML) methods have recently been developed and have shown a high performance in predicting occupant behaviors. However, owing to a lack of comparison of different ML algorithms based on the same database and evaluation indicators, it is difficult to evaluate the positives and disadvantages of different ML algorithms in this research field. To address this gap, in this study, five ML algorithms (extreme gradient boosting, logistic regression, random forest, support vector classification, and k-nearest neighbor) were applied to predict the AC operating behavior under the same standard. The exhaustive method was applied to the five algorithms to detect the application of ML algorithms in different offices. The coupling relationship between the importance ranking method and ML algorithms was also discussed. The predicted results were evaluated using two indices, the open rate (OR) and F1 score. The results revealed that for all five algorithms in two different offices, a ΔOR within ± 5% and an F1 score of larger than 0.8 were achieved. Therefore, from the perspective of the two indices, the five algorithms showed similar performances, although they required different input numbers and parameter combinations. The results also showed that to reach the minimum abs(ΔOR) and the maximum F1 score, the corresponding ML algorithm and input parameters were inconsistent. When selecting the appropriate ML algorithm, determining the indicators that the simulation focuses on might be a key prerequisite. Influenced by the combination of parameters and algorithm chosen, the number of input parameters and model accuracy were not positively correlated. Meanwhile, for predicting the AC operating behaviors, it was difficult to find an important ranking method suitable for all algorithms. In addition, some frequent start-and-stop ACs were detected in the simulated AC on–off profiles, which indicated certain limitations of the ML algorithms.