Short term predictions of occupancy in commercial buildings—Performance analysis for stochastic models and machine learning approaches
Short term predictions of occupancy in commercial buildings—Performance analysis for stochastic models and machine learning approaches
复制标题
DOI：
10.1016/j.enbuild.2017.09.052
复制
发表时间：
2018
影响因子：
6.7
通讯作者：
Zhaoxuan Li;B. Dong
中科院分区：
文献类型：
--
作者：
Zhaoxuan Li;B. Dong
在不久的将来，实时入住率预测是智能建筑的重要组成部分。占用信息，如存在状态和居住者人数，允许对室内环境进行强有力的控制，以提高建筑的能源性能。目前的许多研究都集中在商业建筑的入住率上，大多数研究人员要么建立入住率模型，要么建立入住率模型，而没有对两者的模型潜力进行评估。本研究的重点是：1)提供一个独特的数据集，其中包含了位于美国不同模式品种的办公室的入住率；2)提出两种方法，然后与现有的四种方法进行比较；3)使用本研究提出的方法预测和测试入住率和入住率。在此基础上，提出了一种新的基于变化点分析的移动窗口非齐次马尔可夫模型。在现有模型的基础上，对分层概率抽样模型进行了改进。与先前研究人员的知名模型相比，它们是额外的。在不同时间情景下，包括提前15分钟、提前30分钟、提前1小时和提前24小时，进一步探讨和评估了模型的预测能力。最终结果表明，所提出的马尔可夫模型在15分钟、30分钟和1小时的存在预测方面优于其他方法，最大差异为22%。该马尔可夫模型在所有预测窗口的入住率预测方面也优于其他模型，RMSE和MAE误差分别为0.34和0.23。然而，在24小时前入住率预测模型之间没有太大的性能差异。
Real-time occupancy predictions are essential components for the smart buildings in the imminent future. The occupancy information, such as the presence states and the occupants’ number, allows a robust control of the indoor environment to enhance the building energy performances. With many current studies focusing on the commercial building occupancy, most researchers modeled either the occupancy presence or the occupants’ number without evaluating the model potentials on both of them. This study focuses on 1) providing a unique data set containing the occupancy for the offices located in the U.S with difference pattern varieties, 2) proposing two methods, then comparing them with four existing methods, and 3) both presence of occupancy and occupancy number are predicted and tested using the approaches proposed in this study. In detail, the paper develops a new moving-window inhomogeneous Markov model based on change point analysis. A hierarchical probability sampling model is modified based on existed models. They are additional compared to well-known models from previous researchers. The study further explores and evaluates the predictive power of the models by various temporal scenarios, including 15-min ahead, 30-min ahead, 1-h ahead, and 24-h ahead forecasts. The final results show that the proposed Markov model outperforms the other methods with a max 22% difference in terms of presence forecasts for 15-min, 30 min and 1-h ahead. The proposed Markov model also outperforms other models in occupancy number prediction for all forecast windows with 0.34 RMSE and 0.23 MAE error respectively. However, there is not much performance difference between models for 24-h ahead predictions of occupancy presence forecast.