A Method For Real-time Prediction Of Building Occupancy Based On Hidden Markov-random Forest Model

Accurate occupancy prediction is of great significance to optimize the control of building services systems and to reduce building energy consumption.In current research,the Hidden Markov model has been widely used to predict the number of occupants.In this model,the input environmental parameters were determined by feature selection.Moreover,an underlying assumption was made that the variation in the input environmental parameters accord with the gaussian distribution.However,the actual relationship is changing over time and cannot be described by a single distribution.Besides,existing feature selection method still have some limitations.To address this issue,a filter-wrapper method was first proposed to select environmental parameters as model input parameters.On this basis,a Hidden Markov-Random Forest model was proposed to predict the number of occupants.In the proposed model,to overcome the limitation of the gaussian distribution assumption,the relationship between the number of occupants and environmental parameters was determined directly from data based on the random forest method.To validate the proposed model,the model was applied based on the experimental data from an office room in Hunan University.Then the proposed model was compared with the Hidden Markov model,Classification and Regression Trees Model,Support Vector Machine Model and Artificial Neural Network Model.Finally,the in-depth analysis of model influencing factors was carried out.The main results and conclusions are summarized as follows:(1)Compared with the model results based on the filter method and the wrapper method,the simulated occupancy profile based on the filter-wrapper method is closer to the actual profile.Moreover,the delay time in the simulation result can be significantly reduced.This indicates that the filter-wrapper method can further improve the model prediction performance.(2)The proposed model achieves superior prediction performance over the other four models.Moreover,a comparation with the other four models showed that the proposed model can predict the occupancy more accurately,in terms of estimation accuracy,mean of absolute error and detection accuracy of presence/absence.(3)Among the influencing factors,the simulation time step and the time parameter have a great influence on the simulation performance.Results showed that the proposed model performed best when the simulation step was set as 10 minutes,and the prediction accuracy was increased by 4% when the time parameter was added to the proposed model.Compared with the above two factors,the forgetting factor and the number of decision trees has a smaller influence on the model.In order to further improve the performance of the proposed model,future studies need to focus on the selection of variable steps which are utilized in calculating dynamic occupancy state transition probability,the update of model input parameters and the research of variable forgetting factors.