Development of Occupancy Prediction Model and Performance Comparison According to the Recurrent Neural Network Models

An accurate occupancy prediction is essential for occupant-centric control (OCC) that saves energy while providing a comfortable indoor environment. Various machine learning-based approaches are being tried to develop an occupancy prediction model. Among these approaches, the performance of the recurrent neural network (RNN) based models, showed strength in time series forecasting and were found to be superb. However, studies related to performance comparison between RNN based models are insufficient; although the model performance had possibility for improvement through optimization. Therefore, in this study the RNN, long short-term memory (LSTM), and gated recurrent unit (GRU) models were developed to predict the number of occupants after 15, 30, and 60 minutes. The optimal models for each prediction horizon were derived through optimization and performance evaluation. As a result, the GRU model presented the best performance. The root mean squared error (RMSE) and mean absolute error (MAE) of the prediction model after 15 minutes was 0.8073, 1.5301, the prediction model after 30 minutes was 1.2841, 2.3386, and 2.0769, 3.3685, for the prediction model after 60 minutes. These results show superior performance compared to the existing RNN based models and signify that it is possible to provide accurate values for various prediction horizons. Thus, if outlier supplementation and addition of the adaptation function are implemented through an algorithm in the future, the developed models are expected to be utilized as a key element for OCC.