On the use of Deep Learning Approaches for Occupancy prediction in Energy Efficient Buildings

Abstract

Occupancy forecasting is considered as a crucial input for improving the performance of predictive control strategies in energy efficient buildings. In fact, accurate occupancy forecast is the key enabler for context-drive control of active systems (e.g. heating, ventilation, and lighting). This paper focuses on forecasting occupants' number using real-time measurements of CO2 concentration and its forecasting values. The main aim is to evaluate the accuracy of forecasting occupants' number by applying the steady state model (1) [16] on the CO2 forecast using recent deep learning approaches. The LSTM, a recurrent neural network based deep learning algorithm, is deployed to forecast the CO2 level in a dedicated space, a testlab deployed in our university for conducting experiments and assess approaches for energy efficiency in buildings. Preliminary results show the effectiveness of LSTM in forecasting occupants' number, which reaches 70% in accuracy.