Room occupancy prediction leveraging LSTM - An approach for cognitive and self-adapting buildings

Energy consumption of heating, cooling, ventilation, lighting, and appli- ances, is deeply influenced by human presence in buildings. Accurate room occu- pancy prediction is a key to making buildings cognitive and self-adapting in order to achieve energy efficiency and wastage cut. Instead of using cameras or human tracking devices, a predictive model based on indoor non-intrusive environmen- tal sensors allows mitigating privacy concerns. In such direction, this study aims to develop a data-driven model for occupancy prediction using machine learning techniques based on a combination of temperature, humidity, CO2 concentration, light, and motion sensors. The approach has been designed and realized in a real scenario by leveraging the COGITO platform. The experimental results show that the proposed Long Short-Term Memory neural network is well suited to account for occupancy detection at the current state and occupancy prediction at the future state, respectively, with an overall detection rate of 99,5% and 92,6% on a literature dataset and 99,6% and 94,2% on a real scenario. These outcomes indicate the ability of the proposed model to monitor the occupancy information of spaces both in a real-time and in a short-term way.