Prediction of User-defined Room Temperature Setpoints by LSTM Neural Networks

Buildings are one of the largest consumers of energy in the world and, as such, an opportunity for significant energy savings. Modelling of user behavior is a prerequisite for predicting the building energy consumption. It offers a possibility to adapt the building operation to best suit the user requirements while achieving minimal energy consumption through room air conditioning. This paper presents a deep learning model for prediction of an individual room temperature setpoint provided by the user. Model learning is conducted on the available dataset consisting of external weather and internal climate measurements. External weather conditions include the air temperature, relative humidity and direct, diffuse and global solar irradiation. The internal conditions refer to four offices for which the air temperature, the mode of operation of the fan coil controller and the temperature setpoints are available. The model is based on a neural network with Long Short-Term Memory. The highest accuracy is achieved with the model developed for the inputs of historical setpoints, system operation, day of the week and relative humidity. Presented results show high model accuracy, applicable for real-time model-based predictive building operation.