A Comparison of Time Series Prediction Techniques for the Realization of a Smart Thermostat

The time series prediction problem aimed at the implementation of a 'smart thermostat' is addressed. 'SARIMAX models' are compared with LSTM neural networks. We show that with a low amount of data used for training, SARIMAX models achieve significantly higher accuracy while maintaining high computational efficiency, so in a problem where it becomes necessary to implement the system on low-power embedded devices, these approaches have significant advantages over neural networks.