We present a comprehensive study of state-of-the-art algorithms for the prediction of sensor events and activities of daily living in smart homes. Data have been collected from eight smart homes with real users and 13-17 binary sensors each - including motion, magnetic, and power sensors. We apply two probabilistic methods, namely Sequence Prediction via Enhanced Episode Discovery and Active LeZi, as well as Long Short-Term Memory Recurrent Neural Network, in order to predict the next sensor event in a sequence. We compare these with respect to the required number of preceding sensor events to predict the next, the necessary amount of data to achieve good accuracy and convergence, as well as varying the number of sensors in the dataset. The best-performing method is further improved by including information on the time of occurrence to predict the next sensor event only, and in addition to predict both the next sensor event and the mean time of occurrence in the same model. Subsequently, we apply transfer learning across apartments to investigate its applicability, advantages, and limitations for this setup. Our best implementation achieved an accuracy of 77-87% for predicting the next sensor event, and an accuracy of 73-83% when predicting both the next sensor event and the mean time elapsed to the next sensor event. Finally, we investigate the performance of predicting daily living activities derived from the sensor events. We can predict activities with an accuracy of 61-90%, depending on the apartment.
This item's license is: Attribution 4.0 International