Data mining for residential buildings using smart wifi thermostats

Smart WiFi thermostats are not just a device for controlling heating and cooling comfort in buildings, they also can learn from occupant behaviors and permit occupants to control their comfort remotely. This research seeks to go beyond this state of the art by utilizing smart thermostat WiFi data from detached residences combined with outdoor condition data to develop dynamic models to predict room temperature and cooling/heating demand and then apply these models to new thermostat temperature scheduling scenarios, associated with lower energy cooling/heating. The ultimate objective of this effort is to reduce energy use in residences and demonstrate the ability to respond to peak utility demand events while maintaining thermal comfort within a minimally acceptable range. Back Propagation Neural Network (BPNN), Long-Short Term Memory (LSTM) and Encoder-Decoder LSTM approaches are used to develop these dynamic models. Results demonstrate that LSTM outperforms BPNN and Encoder-Decoder LSTM approach, yielding MAE value on testing data of less than 0.5oC, equal to the resolution error of the measured temperature and MAPE value on testing data of 0.64. Additionally, the models developed are shown to be highly accurate in predicting energy savings from aggressive vithermostat setpoint schedules aimed at yielding deep reduction (up to 14.3%) in heating and cooling energy, as well as energy reduction that cooling or heating could be curtailed in response to a high demand event while maintaining thermal comfort band