Short Term Forecasts of Internal Temperature with Stable Accuracy in Smart Homes

摘要

We forecast internal temperature in two homes, using variants of regression with data from the readings of multiple sensors. We use 48 separate models, where each forecasts mean temperatures that will occur in one future 15-minute interval, to compose a forecast for the next 12 hours. The sensors report internal and external atmospheric and environmental conditions such as temperature, pressure, sunlight, rain and wind, as well as evidence of human activity, including CO2 saturation, motion sensors and electrical load from areas within the house and large appliances. The models use both current and historical sensor values, each of which increases the number of predictors in the linear regression model. We use model simplification techniques including forward stepwise regression, principal component regression, and partial least squares regression. In both houses the forecast accuracy is stable; the mean absolute error over 12 hours is less than 1, while the root mean squared error is less than 1.3. Our accuracy compares favorably to previous work. Our work indicates long sensor histories for forecasts in the next 12 hours do not significantly improve accuracy.