![](https://d3ilqtpdwi981i.cloudfront.net/U-2gkF7pg9q4fBgl8q30eqRPzRE=/425x550/smart/https://bepress-attached-resources.s3.amazonaws.com/uploads/05/e3/8c/05e38ce6-5523-4a7a-92ec-25fca93b874e/thumbnail_99bfa4ba-90b7-4900-9403-581820c41456.jpg)
© 2016 The Authors. We forecast internal temperature in a home with sensors, modeled as a linear function of recent sensor values. When delivering forecasts as a service, two desirable properties are that forecasts have stable accuracy over a variety of forecast horizons - so service levels can be predicted - and that the forecasts rely on a modest amount of sensor history - so forecasting can be restarted soon after any data outage due to, for example, sensor failure. From a publicly available data set, we show that sensor values over the past one or two hours are sufficient to meet these demands. A standard machine learning method based on forward stepwise linear regression with cross validation gives forecasts whose out-of-sample errors increase slowly as the forecast horizon increases, and that are accurate to within one fifth of a degree C over three hours, and to within about one half degree C over six hours, based on one or two hours of history. Previous results from this data achieved errors within one degree C over three hours based on five days of history.
- domotic house,
- forecast accuracy,
- forward stepwise linear regression,
- service level agreement,
- Temperature forecasts
Available at: http://works.bepress.com/feras-al-obeidat/47/