![](https://d3ilqtpdwi981i.cloudfront.net/bSNT_vnNP8E0w_USNaejvm44pPs=/425x550/smart/https://bepress-attached-resources.s3.amazonaws.com/uploads/47/6c/33/476c332b-60a0-46b7-96c9-9721b11b95f0/thumbnail_ec8d6fe0-0c60-441a-8eac-80293efb8770.jpg)
© 2018 The Authors. Published by Elsevier B.V. Model predictive controllers use accurate temperature forecasts to save energy by optimally controlling heating, ventilation and air conditioning equipment while achieving comfort for occupants. In a "smart" building, i.e. one that is outfitted with sensors, temperature forecasts are computed from data gathered by these sensors. Recently, accurate temperature forecasts have been generated using relatively few observations from each sensor. However, long sensor histories are available in smart houses. In this paper we consider improving forecast accuracy by using up to 24 hours of quarter-hourly readings. In particular, we overcome forecast inaccuracy that arises from the "one standard error" heuristic (1SE) in lasso regression. When there are many historical observations, low variance in the error estimations can result in excessively high values for the lasso hyperparameter λ. We propose the midfel refinement of lasso regression, which adjusts λ based on the shape of the error curve, resulting in improved forecast accuracy. We illustrate its effect in a setting where lasso regression is used to select sensors based on forecast accuracy. In this setting, midfel lasso regression using many historical observations has two effects: its improves accuracy and uses fewer sensors. Thus it potentially reduces costs arising both from energy usage and from sensor installation.
- Energy Efficiency,
- Feature Selection,
- Home Sensor Network,
- Internet of Things,
- Lasso regression,
- Model Predictive Control,
- Temperature Forecasting
Available at: http://works.bepress.com/omar-alfandi/10/