On-line learning of indoor temperature forecasting models towards energy efficiency

• The monitoring system of the SMLsystem house is presented.
• On-line learning for estimation of indoor forecasting models is proposed.
• A comparison between Bayesian and gradient descent estimation techniques is shown.
• The major finding is that complex models show worst behavior than simple ones.
• The integration in real hardware devices with low resources is discussed.

The SMLsystem is a house built at the Universidad CEU Cardenal Herrera (CEU-UCH) to participate in the Solar Decathlon 2012 competition. Several technologies have been integrated to reduce power consumption. A predictive module, based on artificial neural networks (ANNs), has been developed using data acquired in Valencia. The module produces short-term forecast of indoor temperature, using as input data captured by a complex monitoring system. The system expects to reduce the power consumption related to Heating, Ventilation and Air Conditioning (HVAC) system, due to the following assumptions: the high power consumption for which HVAC is responsible (53.9% of the overall consumption); and the energy needed to maintain temperature is less than the energy required to lower/increase it. This paper studies the development viability of predictive systems for a totally unknown environment applying on-line learning techniques. The model parameters are estimated starting from a totally random model or from an unbiased a priori knowledge. These forecasting measures could allow the house to adapt itself to future temperature conditions by using home automation in an energy-efficient manner. Experimental results show reasonable forecasting accuracy with simple models, and in relatively short training time (4–5 days).