In this paper we present a room-to-room thermal model used to accurately predict temperatures in residential buildings. We evaluate the accuracy of this model with ground truth data from four occupied family homes (two in the UK and two in the US). The homes have differing construction and a range of heating infrastructure (wall-mounted radiators, underfloor heating, and furnace-driven forced-air). Data was gathered using a network of simple and sparse (one per room) temperature sensors, a gas meter sensor, and an outdoor temperature sensor. We show that our model can predict future indoor temperature trends with a 90th percentile aggregate error between 0.61--1.50°, when given boiler or furnace actuation times and outdoor temperature forecasts. Two existing models were also implemented and then evaluated on our dataset alongside Matchstick. As a proof of concept, we used data from a previous control study to show that when Matchstick is used to predict temperatures (rather than assuming a preset linear heating rate) the possible gas savings increase by up to 3%.

For locating a set of stationary devices, algorithms such as MDS-MAP have been favoured by the sensor network community. This is generally because of their low computational complexity. Whilst comparisons for complexity and performance have been done for other algorithms, non-linear regression (NLR) has been neglected. The authors find that it is not much more expensive than MDS-MAP, yet can yield significantly better accuracy for sensor network localisation.