Forecasting summer-time overheating in UK homes using time series models

2019-12-02T12:42:03Z (GMT) by Matej Gustin
Heatwaves are projected to become more frequent, intense and long-lasting in the UK and the prevalence of overheating in dwellings is set to increase. As a result, occupants will experience increased levels of thermal discomfort, heat stress and heat-related morbidity and mortality. Since the use of mechanical air conditioning in dwellings is unsustainable, and not widely affordable, it is of utmost importance to understand when heat related health risks are anticipated in free-running dwellings. This is crucial for vulnerable occupants, such as the elderly, for whom the accurate detection of future heat risks could prepare them (or their carers) for timely mitigation, for example, through additional window ventilation or the use of shading. Many countries deploy Heat-Health Warning Systems (HHWS) to alert their populations, however, these generally apply to a wide area and are based exclusively on regional weather forecasts. Consequently, HHWSs are unable to identify where, when, or to what extent individual buildings (and their occupants) will be affected.
Previous studies have investigated the use of time series forecasting models, with the majority considering the use of Model Predictive Control. There is, however, no rigorous scientific evidence to support the belief that such models can provide accurate predictions in free-running dwellings during heatwaves and over multi-day forecasting horizons. This thesis therefore examines the use of black-box forecasting models to provide reliable predictions of the impending indoor temperatures in UK homes. Having established the viability of this approach, the application of such models in the context of an indoor Heat-Health Warning System (iHHWS) has been explored.
This research led to five main findings: (i) linear AutoRegressive forecasting models with eXogenous inputs (ARX), i.e. weather forecasts, can provide satisfactory accuracies during heatwaves for time horizons up to 72 h ahead; (ii) more complex semi-parametric Generalized Additive Models (GAMs) were not capable of significantly improving the forecasting accuracy at forecasting horizons over 6 h (iii) logistic GAMs can predict the window opening state with adequate discrimination, however, integration of the window state into forecasting models did not improve their accuracy; (iv) forecasting models could be usefully incorporated within an iHHWS, however, the warning lead-time should be constrained to less than 24 h in order to guarantee high confidence in such a system; (v) a weighted metric such as the Cumulative Heat Index (CHI) could further reduce the risks of false or missed warnings, increasing the dependability of the iHHWS.