Development and integration of stochastic occupant behaviour models in residential building performance simulation and rating tools

Energy policies and regulations play a critical role in reducing energy consumption and carbon emissions in the building sector. One of the widely used methods to evaluate the compliance of building designs with energy regulations is to employ Building Performance Simulations (BPS) to estimate their energy demand. BPS tools need several inputs to simulate how the building will perform, and often assumptions must be made for those variables that are difficult to measure. Occupant behaviour is generally one of those challenging inputs due to the stochastic nature of human behaviour and the difficulty of measuring it, which contributes to uncertainty in BPS results. The user-defined simulation assumptions might over-simplify the realistic occupant behaviour, leading to potentially substantial gaps between real and simulated energy performance.

This thesis first identified research gaps in the field by reviewing the literature on occupant behaviour and building energy performance, including the absence of a clear method to represent the variability in heating and cooling thermostat setting preferences in building simulation tools for building energy rating purposes, the scarcity of equivalent comparisons among different existing window state models for representing window operation state of individual and cohort households and their impact on simulated households’ heating and cooling energy consumption with respect to the current deterministic approaches used in household energy rating methodologies. Following these identified research gaps, the thesis proposed to investigate the stochastic modelling of the following two of impactful user behaviours from the perspective of energy rating assessments (1) heating and cooling thermostat setting preferences and (2) window opening behaviour of occupants in terms of their capacity to produce realistic occupant preferences and practices.

To represent thermostat preferences, two occupant behaviour modelling methods were developed and assessed using experimental datasets collected from 102 households located in three different Australian climate regions. The first method, with a stochastic approach, used the probability distribution of heating and cooling thermostat setpoints to run multiple simulations, while the second method, with a single-setpoint deterministic approach, used weighted averaged values from the same distribution to run a single simulation. Both methods were benchmarked against a baseline simulation, with embedded uncertainty in the user behaviour in the setpoint choice. While both methods equally accurately estimated the resulting heating demands, the stochastic method performed better in estimating the cooling demand, revealing that the deterministic method failed to capture the effects of non-linearities in cooling, which cannot correspond with an average behaviour.

To analyse occupants’ behaviour in operating windows, this thesis developed a set of stochastic window operation models using seven well-established state-of-the-art approaches: Gaussian distribution function, logistic regression, Markov chain, Markov-logit hybrid, classification tree, random forest and artificial neural network. These models were trained and tested using operational data from 50 households in two different Australian climate regions. Models were developed to represent (1) each window individually and (2) a cohort of similarly operating windows (e.g. living room windows) before comparing the model performances against monitored data. While artificial neural network and Markov-logit hybrid models performed well in terms of generating higher accuracy values compared to the rest of the models when trained and used on an individual window, random forest and classification tree models were better when employed as cohort models to represent a population of households. The results indicated that the model selection for BPS should not be made out of convenience but on the suitability of a model for a specific objective (individual/cohort modelling).

To evaluate the impact of these models on energy rating, the selected better-performing stochastic window operation models were then integrated into the BPS tool used for energy rating in Australia (AccuRate). The impact of the stochastic window operation models on the overall energy simulation was quantified against the default deterministic rule-based model, using eight apartment building models as case studies. The thesis also examined the model transferability between climates: this was achieved by training the stochastic models with data from other houses located in a climate different from the one in which the simulated buildings were located. The classification tree and random forest cohort models trained with data from the same climate were the best at reproducing window opening patterns distributed similarly to the monitored buildings. Nevertheless, the transferred models performed still better than the default deterministic rule-based model of the rating tool, even though their performance was slightly lower than the locally trained models.

Employing stochastic window operation models was found to have a significant impact on the estimated buildings’ energy consumption and rating, with an estimated 71% - 480% higher cooling energy consumption than the default rule-based model. The reason for this was attributed to a higher cooling demand due to more sub-optimal use of passive cooling by the realistic user behaviour.

The thesis also explored the possibility of finding a fixed threshold to make the binary decision of ‘open’ or ‘close’ state for the stochastic window operation models that would enable a more realistic deterministic one-time simulation that provides similar energy consumption and rating results to the stochastic one but with less time and effort. It was possible to achieve this with the classification tree (fixed threshold of 0.26) and random forest (fixed threshold of 0.4) models, with energy consumption and rating estimations being very close to the median value of the corresponding stochastic model results.

Overall, the findings of this work provided evidence that data-driven, stochastic occupant behaviour models improve the results of Building Performance Simulation tools, and their use would positively impact the development of new-generation residential building energy rating tools.