Mathematical approaches for improved thermal performance in buildings, simulations, data mining and neural networks

Thermal operations in residential and commercial buildings account for a significant proportion of the primary energy consumption within both Europe and the United States. Owing to the myriad of factors influencing building energy demand, imple menting optimal designs or retro-fitting actions is non-intuitive. Hence, numerical simulations have become popular as decision-making tools in such instances. The large physical scale of buildings along with extended simulation time frames necessi tate compact modelling approaches; simplifying thermal analysis to manage compu tational costs. The trade-off between simulation accuracy and modelling complexity is a product of the discretisation methodology being employed. However, the optimal means of applying discretisation remains unclear. During the course of this thesis a metric is introduced to quantify discretisation effects which is based on energy storage predictions. Importantly, this approach allows for spatial and temporal discretisation effects to be addressed and reported separately. Results are presented in terms of gov erning dimensionless parameters and so are equally applicable to all materials and physical scales. Firstly, guidance on optimal discretisation levels and simulation time steps is presented for linearly-spaced discretisation schemes. Thereafter, an optimised loga rithmic distribution of elements is introduced, yielding up to a fivefold reduction in the number of elements required to achieve desired levels of accuracy when com pared to linearly-spaced elements. The results are then extended to multi-layer walls by introducing a universal layer-by-layer discretisation approach and an effective Biot number calculation procedure. This allows for the relative position of layers to be considered when determining discretisation levels. Finally, guidance is presented for achieving accurate simulation in the widely used software EnergyPlus by selecting optimal discretisation parameters. EnergyPlus-based simulations are then conducted, demonstrating that the reported guidance provides efficient models with excellent prediction accuracies. Many buildings are currently controlled by building automation systems. These systems deploy a multitude of sensors to provide feedback to controllers so that com fortable conditions are maintained. In the process, vast amounts of data about the operation of the buildings is generated, however, much of this currently goes unused. Within this thesis, data mining of building automation system data is undertaken for the purpose of knowledge discovery. Graphical, statistical and machine learning approaches are employed to uncover potential means of improving energy efficiency. Important control settings within such smart automation systems are selected by facility managers or home-owners. These control settings significantly affect sys tem energy efficiency and thermal comfort levels within a building. The potential for improving energy efficiency through optimising these control parameters is well re ported. However, it is difficult to translate potential efficiencies into practice without continuously educating facility managers/users. This thesis presents an application of artificial intelligence to address this issue. The results of the research show that an artificial neural network is capable of learning to autonomously control a full build ing energy system. When compared to a typical human control strategy, the neural network is observed to achieved an approximate 20% reduction in operating costs while maintaining the desired zone temperatures for thermal comfort.