SUMMARY

Buildings are complex structures with dynamic loading and ever-changing usage. Additionally, the need to reduce unnecessary energy consumption in buildings is increasing. As a result, buildings and building energy systems should be designed to conserve energy, and buildings should be monitored and evaluated to ensure that the designs are executed properly and that the buildings are operated correctly. Most building designers now use very adequate energy modeling software such as EnergyPlus, IES, EQUEST, and others to support the design task. However, the problem with the current lineup of programs is that they require extensive inputs for material properties and usage loads; this results in spending extensive amounts of time performing model calibration or having to adjust multiple values (sometimes hundreds) to bring a model in alignment with actual building use. As a consequence, the existing software is complex and awkward for efficient monitoring and evaluation, especially for fault detection and diagnosis. Due to the limitations of current modeling programs, development has begun on rule-based and component-based fault detection by a number of companies. However, a suitable rigorous physics-based model has not been developed for the purpose of fault detection. Consequently, this thesis research will discuss the design, development, evaluation, and testing of a model-based fault detection program and procedure as well as comparisons to state-of-the-art neural networks. Considering how complex some buildings have become, it has become important to make sure the building systems are operating as intended. Some current progress is being done by the large energy service companies in the form of logic-based fault detection for individual components. While component-based fault detection is effective, it relies on accurate sensor readings and does not account for actual building performance. This research herein is the result of the development, testing, and refinement of a simplified but rigorous and complete physics-based model for buildings and building energy systems that is purposely designed and implemented to support fault detection and similar applications. The usefulness and effectiveness of this simplified physics-based model(SPBM) is demonstrated by comparison with the obvious currently available alternative, a state of the art purely data driven neural network black-box model. The models, a simplified physics-based energy model and a neural network, will evaluated total building performance using weather and minimal load data that is common to most buildings to determine, identify, and measure the impact of building faults. Evaluation of performance and accuracy of such a system to a state-of-the-art machine learning model provides substantial insight to current and future fault detection methods. https://bluejeans.com/747317657/5557