The building sector consumes a large proportion of the total energy end-use. In densely built cities such as Hong Kong, more than 55% of the total energy and 90% of electricity is consumed by residential and commercial buildings. Solar energy is a popular renewable energy resource for electricity generation and thermal energy collection. Photovoltaic (PV) systems and daylight-linked lighting controls can contribute significantly to the reduction of building energy use. Solar radiation and daylight data are important for energy-efficient and environmentally friendly building designs. Owing to the unavailability of measurements, estimation models are necessary to obtain such data in some locations. This thesis proposes new modeling techniques and evaluation methods for solar irradiance and daylight illuminance based on long-term measurements and explores their applications in low/zero-energy buildings.

The first part reviews the modeling techniques for predicting global solar radiation and summarizes the current research gaps. Subsequently, it proposes an applicable and flexible approach for estimating the hourly global solar radiation on a horizontal surface based on routine measurements of various meteorological parameters. While previous studies mainly focus on monthly average daily or daily solar radiation, this study develops models for predicting hourly global solar radiation based on a machine-learning technique called multivariate adaptive regression spline (MARS). 16 MARS models were developed and the hourly meteorological data measured from 2010 to 2016 were used for training and testing the proposed models. Logistic regression and artificial neural networks (ANNs) were also used to develop models for comparative study. The results show that the MARS models perform well with regard to prediction accuracy and interpretability. It is necessary to extract the direct or diffuse component of global solar irradiance for different applications. Finally, the study analyzes the global, direct normal, and diffuse components of solar irradiance, and then evaluates three models for predicting direct normal solar irradiance (DNI) based on continuous measurements. The results would be helpful in selecting suitable DNI prediction models for various applications.

The second part focuses on the sky radiance/luminance distribution models which are crucial to the determination of sky-diffuse solar irradiance/daylight illuminance falling on any inclined surfaces, which are important for building energy and daylight simulations. At first, two widely used sky models, the Perez All-weather model and the ISO/CIE Standard General Sky model, were evaluated and compared. Model performance was assessed in terms of vertical global illuminance, indoor daylight illuminance, and lighting and cooling energy savings under corresponding lighting controls. Results show that although the performance of these two models is similar, the ISO/CIE model would be more convenient to use once the sky conditions have been identified. The crucial issue is how to accurately identify the sky conditions. Existing methods for classifying the skies commonly use sky radiance or luminance data, which are unavailable in most places. Thus, a machine-learning technique called Random Forest (RF) was used to correlate the identified sky types with 16 variables, including solar irradiance indices and other meteorological parameters. Four RFs were developed to classify the 15 Standard Skies. The performance of the proposed method was validated against a well-acknowledged model with new measurements of vertical solar irradiance and daylight illuminance data. The results show that the developed RFs can efficiently classify sky conditions and predict solar irradiance and illuminance data with high accuracy. These RFs only require horizontal measurements of solar irradiance and meteorological data that are readily available in most weather databases.

The third part proposes an efficient approach to determine vertical sky-diffuse solar irradiance and daylight illuminance on building façades under complex obstructed environments for various ISO/CIE Standard Skies. The work is based on a previous study which proposed a mathematical model to estimate the obstructed vertical sky component (OVSC) under various skies by assuming that the obstructions are infinitely long. However, more often, in actual cases, obstructions are irregular surfaces. Therefore, this study evaluates the OVSC under irregular skyline patterns. The possible obstructed zone was divided into fragments with the same angular interval. Various different configurations of irregular obstructions were identified, and the OVSCs were simulated using RADIANCE package. The simulated results were correlated with those of infinitely-long urban canyons so that the irregular obstructions could be described by the obstruction angle of equivalent infinitely-long cases. Both ANN-based models and simple regression models were proposed to correlate the OVSCs under various sky conditions. New irregular obstructions were generated for testing the proposed models. The R² values of the ANN model and regression models on the testing data set were above 0.99 and 0.96, respectively.

The fourth part discusses the applications of solar irradiance estimation in Photovoltaic (PV) energy output prediction and low/zero-energy building designs. PV panels are one of the most popular devices for harvesting solar energy and generating electricity. Based on measurements of a series of PV projects, a machine learning based empirical model was built to predict the PV energy output according to the weather data. The annual energy outputs of different PV installations were estimated based on the proposed empirical model. Finally, a school building was selected as a case study for the application of PV systems to achieve zero energy consumption.