Conference Agenda

Employing data-driven models and machine learning techniques for predict thermal comfort of individuals is becoming increasingly popular. However, collecting sufficient quantity and quality of data to develop and train the models is challenging in real buildings. This work addresses this important challenge by introducing a novel meta-learning approach that leverages big data from the ASHRAE RP-884 database to train an artificial neural network (ANN) structure, and applies the Bayesian approach to encode domain knowledge for the unmeasured variables. The model uses general thermal comfort factors and personal thermal characteristics (considered as latent variables) as inputs and predicts thermal comfort of individuals. Using the pre-trained ANN structure and limited data collected from the target occupants, their personal thermal preference characteristics and the unmeasured variables are estimated using Bayesian inference and then used to predict their personal thermal comfort under variable indoor conditions. The results for various preference profiles showed that our approach outperforms existing methods, especially when using limited data, making it practical and data-efficient for realistic applications in buildings.

Human-centered daylighting operation remains a challenge due to (i) practical difficulties with real-time luminance monitoring within the human field of view (FOV) and (ii) the inability of common physical lighting parameters and scaled responses to characterize daylight preferences. This paper addresses both of these challenges by (i) introducing a novel deep-learning-based framework method to demonstrate that meaningful features in the human FOV can be extracted without invasive measurements and (ii) demonstrating a method for inferring personal daylight preferences using image pixelwise similarity analysis and deep learning.

For the first part, a Conditional Generative Adversarial Network (CGAN), pix2pix is used to transfer information from non-intrusive images to FOV images. Two datasets were collected with low-cost programmable cameras installed at two locations (a wall or a monitor) in a large office, to separately train two pix2pix models with the same target FOV images. The results show that the generated FOV images closely resemble the measured FOV images in terms of luminance errors and structural similarity. For the second part, we collected successive image-based luminance maps in the office space with simultaneous comparative preference responses. The dataset was converted to pairwise luminance similarity index maps which were used to train a convolutional neural network (CNN) model, able to classify and rank occupant’s visual preferences with great accuracy. This is the first study that demonstrates that it is possible to enable human-centered daylighting operation without intrusive luminance monitoring, by employing the full potential of image sensing and deep learning techniques.

Previous studies have proved that glazed spaces can serve as climate buffer zones for buildings by absorbing solar energy, potentially enhancing thermal and energy performance. However, in the high north and Arctic regions, where sunlight distribution is uneven throughout the year, the performance of the glazed spaces remains uncertain. This study aims to investigate the performance of the glazed balcony integrated with a room-based ventilation unit installed for buildings in the Arctic. It examines energy-saving potential, overheating risk, and hygrothermal behavior through field measurements and simulations. The findings indicate that the glazed balconies maintain higher temperatures than outdoor conditions all year round, with an average temperature difference of 2.4 °C. The thermal performance is particularly advantageous during the shoulder seasons (spring and autumn), when there is considerable heating demand, while sunlight is sufficiently available for utilization. Additionally, the study highlights the importance of ensuring adequate airtightness in the design to mitigate the risks of humidity and pollutant accumulation within the glazed space.

Daylight and solar irradiance simulations are crucial for evaluating natural lighting, visual comfort, and solar exposure on building envelopes. Traditional sky models, such as the CIE and Perez models, have limitations in accurately representing real-world conditions, especially in urban environments. As an alternative, Image-Based Lighting (IBL) has been proposed to enhance realism and prediction accuracy by using measured luminance distribution images. However, conventional IBL simulations often rely on the RADIANCE command-line interface (CLI), which poses usability challenges. This study introduces a user-friendly IBL simulation workflow within the Rhino–Grasshopper environment, ensuring accuracy while simplifying the process. The workflow facilitates large-scale simulations and enables the use of HDR images for realistic lighting analysis. Preliminary tests show that the proposed method is robust, aligning well with baseline simulations. Future work includes creating a 3D luminance distribution model using stereo vision and integrating it with visual comfort metrics and real-time lighting control systems.

Thermophysiological comfort models like Berkeley model, Fiala model, JOS-3 model, or 33-node comfort model (NOODEL) can be used to assess the thermal sensation and comfort in transient and asymmetric thermal environments. Those models calculate first the local body part temperatures within the physiological sub-model depending on the thermal boundary conditions. Then, the thermal sensation and comfort are calculated by the psychological sub-model depending on the body part temperatures. To generate reliable statements on thermal sensation and comfort, a precise physiological sub-model is necessary. Since high-quality experimental data are rare and are often used directly for calibration, mostly literature data are used for validation. However, the number of body parts of the models and the number of body parts on which temperatures were measured may differ. Therefore, mean skin temperatures are often used for a global validation of the models. Depending on the modelled or measured number of body parts and their weighting factors, their calculation can lead to very different results.

In this paper we analyze 16 different mean skin temperature calculation approaches for an experimental data set of 48 subjects which each were equipped with 25 PT100 resistance thermometers. The tests were performed in the Aachen Comfort Cube (ACCu) over 4 hours (1 h at 18 °C, 2 h at 28 °C, 1 h at 18°C).

The reference mean skin temperature is calculated by weighting the surface area of the body parts where the temperature sensors where located. Maximum deviations vary between 0.3 and 3 K. RMSE varies between 0.2 and 1.1 K. Based on the sensor number we also calculated optimized sensor positions and weighting factors to minimize the RMSE for the test data set. Results show that the validation and performance comparison of physiological sub-models, it is necessary to check how the mean skin temperature was determined.