PREDICTION OF WORKSPACE THERMAL COMFORT IN UNIVERSITY BUILDINGS USING DEEP NEURAL NETWORK METHOD

• Main Author: Faniama Nurrahmayati, Virara
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/84384
• Institution: Institut Teknologi Bandung

Thermal comfort of buildings becomes very important in the implementation of energy conservation standards, as global warming can reduce the thermal comfort index. Thermal comfort is important in building design and management, especially in university environments as it has a significant effect on occupants' health, productivity and concentration. Future smart buildings will use artificial intelligence to maintain thermal comfort through prediction and adjustment of air conditioning operations. Various approaches have been taken to predict thermal comfort, but variations in comfort perception are still difficult to predict from a psychological, physiological, or experimental perspective. Data-driven thermal comfort prediction models are an effective approach to improve occupant comfort and energy savings. These data-driven models reveal the relationship between various factors to predict thermal comfort levels and can be optimized using neural networks. New opportunities to predict thermal comfort more accurately and efficiently are emerging as a result of advances in artificial intelligence technology as it can process various environmental parameters such as air temperature, humidity, air velocity, and thermal radiation and can also process other factors such as occupant activity levels and the type of clothing worn. This study aims to develop a prediction model of thermal comfort space using Deep Neural Network method by utilizing environmental data available on university buildings with the output of thermal comfort level prediction based on the pattern found. The research data collection method was obtained by taking direct measurements using measuring instruments. Room thermal data collection involved several important parameters including air temperature, relative humidity, air velocity, MRT, thermal preferences such as activity, clothing insulation, and measured PMV index to obtain environmental data along with Actual Mean vote validation. The evaluation parameters used are RMSE, MAE, and R² to predict the measured PMV. The model showed variable performance in predicting the actual PMV value and had excellent performance in predicting the calculated PMV value with the prediction curve always being close to the actual curve of Instrument PMV. The DNN model can learn from the data to calculate the PMV of the instrument. The best results are at the input of the important parameters of air temperature, humidity, MRT, air velocity, and activity, with RMSE = 0.0445, MAE = 0.0329, and R² = 0.9984. Simultaneously, the PMV validation data predicted by the model is very close to the actual PMV value. The graph shows a strong relationship between the PMV model instrument (Thermal comfort meter) and the PMV Actual Mean Vote (Observation) as evidenced by the high R² score (0.9430). This model evaluation shows that the PMV Actual Mean Vote is a good predictor of the Measured PMV model (Thermal comfort meter). This is also evidenced by the low MSE value of 0.083, although there are some outliers, the error is still acceptable. Statistical tests showed no significant difference between the means of the two data sets indicating that the measurement method through Actual Mean Vote (Observation) can be considered valid to replace direct measurement of PMV. The results of this study are expected to optimize air conditioning settings based on ASHRAE standard analysis and improve thermal comfort, make buildings more energy efficient, and assist in the development of intelligent building research in the future. Keywords: Thermal Comfort, Deep Neural Network, Smart Building, Actual Mean Vote ?