Abstract:
Heating, Ventilation, and Air Conditioning (HVAC) systems play a vital role in building
energy management by controlling the indoor environment and ensuring the occupant’s comfort.
These systems are responsible for regulating the temperature and air quality inside buildings,
thereby creating a comfortable and healthy indoor environment for occupants. However, the
energy consumption of HVACs contributes significantly towards overall energy usage of a
building and carbon footprint creating a challenge for building energy management. To address
this challenge, this research proposes the development of a predictive model for HVAC
temperature forecasting using Machine Learning (ML) algorithms to optimize energy efficiency
while maintaining thermal comfort in buildings. The study focuses on comparing the performance
of Transformer Neural Networks and CNN-LSTM, a seq2seq model combining Convolutional
Neural Networks (CNN) and Long-Short Term Memory (LSTM) on multiple forecasting horizons.
Both models are validated using data obtained from multiple devices which are deployed in a room
verified by feedback survey forms filled by occupants. The transformer model outperformed,
achieving an R2 score of 0.936 at a 1 minute forecasting horizon, surpassing the performance of
CNN-LSTM model at all tested forecasting horizons. The transformer model yielded significant
energy savings thereby reducing energy consumption by almost 50 percent compared to the nonAI conventional methods, particularly at forecasting horizons of 1 minute and 60 minutes, while
the occupant survey also favored a 60-minute forecasting horizon indicating that the proposed
model can effectively balance energy efficiency with occupant comfort. The performance of
transformer model particularly with a 60-minute forecasting horizon underscores its potential to
optimize energy efficiency while ensuring thermal comfort in building energy management
systems.
