| Title |
Evaluation of Summer Zone Temperature and Load Forecasting Performance of Transformer Architecture according to Training Dataset Size Change |
| DOI |
https://doi.org/10.5659/JAIK.2023.39.2.237 |
| Keywords |
Machine Learning; Deep Learning; Time-Series Forecasting; Model Predictive Control(MPC); Transformer; Attention Mechanism; Recurrent Neural Network(RNN); Long Short-Term Memory(LSTM); Multilayer Perceptron(MLP) |
| Abstract |
The data that fully reflects the dynamics of a building can only be collected after the building is completed. Therefore, the data for training
machine learning models are not sufficient at the operation stage of buildings. In addition, the dynamics of buildings and energy systems
frequently change due to age deterioration, commissioning, component replacement, and retrofitting. Thus, the retraining of deep learning
models to reflect the changed system dynamics is required. Therefore, the performance benchmark of deep learning architecture should be
designed in consideration of these specificities of the building-energy field. This study benchmarks the time-series forecasting performance of
three deep learning architectures: the multilayer perceptron (MLP) and long short-term memory (LSTM), which are widely used architectures,
and the transformer, which is relatively recently developed but has high potential. For reproducible benchmarks, a publicly accessible data
generator and the open-source Python library DeepTimeSeries was developed. The performance dependence according to the training dataset
size was evaluated by changing the training dataset size from 0.3 to 0.9 years. Forecasting targets were the zone air temperatures and
thermal loads. Among the three architectures, the transformer had the best performance. In particular, when the training dataset size was
small, the transformer exhibited better performance than other architectures in forecasting peaks and dips. Other architectures displayed
unstable performance when the training dataset size was small. The results suggest that the transformer has a high potential for time series
forecasting in the field of building energy, where the amount of data is limited in most cases. |