Performance evaluation of deep learning architectures for load and temperature forecasting under dataset size constraints and seasonality

Abstract

Buildings and their energy systems have unique characteristics and temporally changing dynamics. Additionally, building data have strong seasonality. Therefore, a deep learning model with good adaptability (i.e., high performance with little data) is necessary for model predictive control of buildings. Considering the specificity of the building-energy field, we benchmarked six deep learning architectures: the multilayer perceptron, simple recurrent neural network (RNN), long short-term memory, gated recurrent unit (GRU), dilated convolutional neural network, and transformer. Moreover, the data similarity analysis method was developed to analyze the effect of data seasonality on forecasting performance. A publicly accessible data generator and the open-source Python library DeepTimeSeries were used to design a reproducible benchmark. The zone temperatures and thermal loads over a future 24-h period were the forecasting targets. The benchmark results, with varying training dataset sizes from 0.3 to 0.9 y, demonstrated that the transformer architecture performs best, especially on small training datasets. The GRU and RNN were the second- and third-best performers, whereas the rank of other architectures varied depending on the training dataset size and forecasting targets. The similarity analysis revealed that simply increasing the dataset size does not necessarily improve performance, indicating the importance of training with highly similar data for the forecasting period.