Short-term cooling and heating loads forecasting of building district energy system based on data-driven models

• Published in: Energy and buildings 2023-11, Vol.298, p.113513, Article 113513
• Main Authors: Yu, Hanfei, Zhong, Fan, Du, Yuji, Xie, Xiang'e, Wang, Yibin, Zhang, Xiaosong, Huang, Shifang
• Format: Article
• Language: English
• Subjects: Attention mechanism; Building district energy system; Data-driven model; Interpretability; Load forecasting
• ISSN: 0378-7788
• DOI: 10.1016/j.enbuild.2023.113513

Accurate forecasting of cooling and heating loads is critical for optimizing the energy usage of devices and planning for energy storage in building district energy systems (BDESs). Data-driven models offering high prediction accuracy and efficiency are receiving significant attention for addressing short-term load prediction problems. However, load forecasting for district building levels, as well as interpretable studies of deep learning models, are still the minority of existed research. To satisfy the cooling and heating loads forecasting of an actual BDES, a framework that includes raw data acquisition, data processing, model development, and evaluation is developed and analyzed. The performance of a multi-input-multi-output prediction strategy evaluated using five data-driven models, namely extreme gradient boost (XGB), long short-term memory (LSTM), gated recurrent units (GRU), and attention-added LSTM/GRU, is investigated. The results show that XGB performs the best, with CV-RMSE values of 14.51% and 11.95% for the one-hour-ahead forecasting of cooling and heating loads, respectively. In terms of 24-hour ahead prediction, the most accurate condition is “48–24” attention-added LSTM in both cooling and heating loads forecasting, where the CV-RMSE values recorded are 26.61% and 27.58%, respectively. The accuracy and interpretability of the models improved after an attention mechanism is introduced. The most significant enhancement in the prediction accuracy is reflected by a reduction in the CV-RMSE by 2.01%. The visualized heat maps of attention weights indicate the mechanism by which the models learn and extract the load patterns, where LSTM and GRU indicate different priorities. Further analysis of the attention weight distribution reveals the effects of building thermal inertia and load periodicity.