EUVIMEDEUVIMED entdecken
Crossref

District Heating Daily Gas Consumption Prediction Based on Transformer with Thermal Inertia Prior Position Encoding: A Case Study of Urumqi

Zhang Shuya, Li Zhenfei

Journal of Energy Science and Technology · 02.11.2026 · 5 (2) · 26-38

DOI: 10.57237/j.jest.2026.02.002 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Journal of Energy Science and Technology, online erschienen am 02.11.2026. Verfasst von Zhang Shuya und 1 weiteren Autor:innen. Die Arbeit nennt 22 Literaturangaben.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

In recent years, with the increasing attention to ecological protection, the Chinese government has set the carbon neutrality development target. The energy supply of district heating systems in northern cities is gradually shifting from coal to cleaner natural gas. Accurate prediction of natural gas load is critical for energy dispatch planning and demand response. Nevertheless, natural gas load is not only affected by random fluctuations of meteorological factors, but also exhibits complex nonlinearity and significant hysteresis characteristics due to the inherent thermal inertia of heating systems, making it difficult for existing methods to meet the prediction accuracy requirements in practical applications. To address this issue, this paper adopts measured operational data from three heating enterprises in a city and proposes a natural gas load prediction algorithm for district heating systems based on a sequential Transformer model (TI-STM) with thermal inertia position encoding. According to the physical law that recent observations exert stronger impacts while distant ones exert weaker impacts on heating load, an exponentially decaying position encoding is designed. Dimension-wise learnable decay coefficients are adopted to characterize the temporal differences of thermal inertia, replacing the traditional sinusoidal position encoding. Meanwhile, meteorological variables of future time steps are fed into the decoder as known covariates to guide the model to utilize temperature priors. With the multi-head attention mechanism, the model mines complex implicit features within the load sequence and quickly focuses on different key segments of the sequence, mitigating long-range dependency problems and reducing information loss. Comprehensive comparative experiments are carried out against advanced algorithms including SVM, LSTM, MLP, RNN, TCN and TI-STM. The results demonstrate that the proposed model achieves superior prediction accuracy, with mean absolute percentage errors (MAPE) of 1.77%, 2.05% and 2.65%, respectively. 近年来,随着生态保护日益受到重视,中国政府确立了“碳中和”的发展目标,北方城市区域供热系统的能源供给正逐步由煤炭转向更清洁的天然气。天然气负荷的准确预测对能源调度规划与需求响应至关重要。然而,天然气负荷不仅受气象因素随机波动的影响,还因供热系统固有的热惰性而呈现出复杂的非线性与显著滞后特性,导致现有方法的预测精度难以满足实际应用需求。为此,本文以某市三家供热企业的实测运行数据为对象,提出一种基于热惰性位置编码序列Transformer模型(TI-STM)的区域供热系统天然气负荷预测算法。该方法针对供热负荷“近端影响强、远端影响弱”的物理规律,设计指数衰减形式的位置编码,以逐维可学习的衰减系数刻画热惰性在时间尺度上的差异,替代传统正弦位置编码;同时将未来时刻的气象要素作为已知协变量注入解码器,引导模型利用温度先验。模型借助多头注意力机制挖掘负荷序列中隐含的复杂特征,快速聚焦序列中不同的关键部分,在避免长距离依赖问题的同时减少信息损失。本文与SVM、LSTM、MLP、RNN、TCN和STM等先进算法进行了详细的对比实验。结果表明,所提模型具有更高的预测精度,其平均绝对百分比误差分别达到1.77%、2.05%和2.65%。

Zitieren

APA

Shuya, Z., & Zhenfei, L. (2026). District Heating Daily Gas Consumption Prediction Based on Transformer with Thermal Inertia Prior Position Encoding: A Case Study of Urumqi. Journal of Energy Science and Technology, 5(2), 26-38. https://doi.org/10.57237/j.jest.2026.02.002

Vancouver

Shuya Z, Zhenfei L. District Heating Daily Gas Consumption Prediction Based on Transformer with Thermal Inertia Prior Position Encoding: A Case Study of Urumqi. Journal of Energy Science and Technology. 2026;5(2):26-38. doi:10.57237/j.jest.2026.02.002

BibTeX
@article{Shuya2026,
  author = {Shuya, Zhang and Zhenfei, Li},
  title = {District Heating Daily Gas Consumption Prediction Based on Transformer with Thermal Inertia Prior Position Encoding: A Case Study of Urumqi},
  journal = {Journal of Energy Science and Technology},
  year = {2026},
  volume = {5},
  number = {2},
  pages = {26--38},
  doi = {10.57237/j.jest.2026.02.002}
}
Kostenlos starten →