Journal
ENERGY
Volume 263, Issue -, Pages -Publisher
PERGAMON-ELSEVIER SCIENCE LTD
DOI: 10.1016/j.energy.2022.125976
Keywords
Natural gas consumption; Daily prediction; Domain knowledge; Temporal -spatial correlations; Deep learning
Categories
Ask authors/readers for more resources
Accurate prediction of natural gas consumption is crucial for long-distance pipeline system production and transportation planning, especially in the current political situation. Existing prediction methods lack consideration of spatial correlations and domain knowledge. This study proposes a novel deep learning prediction method (KE-GB-TSN) that integrates domain knowledge into association graph construction and captures temporal-spatial features, leading to more accurate and efficient predicted results compared to advanced models.
The accurate prediction of natural gas consumption plays a central role in long-distance pipeline system production and transportation planning, and it becomes even more important during present political situation. The existing prediction methods for natural gas consumption barely consider spatial correlations and domain knowledge. As a result, the study proposes a novel deep learning prediction method (knowledge-enhanced graphbased temporal-spatial network, abbreviated to KE-GB-TSN) for predicting natural gas consumption by integrating domain knowledge into association graph construction and capturing temporal-spatial features via a hybrid deep learning network. This study first applies the domain knowledge that analyses the operation technique of the natural gas pipeline network and combines the historical data to establish an association graph. Subsequently, the historical data and association graphs are input to a hybrid deep learning network to predict natural gas consumption. The comparative experiments are conducted by taking real-world cases of natural gas consumption as examples. At last, a sensitivity analysis of different components combination is carried out to exhibit the significance of each component in the proposed model. The results prove that the proposed model is capable of achieving more accurate and efficient predicted results compared to the advanced models, such as decision trees and gated recurrent units. The Mean Absolute Relative Errors and Root Mean Squared Relative Errors gotten by the proposed model are less than 0.11 and 0.14 in all cases, indicating an improvement compared to previous works. Additionally, it is also suggested that domain knowledge and temporal-spatial correlations are crucial for the excellent performance of the prediction model.
Authors
I am an author on this paper
Click your name to claim this paper and add it to your profile.
Reviews
Recommended
No Data Available