A Novel Spatial-Temporal Hybrid Framework for Natural Gas Consumption Prediction

Abstract Efficient scheduling of natural gas pipeline networks and reliable gas supply are contingent upon the accuracy and dependability of consumption forecasts. In light of this, a novel point-interval prediction framework for natural gas consumption is proposed. Firstly, a high-quality correlation graph is constructed based on the topological structure of the pipeline network and historical natural gas consumption data, and a graph convolutional network is established for topological structure learning. Subsequently, time series data is collected, and gated recurrent units are employed to extract temporal features. Next, historical time series data and the correlation graph are fed into a hybrid deep learning network to forecast natural gas consumption. Finally, adaptive bandwidth kernel density estimation (ABKDE) is employed to compute prediction intervals for natural gas consumption at various confidence levels. The model is validated through comparative experiments using real-world natural gas consumption cases. The results indicate that in all scenarios, the model’s mean absolute relative error and root mean square relative error consistently remain below 0.11 and 0.14, respectively. Moreover, the predictive interval coverage probability and predictive interval normalized average width surpass those of other advanced models, such as decision trees and gated recurrent units, demonstrating significant improvements. This work can provide decision-making support for the scheduling and operational optimization of natural gas pipeline networks.