The rapid advancement of energy digitization and intelligence has engendered an escalating demand for accurate and timely prediction of natural gas consumption across diverse temporal scales. This exigency finds solutions in the swift development of intelligent algorithms. Most of the current short-term prediction algorithms are season-dependent and can achieve high prediction accuracy in temperate regions. However, in tropical and subtropical regions where seasonal patterns are not obvious, the prediction accuracy of related studies tends to decrease significantly. To address this challenge, we propose a natural gas hybrid prediction model that is applicable to temperate, subtropical, and tropical regions. This model integrates mode decomposition, feature selection and prediction. Specifically, the Feature Matching Maximization (FMM) optimization algorithm and Variational Mode Decomposition (VMD) are proposed to decompose the natural gas consumption into three sub-series: trend component, feature component, and residue component, thereby enhancing the feature representation capability for complex nonlinear characteristics. Furthermore, the Transformer model based on multi-head attention is enhanced to predict the temporal data of each sub-series. This augmentation capitalizes on the Transformer model's robust long-term dependence modeling capability, enabling it to extract valuable information from historical data. The hybrid model is experimentally validated on the natural gas consumption dataset of Xiamen, a subtropical region. Compared to other decomposition-based prediction algorithms, our findings possess superior feature extraction capabilities and higher prediction accuracy, especially in the case of nonlinear and non-stationary data. Finally, the robustness of the model is validated using datasets from multiple cities and different time spans.

FMM-VMD-Transformer: A hybrid deep learning model for predicting natural gas consumption