Forecasting natural gas prices using a novel hybrid model: Comparative study of different sliding windows

As the expansion of global energy trade, accurate prediction of natural gas prices has become increasingly important. Many uncertainties pose challenges to the current natural gas price forecasting, this paper proposed a new hybrid model ICEEMDAN-STLSTM-TCN-CBAM without IMF1. The natural gas prices are decomposed into sublayers of different frequencies by an improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN), and the disordered highest frequency IMF1 is eliminated. A special tanh long short-term memory, a temporal convolutional network and a convolutional block attention module (STLSTM-TCN-CBAM) are used to jointly forecast all sublayers. The experimental results based on two natural gas prices shown the model outperforms other relevant comparative models, which indicated that the proposed model has a good performance. In addition, the constructed model had good prediction performance under different sliding windows and outperformed other models. This study could proposal guidelines for natural gas market risk management and offer different insight into the era of global energy turbulence.