Objective Pipeline network companies need to forecast the short-term natural gas loads of downstream users to ensure a stable gas supply. However, traditional load forecasting methods have various problems such as poor fitting effect and low forecasting accuracy. A combined short-term natural gas load forecasting model based on the Complementary Ensemble of Empirical Mode Decomposition (CEEMD)and the Long Short-Term Memory (LSTM) network is proposed to improve the accuracy of short-term natural gas load forecasting.

Methods To fully explore the internal hidden feature information of the load series and to avoid the mutual interference between different component features and additional noise, the original load series was decomposed into a finite number of Intrinsic Mode Function (IMF) components through CEEMD. Different IMF components were input into the LSTM model for iterative multi-step forecasting, the hyperparameters of the LSTM model were optimized based on the Bayesian algorithm to improve the learning effect and the forecasting accuracy, and the forecasting results of individual components were superimposed and reconstructed to obtain the final forecasting results.

Results The comparison of forecasting results obtained by different combined models indicates that the errors in the values forecast by CEEMD-LSTM based on the real values are smaller, proving that the forecasting accuracy is higher. The Bayesian hyperparameter tuning further improves the forecasting accuracy but takes longer forecasting time.

Conclusion Compared with other forecasting models, the combined CEEMDLSTM forecasting model can effectively extract the time order information of load series, eliminate the influence of nonlinear factors, and reduce reconstruction errors while suppressing modal superposition, thus improving forecasting accuracy. This method can provide a reference for the scheduling management and operation optimization of natural gas pipeline networks.