Natural Gas Load forecasting based on the fusion of VMD-SE and CNN-BiLSTM-Attention

Abstract

The rational dispatch and steady supply of natural gas resources can be effectively supported by accurate forecasting of the natural gas load. The accuracy of natural gas load forecasting is, however, low due to the fundamental characteristics of the non-stationary and non-linear natural gas load. Based on this, this paper proposes a new natural gas load forecasting model based on the fusion of VMD-SE and CNN-BiLSTM-Attention. First, the average instantaneous frequency method is used to determine the optimal K value of VMD, and then the sample entropy method is introduced to reconstruct the sub-sequence of VMD decomposition, which effectively reduces the computational overhead of the algorithm. Then, using CNN to mine the level information of multi-dimensional time series variables, using BiLSTM to mine the time series feature information, and introducing the Attention mechanism to give corresponding weight to different feature factors, highlighting the influence of key factors, the anticipated values of each reconstructed sub-series are superimposed to provide the final predicted value of the natural gas load. The model proposed in this paper has lower error than other forecasting models, can effectively improve the precision of natural gas load forecasting, and has a better forecasting effect, according to the results of establishing various forecasting models using the natural gas load data of one region of China as an example.