A combination model based on wavelet transform for predicting the difference between monthly natural gas production and consumption of US

The prediction model's performance in view of the wavelet transform (WT) is affected because the wavelet basis function (WBF) and its orders and layers are determined randomly. To solve this problem, this research devises a crossover experiment with 160 components of each WBF (Coiflets and Symlets) and forecasts 320 schemes with sparse autoencoder (SAE) and long short-term memory (LSTM), developing a combination model with WT, SAE, and LSTM. Furthermore, to verify the performance of the combination prediction model, the difference between the natural gas production and consumption in the U.S. is determined which is taken as an example. The results indicate that the SAE-LSTM exceeds other AI models (e.g. ELM), and WT outperforms other preprocessing algorithms (e.g. EMD) based on forecasting accuracy. The best performance of the established model is obtained by using the two orders six layers of Coiflets, and six orders seven layers of Symflets for natural gas production and consumption. In addition, the average difference between consumption and production of natural gas is 10.6809 Bcfpd. To make up for such a gap, some import methods can be adopted. It can be concluded that this study can provide a reference for other time-series prediction and natural gas policymakers. (c) 2021 Elsevier Ltd. All rights reserved.