Short-term natural gas consumption prediction based on wavelet transform and bidirectional long short-term memory optimized by Bayesian network

In the energy structure, natural gas is an important clean energy, and its consumption forecast is of great significance for energy policy formulation, pipeline network planning design, and peak shaving capacity determination. In the field of forecasting, wavelet transform (WT) is widely used to process time series, especially natural gas consumption. However, in the application of the WT, the wavelet's orders and levels are randomly fixed. To solve this problem, this study proposes a hybrid model, which is based on bidirectional long short-term memory (BiLSTM). In addition, Bayesian networks are used to optimize the hyperparameters of BiLSTM. And the different wavelets' orders and levels are utilized to deal with the daily natural gas consumption (DNLGSCN). When each level of Coiflets wavelets' four and five orders are used to process the DNLGSCN, the fifth level of Coiflets wavelets' four and five orders have the highest prediction accuracy, respectively, which is 0.3318 and 0.1773, taking U1 as an example. Using Symflets wavelet to decompose the DNLGSCN also gives the same result. Compared with the decomposition method (e.g., ensemble empirical mode decomposition), the prediction performance of the above four wavelet decomposition methods is better. The main conclusions are that decomposing the DNLGSCN with the fifth level of Coiflets wavelet's five orders has the best prediction performance.