Multi-step Natural Gas Price Forecasting using Ensemble Empirical Mode Decomposition and Long Short-Term Memory Hybrid Model

Abstract

With the characteristic of natural gas as a clean, non-toxic, and valuable energy source, its use has been increasing in recent years. Thus, maintaining stable natural gas security requires a reliable long-step price forecasting indicator with less error. We propose a hybrid theory of Ensemble Empirical Mode Decomposition (EEMD) with Long Short-Term Memory (LSTM) to perform multi-step forecasting focusing on 30 to 90 steps of the daily Henry Hub natural gas price as a dataset. Using four widespread error measurements, the proposed model provides excellent results compared to no-decomposition as the benchmark model. The proposed model provides 50% lower error results than the single LSTM. EEMD_LSTM brings values below 10 in the MAPE indicator, even up to 90-step prediction. The Diebold-Mariano test also confirms that EEMD_LSTM outperforms the single LSTM on every step with the majority of 90% confidence level. We also simulated the model by analysing the box and whiskers plot of RMSE, which shows that the variance of predicted values ranges between 1.11%. These results show that the proposed forecasting model provides robust results for the case of medium-term natural gas prices with excellent forecasting results.