A residual driven ensemble machine learning approach for forecasting natural gas prices: analyses for pre-and during-COVID-19 phases