Deep hybrid models for daily natural gas consumption forecasting and complexity measuring

Accurately forecasting daily natural gas consumption (NGC) is challenging due to its complex time-varying high-dimensional features. Therefore, this paper proposes a novel hybrid model for NGC consisting of two stages. In Phase I, the parallel local outlier factor-isolation forest-based data preprocessing algorithm applied denoise input and preserve valuable II, convolutional neural network (CNN)–stacked long short-term memory (SLSTM) can extract spatio-temporal features deep learning forecaster. The influencing mechanism different factors on prediction accuracy has not been mathematically experimentally revealed. Thus, complexity measure ( Γ NG ${\Gamma }_{\mathrm{NG}}$ ) developed by combining variation coefficient analysis, fluctuation kurtosis analysis. historical sets three representative cities were collected in case studies, six advanced models selected comparison. results reveal that high-quality preprocessed I enables better performance all models. CNN-SLSTM outperforms other algorithms with an average improvement about 26%–49%. Prediction perform low periods predominantly industrial users their lower complexity.