Day-ahead city natural gas load forecasting based on decomposition-fusion technique and diversified ensemble learning model

• Published in: Applied energy 2021-12, Vol.303, p.117623, Article 117623
• Main Authors: Li, Fengyun, Zheng, Haofeng, Li, Xingmei, Yang, Fei
• Format: Article
• Language: English
• Subjects: Diversified base-learner; Fast ensemble empirical mode decomposition; Feature contribution analysis; LightGBM; Natural gas load forecasting; Stacking ensemble learning approach
• ISSN: 0306-2619; 1872-9118
• DOI: 10.1016/j.apenergy.2021.117623

•This paper proposed a novel model for city natural gas forecasting.•We adopted the fast ensemble empirical mode with data replacement function.•The mode decomposition-fusion technique is exploited to denoise original sequence.•We considered the impact of the base-learner’s capability and diversity.•The accuracy, adaptability, and stability of the developed models are superior. Accompanying the trend of low-carbon energy consumption, natural gas has ushered in the energy transition era’s development. However, rapid growth has thrown off the balance of urban natural gas supply and demand, resulting in gas shortages in many Chinese cities for several consecutive years. This work proposes a novel model for short-term load forecasting that combined the decomposition-fusion technique with a replacement data function, feature selection, and a diversified Stacking ensemble learning model. First, fast ensemble empirical mode decomposition is used to disintegrate the original complex nonstationary time series data into several modes. To ensure accurate information and computational efficiency while preventing excessive decomposition, the Pearson coefficient is used to fuse highly correlated patterns further. Second, hybrid feature engineering is used to select high contribution candidate input variables. Finally, K-Flod cross-validation is performed in each base-learner to enhance generalization capability during the training process. The empirical results prove that the base-learners’ capabilities and discrepancy will significantly impact the model (MAE = 167.409, MAPE = 3.125, RMSE = 234.654). Even if different types of city data are used, the proposed model’s effectiveness and robustness in gas load forecasting is not weakened, and decomposition-fusion technology can boost the model’s effectiveness. However, it cannot play a decisive role; the ensemble learning approach can integrate the ascendancy of the single model while effectively reducing the risk of falling into a local minimum. The developed model has good application prospects in natural gas dispatch and control systems as it outperforms alternative models in prediction accuracy, adaptability, stability, and generalization ability.