Short-term natural gas load forecasting based on EL-VMD-Transformer-ResLSTM.

Area of Science:

• Energy Systems Analysis
• Computational Intelligence
• Time Series Forecasting

Background:

• Urban natural gas consumption is influenced by evolving resident habits and lifestyles.
• Accurate forecasting of natural gas demand is essential for efficient energy management and grid stability.
• Existing forecasting models often struggle to capture complex patterns in natural gas consumption data.

Purpose of the Study:

• To develop a novel hybrid forecasting model for natural gas consumption.
• To enhance prediction accuracy by integrating advanced machine learning techniques.
• To address the limitations of traditional forecasting methods in dynamic urban environments.

Main Methods:

• Ensemble Learning (EL) using XGBoost, CatBoost, and LightGBM as base learners.
• Variational Mode Decomposition (VMD) to decompose the natural gas load sequence into intrinsic mode functions (IMFs).
• A novel Transformer-ResLSTM network with a modified decoder structure and residual connections for prediction.

Main Results:

• The proposed hybrid model demonstrated superior performance compared to ARIMA, Transformer, GRU, and LSTM.
• Significant reductions in Mean Squared Error (MSE) by 92-98% and Mean Absolute Error (MAE) by 74-83%.
• The model effectively extracts inherent features and improves the accuracy of natural gas load forecasting.

Conclusions:

• The hybrid EL-VMD-Transformer-ResLSTM model offers a significant advancement in natural gas consumption forecasting.
• The method shows substantial potential for practical application in energy management and urban planning.
• Accurate forecasting supports optimized resource allocation and grid reliability.