Comparative Analysis of Hybrid SVMD-LSTM Models in Energy Demand Predicting for Renewable and Non-Renewable Sources in China

Zekai Nie¹Luhao Cao^2*Dr. Sarita Gajbhiye Meshram³

• ¹INTI International University, Nilai, Malaysia
• ²Hunan University of Technology, Zhuzhou, China
• ³Indian Institute of Management Nagpur, Nagpur, India

Predicting energy demand in China using hybrid models is vital for effective policy-making, economic planning, and sustainable development. These advanced models offer the accuracy and flexibility necessary to navigate China's complex and rapidly evolving energy landscape, supporting both national energy security and the global transition to cleaner energy sources. This research proposes a flexible framework that combines renewable and non-renewable energy sources with machine learning techniques to predict China's monthly overall energy demand. Two distinct scenarios were examined: the first focused on renewable energy sources—hydropower, bioenergy, wind, and solar—while the second encompassed fossil-based energies, specifically coal and natural gas. Data spanning 2015 to 2025 underwent preprocessing and were subsequently partitioned into training (70%) and testing (30%) sets via random sampling. To predict monthly energy demand, two hybrid models—Long Short-Term Memory with Sequential Variable Mode Decomposition (SVMD-LSTM) and Convolutional Neural Network with a Long Short-Term Memory (CNN-LSTM), along with a standalone Support Vector Regression (SVR) model —were implemented. Performance was assessed using visual plots and quantitative metrics, including the coefficient of determination (R²), root mean square error (RMSE), Nash-Sutcliffe efficiency (NSE), and mean absolute percentage error (MAPE). The findings indicated that the SVMD-LSTM hybrid model outperformed both the CNN-LSTM and SVR models across the two evaluated scenarios. Under the first scenario, it achieved an RMSE of 19.49 TWh, an R² value of 0.979, NSE of 0.978, and a MAPE of 2.74%. In the second scenario, the model yielded an RMSE of 23.69 TWh, R² of 0.974, NSE of 0.967, and MAPE of 2.611%. Through comparative analysis using the proposed hybrid model, the influence and significance of each scenario on total energy demand can be effectively assessed, enabling data-driven and reliable decision-making.