Sales forecast of new energy vehicles in China based on multi-source information fusion and link prediction

This study develops an integrated forecasting framework that leverages multi-source information fusion to improve predictions of new energy vehicle (NEV) sales in China. The dataset incorporates historical sales records, key influencing factors, and unstructured Baidu Index (BI) data. Our framework first applies complementary ensemble empirical mode decomposition with adaptive noise (CEEMDAN) combined with sample entropy (SE) reconstruction. This procedure decomposes the original sequences into high-, medium-, and low-frequency components, yielding data that are more regular and informative while retaining underlying dynamics. Forecasting is then conducted through a structured combination of methods tailored to each frequency band. High-frequency components are modeled by converting the time series into directed visibility graphs (DVGs) and performing prediction as a link-prediction task on the resulting complex networks, implemented using a particle-swarm-optimized support vector regression (PSO-SVR) model. Medium- and low-frequency components are predicted using backpropagation neural networks (BPNN) and Lasso regression, respectively. Empirical evaluations show that integrating multi-source information, CEEMDAN-SE decomposition, DVG-based modeling, and frequency-specific predictors substantially improves forecasting accuracy compared with benchmark approaches. The findings show the potential evolution of China's NEV market and highlight the drivers of sales dynamics, providing evidence to inform policy formulation and sectoral strategic planning.