This study is meant to investigate the main driving factors of energy consumption structure (ECS) in China and construct a hybrid prediction model with higher accuracy. In this paper, Logarithmic Mean Divisia Index (LMDI) is applied to study the factors of energy consumption structure of China covering 1980 to 2016 and calculate the contribution of each factor. The factors that have large contribution rate are selected as the main factors of energy consumption structure to simplify prediction model. And then, the original data series of energy consumption structure and main factors are decomposed into different components by empirical mode decomposition (EMD) model to deal with the intrinsic complexity and irregularity of sequences. Finally, the components are used as input data to construct the prediction model, least square support vector machine (LSSVM), the coefficients of which are optimized by the particle swarm optimization (PSO) algorithm. The results show that prediction error of EMD-PSOLSSVM model is significantly lower than benchmark models of grey model neural network and PSOLSSVM model. The constructed model fully exploits the advantages of each algorithm of the hybrid model and improves the prediction accuracy of energy consumption structure.