Coal-fired power plants serve as primary sources of NO_x emissions, and the efficient operation of SCR denitrification systems is crucial for reducing pollutant emissions. However, the highly dynamic changes of data during NO_x prediction processes limit the accuracy of predictive models. Therefore, a hybrid prediction framework based on modal energy difference and sample entropy, which combining variational mode decomposition (MEVMD) with genetic algorithm (GA) to optimize convolutional neural network (CNN) and long short-term memory network (LSTM), is proposed. Firstly, abnormal data are corrected using the 3σ criterion; 20 key auxiliary variables are selected via Pearson correlation coefficients. The maximum information coefficient (MIC) is employed to determine the delay time for each variable, achieving temporal alignment between features and target variables. Secondly, adaptive variational mode decomposition (VMD) precisely extracts multi-frequency features from NO_x time-series signals. Hyperparameters are optimized via GA to achieve adaptive modeling of multiple sub-modes. Finally, prediction results are generated through data reconstruction. Experimental results demonstrate that the proposed model achieves RMSE of 0.9492, MAE of 0.4969, and R² of 0.9767, outperforming comparison models.