Non-residential buildings use a significant portion of global energy, which is expected to rise sharply in the coming years due to rapid population growth. Therefore, energy consumption (EC) in this sector needs to be accurately forecasted to help reduce climate change impacts while ensuring a balance between energy supply and demand. Although several studies have recently been presented to address this problem, most still struggle to accurately capture the highly irregular and disordered patterns, resulting in suboptimal efficiency. Therefore, this study presents a new robust forecasting approach, namely Vote-CNGRU-CEEMDAN, for predicting the EC in this sector with more effective accuracy. This approach consists of three components: the CEEMDAN signal decomposition technique, a newly proposed voting regression model, and a newly proposed deep learning (DL) model. The first component decomposes the original time series data into multiple IMFs and a residual to capture nonstationary and nonlinear features accurately. The second component is used to improve the prediction of the first IMF, which is the most irregular and disordered. This component combines random forest and gradient boosting regression to create a new voting regressor capable of handling complex patterns and noise in the first IMF. The final component employs a novel DL model that more accurately predicts remaining IMFs by integrating a convolutional layer with a gated recurrent unit, effectively extracting new features and capturing temporal patterns from the input data. The proposed model is tested and validated using five common non-residential buildings and compared to several state-of-the-art models in terms of multiple performance metrics and the Wilcoxon rank-sum test to demonstrate its effectiveness and statistical significance. According to the simulation results, Vote-CNGRU-CEEMDAN significantly outperforms all compared models, with an improvement rate ranging from approximately 4.9% to 71.3% on the MAPE metric, indicating that it is a highly effective alternative for accurate EC prediction in non-residential buildings.