Accurate power consumption forecasting is essential for optimizing resource management and advancing sustainability in residential buildings. To address this challenge, this study introduces a hybrid framework integrating Variational Mode Decomposition (VMD) and a Gated Recurrent Unit (GRU) neural network to predict energy demand in Moroccan households. Specifically, the methodology tackles the non-stationarity and noise inherent in historical consumption data by decomposing the signal into interpretable Intrinsic Mode Functions (IMFs). This process isolates critical components such as seasonal trends, consumption peaks, and residual noise. To validate the framework, a dataset of 3 Moroccan households across diverse climatic regions and building types was used. The results demonstrate state-of-the-art performance, with MAE = 3.177 Wh, RMSE = 8.295 Wh, and a coefficient of determination of 0.999. Furthermore, rigorous testing on unseen households confirms the model’s robustness in handling heterogeneous consumption patterns. These findings underscore the synergistic potential of combining signal decomposition with deep learning, thereby offering a scalable tool for energy planners to enhance grid stability, reduce operational costs, and support Morocco’s transition toward sustainable energy management.