Powerful planning and operations of smart cities require precise estimations of the future demand of urban energy. This paper proposes a hybrid forecasting model that integrates Variational Mode Decomposition (VMD) and a Dynamic Spatio-Temporal Graph Neural Network (DST-GNN) where the implementation of the latter was achieved with PyTorch Geometric (PyG). VMD intelligently separates time-series of energy into trend, seasonal and residual signals, which are essentially the different temporal patterns. At the same time, DST-GNN models the time-varying spatial connection among regions in the city, graph-structured data is learned dynamically. Combination of these elements allows the model to deal with the non-stationarity and spatial heterogeneity that is inherent in the data on urban energy. An analysis of the method conducted on real-life datasets indicates that the proposed approach yields better performance than the state-of-the-art baselines in predicting many different forecasting horizons. The outcomes support the fact that the adaptive decomposition, matched with graph embeddings, is an efficient approach to optimize the accuracy of forecasts. The proposed approach provides a highly scalable, understandable solution to the energy management problem which would find use in urban settings, and it would help to develop intelligent infrastructure systems.