In the Industry 4.0 era, intelligent energy management is emerging as a critical aspect to optimize efficiency and sustainability in smart spaces. Real-time energy demand predictions are also crucial for the emergence of adaptive control systems and the stochastic allocations of resources at the appliance level. Motivated by this, this paper introduces a hybrid deep learning based framework that leverages the power of both Transformer encoders and Long Short-Term Memory (LSTM) models in the context of appliance level energy consumption forecasting. The Transformer block models cross-interactions and inter-feature correlations among multivariate time-series data as well as learns long-range dependencies, where the LSTM can refine sequential temporal patterns to improve the forecasting performance. The model is trained on time-stamped input sequences with engineered temporal signatures based on sinusoidal encoding. The customized preprocessing pipeline enables proper normalisation and windowed sequence formation for real-time forecasting purposes. This architecture, in which the proposed architecture consisting of multiple federations, is intended to be a lightweight, modular, and flexible for usage in such energy-aware applications as smart home, industrial IoT systems, etc. This research is a key step in pushing the boundary of predictive models in the smart energy systems, and leads to the possibilities of future extension in such aspects as multi-step prediction and context data fusion.