The modern power systems have been reshaped into sophisticated, data-intensive ecosystems with high requirements in accurate and adaptive control of energy through rapid urban electrification and renewable integration. Patronizing the traditional deep and reinforcement learning models, despite their effectiveness in pattern recognition, are typically unable to represent high-dimensional dependencies and stochastic variations of large-scale smart grids. In order to overcome the preceding, the paper presents a Quantum-Enhanced Machine Learning (QEML) framework that combines the use of quantum feature encoding, hybridized variation circuits, and reinforcement-based optimization to oppose predictive sustainability and real-time energy control. The model maps correlated grid variables into quantum states using amplitude phase encoding and jointly trains a classical LSTM layer using adaptive gradient exchange which optimizes dynamically and has low computational latency. The evidence of experimental analysis indicates that QEML reduces the cost of prediction error by 35 percent and increases the energy-saving efficiency by 17 percent compared to the state-of-the-art approaches. The adaptive quantum kernel in the framework also improves a quicker convergence as well as lowers the amount of carbon footprint. This study provides QEML as a practical solution to the concept of quantum-intelligent, self-optimizing smart grids and the connection between the computational innovation and the sustainable energy transitions.