Abstract:
The growing integration of renewable energy sources and smart appliances into residential grids has increased the importance of Intelligent Home Energy Management Systems (HEMS). These systems aim to reduce electricity costs while maintaining user comfort, yet designing autonomous schedulers that can perform reliably in uncertain and dynamic environments remains a significant challenge. The complexity arises from variable renewable generation, time-varying tariffs, and heterogeneous appliance demands that must be coordinated in real time. This thesis develops and evaluates a hybrid AI-based HEMS controller within a high-fidelity simulation framework constructed from real-world, per-minute household appliance load and solar irradiation data. The research followed a staged methodology. A baseline reinforcement learning (RL) agent was first implemented, demonstrating significant cost reductions relative to an unmanaged system. The framework was then extended into a proactive multi-appliance scheduler that shifted high-power loads in anticipation of solar availability and tariff fluctuations, further enhancing efficiency. The principal contributions of this work are threefold: (i) the development of a high-resolution simulation environment for testing HEMS algorithms under realistic operating conditions, (ii) the design of a hybrid GA-RL scheduling framework that evolves an RL reward function using a genetic algorithm as a meta-optimizer, and (iii) the demonstration of scalable multi-appliance scheduling that balances economic and operational objectives. Across three months of simulation, the GA-optimized controller consistently managed six to eight high-power appliances, achieving reliable task completion under all tested scenarios. Compared with an unmanaged baseline, the optimized system achieved annual cost reductions of 70–75% under a net metering policy. These findings suggest that GA-enhanced RL offers a powerful strategy for addressing the scheduling challenges of residential energy management. Beyond performance validation, the proposed framework provides a flexible platform for exploring the interplay between hardware constraints, control strategies, and economic incentives. Taken together, the results highlight both the practical potential of hybrid AI-based HEMS and the opportunities for future research in real-world deployment, demand-response integration, and adaptive policy-driven optimization.
