International Journal of Scientific Research and Engineering Development

Demand response (DR) optimization in smart grids faces escalating complexity due to heterogeneous distributed energy resources, uncertain renewable generation, and conflicting stakeholder objectives. Traditional DR methods—rule-based systems and model predictive control—suffer from poor scalability, myopic decision-making, and inability to handle stochastic uncertainties. This paper proposes a novel Multi-Objective Deep Reinforcement Learning (MODRL) framework integrating Proximal Policy Optimization with Lagrangian constraint handling to simultaneously minimize operational costs, peak demand, and user discomfort while maximizing renewable energy utilization. The framework models DR as a constrained Markov Decision Process, employing neural network-based policy and value approximators to navigate the vast state-action space. Experimental validation on a synthetic 100-prosumer grid with real-world weather data demonstrates 23.4% cost reduction, 31.2% peak load mitigation, and 18.7% renewable penetration improvement compared to model predictive control baselines, with discomfort indices below 0.15. The proposed constraint-aware PPO variant ensures operational feasibility through adaptive penalty coefficients. Results confirm the framework's efficacy in addressing multi-stakeholder requirements, offering a scalable pathway toward intelligent, autonomous grid management.