The intermittent nature of solar and wind generation and the dynamic load fluctuations make optimal energy management in renewable integrated smart microgrids a complex stochastic optimization problem. Traditional rule-based and model predictive control (MPC) methods are based on accurate system models and cannot adapt to real-time uncertainty. This paper proposes a Deep Q-Network (DQN) based energy management system (EMS) for an islanded microgrid consisting of a 10-kW solar PV array, 5 kW wind turbine and 20 kWh lithium-ion battery storage, supplying a mixed residential-commercial load. The DQN agent is trained on a Simulink-based microgrid environment using real data of irradiance and wind speed from Nagpur, India. The experimental results demonstrate that the proposed DQN-EMS reduces the daily operational cost by 52.2% and 35.7% and CO2 emissions by 56.8% and 41.9% compared to rule-based control and MPC, respectively. The DQN agent converges in 350 training episodes and keeps the battery state-of-charge (SOC) within safe limits (20%–80%) for 98.4% of the operating hours.