Aiming at the defect that traditional methods in green building environmental control systems cannot dynamically optimize multi-objective parameters, this paper proposes an adaptive control framework based on the proximal policy optimization (PPO) reinforcement learning algorithm. By constructing an 18-dimensional state space including temperature, humidity, CO₂ concentration and light intensity and a 6-dimensional continuous action space (covering air conditioning set temperature, fresh air volume and sunshade opening, etc.), a composite reward function combining energy consumption penalty (based on the real-time electricity price model) and comfort reward (integrating PMV index and CO₂ deviation) is designed, and a long short-term memory network (LSTM) is introduced to process the characteristics of sensor time series data. After completing 150,000 steps of iterative training on the EnergyPlus and Python joint simulation platform, the measured data shows that the system saves 20% energy compared with the model predictive control (MPC) method in typical winter scenarios (the average daily energy consumption is reduced from 41.8 to 33.4 kWh); the thermal comfort compliance rate is increased to 98.2%; and the strategy update time is shortened to 270 ms, which verifies the efficiency and real-time performance of the PPO algorithm in multi-objective collaborative optimization of complex building environments, and provides a scalable solution for reducing the carbon footprint of building operations.