Application Research of Deep Reinforcement Learning in Low-carbon Scheduling Mechanism for Power Grid Construction Tasks

Abstract

This paper proposes a low-carbon scheduling and carbon emission prediction framework for power grid construction machinery by integrating an improved genetic algorithm (IGA), deep reinforcement learning (DRL), and a gated recurrent unit (GRU) neural network. The proposed approach is applicable to engineering scenarios involving power transmission facilities, electromagnetic infrastructure, and communication-support systems where energy efficiency and carbon reduction are critical. A multi-objective scheduling model is established based on task dependency relationships, equipment operating characteristics, and carbon emission factors, while the GRU network is employed for real-time carbon emission forecasting. Experimental results demonstrate that the proposed method outperforms traditional GA and PSO approaches, reducing average scheduling time by 35.37% and total carbon emissions by 24.69%. The GRU model achieves an RMSE of 2.89 kgCO/h, showing superior prediction accuracy compared with RF and SVR models. The results verify the effectiveness, stability, and practical value of the proposed framework for intelligent low-carbon scheduling in complex engineering environments.