This paper proposes a hybrid framework that combines a Transformer-based multi-output forecasting model with a proximal policy optimization (PPO) agent for intelligent energy management in renewable-based microgrids. The Transformer forecaster captures long-range temporal dependencies in load, photovoltaic (PV), wind generation, and electricity price time series, providing 24-h-ahead predictions that serve as state inputs to the control layer. The PPO agent then learns an adaptive energy management policy that jointly optimizes three conflicting objectives: maximizing economic profit, minimizing carbon emissions associated with grid electricity, and reducing dependency on external power imports. A weighted multi-objective reward function with tunable coefficients Formula: see text is designed to flexibly adjust the trade-offs among these goals according to different operational priorities and regulatory contexts. In addition, a quantitative sensitivity analysis is conducted on both the reward weights and the Transformer input window length to investigate their impact on forecasting accuracy, policy stability, and overall microgrid performance. Simulation results on a renewable-rich microgrid scenario show that, under a representative configuration with Formula: see text, Formula: see text, and Formula: see text, the proposed framework increases operational profit, significantly reduces CO₂ emissions, and lowers grid dependence compared with baseline methods. These findings demonstrate that integrating advanced sequence modeling with reinforcement learning provides a flexible and scalable solution for sustainable, data-driven microgrid operation under uncertainty.
Sadeghpour et al. (Fri,) studied this question.