Reinforcement Learning-enhanced Policy-aware Modeling of Smart Grid Efficiency under Carbon Constraints: Integration of SBM DEA and Dynamic Policy Response Simulation

Artificial intelligence-based sensing, forecasting, and decision optimization are being rapidly integrated into smart grid operations. Although reinforcement learning has created new opportunities for dispatch optimization and low-carbon transition under carbon constraints, most existing studies focus primarily on short-term economic or operational objectives and rarely incorporate system-level carbon reduction efficiency benchmarks into the learning process. To address this gap, this study proposes an integrated SBM-DEA and reinforcement learning framework for policy-aware smart-grid dispatch, in which carbon reduction efficiency scores are transformed from static evaluation results into dynamic learning signals for dispatch optimization. Using panel data from 30 provinces in China over the period 2011–2022, this study develops an indicator system covering capital input, labor input, electricity service output, and electricity-related carbon dioxide emissions. An SBM-DEA model with undesirable outputs is employed to measure the carbon reduction efficiency of smart grids. The estimated efficiency scores are then embedded into both the state representation and reward function of a reinforcement learning framework, where the agent learns dispatch policies that balance economic performance, carbon constraints, and efficiency improvement. A dynamic policy response simulation environment is further constructed, incorporating a hybrid energy storage system comprising battery storage and pumped hydro storage. The results show that the carbon reduction efficiency of smart grids in China exhibits stage-specific fluctuations, with annual average values ranging from 0.505 to 0.568 and pronounced interprovincial disparities. In the simulation experiments, the reinforcement learning agent trained with efficiency-based penalties achieves 7.3% lower operational costs and 8.5% higher average efficiency compared to an economic-only agent. The trained policies also exhibit clear policy-responsive behavior: when carbon prices rise, hybrid storage utilization increases and coal-fired generation declines. The main innovation of this study is that it integrates historical efficiency benchmarking with reinforcement learning-based dispatch optimization, providing a policy-aware and efficiency-guided decision-support framework for carbon-constrained smart grids.