Multiobjective optimization of rural multienergy microgrids using offline deep reinforcement learning

Amid the global transition toward renewable energy, rural microgrids offer a vital pathway to sustainable and accessible electricity supply. This study models the operational and carbon emission characteristics of biomass, biogas and diesel generator sets, proposing a unified multi-dimensional dynamic coupling optimisation framework. This framework incorporates carbon trading costs (65–70 RMB/tonne), government subsidies (0.04–0.05 RMB/kWh), waste recovery economic, and real-time electricity prices into the objective function. Adjustable parameters enable adaptation to rural microgrids of varying scales. To address the scarcity of rural scenario samples and the prevalence of offline data, off-policy deep reinforcement learning algorithm (SAC, TD3, DDPG) is employed for coordinated scheduling within the MDP framework, simultaneously optimising total system cost, carbon emissions, and power balance. Experiments demonstrate that during the low-sample initial phase, SAC converges faster due to its maximum entropy policy and randomised action sampling. TD3's dual-Q structure effectively mitigates Q-value overestimation and exhibits marginally superior power balance control in the early stages. With sufficient training, all three methods converge to nearly identical operational costs and zero power imbalance, validating the efficacy of offline DRL in multi-objective rural microgrid optimisation. Sensitivity analysis indicates a cost elasticity of approximately +1.96 for carbon pricing and -0.93 for subsidies. Under a combined shock of ±20%, the cost fluctuation range spans from -42% to +62%, rather than ¥501,730.74, providing actionable policy recommendations for subsidy/carbon price design. This research provides a reusable modelling and optimisation pathway for low-carbon, low-cost, and reliable operation of rural microgrids.