Optimization of hydropower's clean attributes: a multi-objective scheduling framework combining deep reinforcement learning and life-cycle assessment for sustainable cascade reservoir management

Climate change presents a formidable challenge to global sustainable development. Hydropower reservoir systems, while serving as critical infrastructure for clean energy and fulfilling essential roles in renewable electricity supply, simultaneously raise environmental concerns due to their substantial greenhouse gas (GHG) emissions. This dual role has sparked controversy over the clean attributes of hydropower. To address this, the present study develops an integrated multi-objective scheduling framework for cascade hydropower systems that systematically tackles operational complexity and environmental impacts by combining deep reinforcement learning (DRL) with life-cycle assessment (LCA). The proposed ε-DRLMOEA/D algorithm, driven by DRL and featuring an adaptive operator selection mechanism, significantly improves search efficiency and solution diversity compared to traditional multi-objective evolutionary algorithms. The framework couples multiple GHG emissions and carbon burial dynamics at the water-soil and water-gas interfaces and employs a minimum information gap decision model (MIGDM) to effectively balance power generation, flood control, and net GHG emissions under various hydrological scenarios. Optimized scheduling improves GHG emission benefits during the operation of cascade reservoirs, and the optimized results are subsequently linked with carbon accounting in the operational phase of the reservoir's life-cycle, achieving a comprehensive life-cycle carbon footprint assessment. Applied to the cascade reservoir complex in the lower Jinsha River basin, results indicate a 5.19 % increase in annual power generation, a 2.25 % reduction in net GHG emissions (CO 2 -equivalent), and a 4.85 % decrease in life-cycle carbon intensity. The study reveals the complex coupling between hydropower operations and greenhouse gas dynamics, demonstrating that optimizing reservoir scheduling can effectively reduce life-cycle carbon intensity. Additionally, the proposed ε-DRLMOEA/D algorithm exhibits excellent adaptability in reservoir scheduling problems across different spatial and temporal scales. This research highlights the key role of combining DRL with evolutionary algorithms in solving the dynamic evolution of hydropower's clean energy attributes and in driving low-carbon transitions and sustainable river basin management.